OLDGABE — research log

Does the CAUTION-recovery re-ranker actually beat the production tier scorer? — head-to-head + de-confounding

The framing. OLDGABE's fusion stage sorts every candidate fire into tiers; the great majority land in a low-confidence CAUTION tier and are never surfaced. The question that had never been measured: among those CAUTION events, can a learned recovery model rank the genuine hidden fires better than the production tier/confidence scorer does?

BLUF. On the operationally meaningful metrics the recovery model decisively beats production. At an 80%-precision operating point it recovers 60.5% of confirmed real fires vs production's 9.5% (Truth1), and 27.2% vs 0% under a second, HMS-independent truth (Truth2). Production holds a small global-AUC edge on Truth1 (0.951 vs 0.936) — but that edge is a confound: production uses the NOAA HMS smoke channel, and Truth1 is 91.5% HMS-confirmed, so it is partly scoring HMS against HMS. When the truth is switched to an independent polar-satellite corroboration, production's AUC edge reverses (recovery 0.845 vs production 0.786).

Populations & truths (reproducible). Candidate set = every canonical event with event_corroboration_tier.tier='CAUTION' that has an event_ground_truth row (n = 197,325). Truth1 (multi-source) = event_ground_truth.confirmed (6,867 positive, 3.48%; for CAUTION this channel is 91.5% NOAA-HMS-confirmed). Truth2 (neutral polar) = a VIIRS/MODIS active-fire detection from a different canonical event within 5 km and −1..+14 days of the event's first detection (7,791 positive, 3.95%). The two truths overlap only partially (both-positive 1,753; Truth1-only 5,114; Truth2-only 6,038), so Truth2 is a genuine cross-check, not a relabelling.

Features (non-leaky, detection-time only). Evidence is joined excluding the confirmation channels IRWIN, HMS, SAR and S2. Per event: n_det, n_fam, max/mean_frp, max/mean_conf, span_h, n_days, max_ti4, ti4−ti5 contrast, frac_night, mean_pixel_area, max_ti5, abi_temp_k, abi_area, modis_brightness, mean_hour, fp_dist_km, frp_per_pixel. Model: gradient-boosted trees (HistGradientBoostingClassifier, class-weight balanced), scored out-of-fold under 5-fold GroupKFold grouped on 0.2° spatial cells (a held-out cell is never in training). The production scorer is the fixed tier_and_conf() confidence; its value is pinned at 0.02 for ~94% of CAUTION events (the GOES/ABI-only majority), which is why it cannot rank within the tier.

Result.

De-confounding (does the win survive removing the reporting channel?). Feature-group AUCs, per truth — month-only / geography-only (lat,lon) / detection-only / all:

Reads: (1) Season is not a confound — the candidate window is summer-only, so the summer-share of positives and of negatives is identical (gap = 0) and month-only ranking is near-chance (~0.57). (2) Geography is a large prior under both truths (0.903 vs 0.877), which already argues it is real fire geography rather than pure reporting bias — a reporting-only signal would predict the HMS truth far better than the polar truth. (3) Production's Truth1 AUC lead does not survive the neutral truth. The honest operating-point story is unchanged: production cannot deliver a high-precision set within CAUTION; the recovery model can.

Adding geography to the recovery re-ranker — a defensible lift that grows under the neutral truth

The framing. Entry 0001 showed geography alone ranks CAUTION fires surprisingly well. Wildfires cluster by fuel, terrain and climate, so a location prior should carry real signal — but a location prior can also just memorise where fires get reported. This entry adds geography to the recovery model and tests, honestly, whether the lift is real fire behaviour or a reporting/coverage artifact.

BLUF. Adding raw lat/lon to the detection feature set (recovery+geo) is a defensible improvement over both the detection-only recovery model and production, on both truths and every metric. The decisive de-confounding result: the geo lift is larger under the HMS-independent polar truth (+0.081 AUC) than under the reporting-heavy truth (+0.036 AUC). If geography were merely encoding reporting bias, its benefit would shrink when the truth stops depending on reporting — instead it grows. So the geographic signal is real fire geography.

What "geo" is (hard data bound). OLDGABE currently has no populated land-cover / fuel / terrain / climatology layer — the fire-weather cell table is empty, weather stations are a current snapshot (temporally invalid for historical events), and the raw detection JSON carries no geographic fields. So the only non-leaky geographic priors available are raw lat/lon + distance-to-known-false-source. A Truth1-only supplementary that adds an out-of-neighbourhood polar-thermal climatology density (a fuel/fire-proneness proxy; circular against Truth2, hence Truth1-only) lifts a little further (AUC 0.978, recall@80% 0.771), indicating headroom if a real independent fuel/climatology layer were ingested.

Method delta from entry 0001. Same populations (n = 202,005; the DB ingested more June/July events between runs), same non-leaky features, same 5-fold GroupKFold on 0.2° cells, same two truths. New models: recovery+geo = detection features + lat + lon; geo-only = lat,lon; all = detection + lat + lon + month. A June↔July temporal-transport split is added as the only distribution shift the data permits.

Result — head-to-head.

The geo lift over detection-only is +0.036 AUC on Truth1 and +0.081 AUC on Truth2 — it grows under the neutral truth. recovery+geo also closes and reverses the one metric where production had led (Truth1 AUC, now 0.973 vs 0.950). Detection and geography are complementary: recovery+geo (0.941) beats both detection-only (0.860) and geo-only (0.886), so neither is redundant. Month adds nothing (all ≈ recovery+geo).

Result — temporal transport (train one month, test the other).

recovery+geo beats detection-only in all four cells and production in three of four (it loses only Jun→Jul on the HMS-confounded Truth1, by 0.007). Under the neutral truth it wins both directions by wide margins, and it stabilises the hard direction (Jun→Jul detection-only falls to 0.888; recovery+geo holds 0.934). Within the available window, the geographic prior transports across time and does not hurt.

The cross-sensor fusion safety-net: two independent satellites agreeing is 4.5× better than one

The framing. Each satellite makes its own noisy fire hunches. OLDGABE's fusion daemon is the “two-witnesses” rule: it links detections of the same fire across sensors and only trusts a candidate strongly when independent instruments agree. This entry measures how much that agreement is worth.

BLUF. OLDGABE ingests ~1.80 million raw per-sensor fire detections. The fusion daemon cross-links them into 11,441 multi-sensor clusters and grades them. Requiring ≥2 independent CORE thermal sensors to agree (the “confirmed” tier, n=620) raises the IRWIN/perimeter match rate to 37.7%, versus 8.4% for a random raw single detection — a 4.5× precision lift — while GOES-only single-frame candidates (the “early” tier) match just 5.5%. Separately, the false-positive catalog + status logic demote 620,160 raw candidates (34.4%) to masked-fp so they never reach the map.

Result.

Dozier subpixel strict-gate: a clean precision-for-recall trade, and why we keep it out of fusion

The framing. The Dozier subpixel method estimates what fraction of a coarse GOES pixel is actually burning (p_subpixel). Tightening a threshold on that fraction is a dial between “catch everything, mostly noise” and “only the unambiguous, but almost nothing.”

BLUF. Gating on p_subpixel trades recall for precision cleanly: the IRWIN match rate rises from 2.8% (ungated) to ~18–27% at p_subpixel ≥ 0.005–0.01, but keeps only ~10–19% of Dozier's own positives. Because the high-precision operating point yields so few detections — and essentially none the polar sensors don't already provide — OLDGABE's fusion daemon deliberately excludes Dozier. A measured “keep it, but don't fuse it” decision, not an oversight.

Result (60,000 sampled Dozier detections; 1,681 IRWIN-positive = 2.8% base):

The deployed strict gate independently scores 26.3% precision (G19) / 48.3% (G18) at ~1.3% whole-fire recall in the daily-scorer (entry 0007) — consistent with this sweep.

Negative result: GOES temporal persistence does not add defensible recall beyond polar

The framing. A geostationary satellite stares at the same spot every few minutes, so a tempting idea is: “if GOES keeps seeing a hotspot in one place for hours, that persistence is extra evidence of a real fire the polar satellites missed between overpasses.” We tested it. It does not hold up.

BLUF. GOES near-match “recall” is saturated by coverage: 92.4% of IRWIN fire cells have some GOES detection within ~5 km, simply because GOES-ABI emits candidate hotspots almost everywhere over CONUS (its measured precision is only ~15% — 85% are not fire). Requiring temporal persistence (a cell seen in ≥3 distinct hours) only drops this coverage-inflated number (92.4% → 66.1%); it never adds genuine recall. The apparent 43.8% “marginal over polar” is the same coverage/false-positive floor, not independent detections. Conclusion: GOES temporal persistence is not a validated recall lever — OLDGABE's defensible cross-sensor recall comes from polar corroboration and fusion (entry 0003).

Result (45-day window; 2,592 IRWIN truth cells at 0.05°):

Wildfire-camera smoke detection: highest precision of any source, but coverage/visibility-bound

The framing. OLDGABE runs a smoke/flame classifier on public wildfire-camera frames and, when two cameras see the same plume, triangulates the bearings into a map location. Cameras see smoke a satellite can't — but only where a camera is pointed, in daylight, through clear air, within range.

BLUF. When a camera fires on a fire IRWIN also records, it is usually right — daily-scorer precision 66.7%, the highest of any single OLDGABE source. But it matches only 1.2% of IRWIN fires: coverage and visibility (line-of-sight, daylight, haze, range) are the binding constraint, not the classifier. Honest status: a high-precision corroborator where cameras can see, with no proven standalone field value yet at network scale.

Result (daily-scorer, vs IRWIN 10 nm / 24 h):

Per-sensor scorecard vs NIFC IRWIN — what each US datalink actually contributes

The framing. Rather than argue which feed matters, OLDGABE runs a daily replay scorer that matches every sensor's detections to NIFC IRWIN ground truth and records precision, recall and count per source. This entry is that scorecard, unedited.

BLUF. Polar VIIRS is the best single recall+precision workhorse; the fused OLDGABE output leads recall (72.4%); cameras lead precision (66.7%); GOES-ABI is high-volume/low-precision; SAR and Sentinel-2 SWIR contributed nothing in this window. This is the empirical basis for weighting the fusion daemon (entry 0003) toward polar + agreement.

Result (daily-scorer run 18, 24 h window ending 2026-07-08, match 10 nm / 24 h vs IRWIN):

Reframing the camera 1.2%: coverage-conditional recall — cameras detect up to ~25% of fires they can actually see

The framing. Entry 0006's 1.2% camera recall was measured against every US fire — but a camera can only see fires inside its viewshed. The fair question is: where a camera can physically see, how often does it catch the fire? And is the gap infrastructure (coverage) or algorithm (detection)?

BLUF. Conditioning on coverage reframes the number by 4–20×: within 60 km of an online camera, cameras detect 24.7% of IRWIN fires; within 20–40 km, 8.8–19.0%. But two gaps remain and split cleanly: coverage — only 28–59% of fires are within R km of an online camera — and detection-in-coverage — even where a camera can see, only ~5–25% of fires are caught (night, haze, range, azimuth). Precision rises with R to 72% (matching entry 0006/0007). Cameras remain a high-precision corroborator; the primary lever is detection-in-coverage, but it is capped by coverage.

Result (3,903 deduped IRWIN incidents; 10,227 online cameras; 1,016 smoke≥0.5 hits):

Repositioning Dozier: a median 6-hour early cue, and a multi-feature gate that dominates the single threshold

The framing. Entry 0004 showed the single-p_subpixel Dozier gate is high-precision but low-yield, and excluded from fusion. Two follow-ups ask whether Dozier is worth more than that: is it early, and can a smarter gate move the curve out rather than just slide along it?

BLUF. Both hold. (1) Lead-time: for fires seen by both Dozier and a polar sensor (n=472), Dozier is earlier 53% of the time, median ~6 hours earlier — an early-cue / lead-time tier, not a recall source. (2) Multi-feature gate: replacing the single threshold with a 13-feature classifier lifts PR-AUC from 0.044 to 0.179 (4×) and precision-at-recall from 8.0% to 39.4% (at 10% recall). The multi-feature curve dominates the single threshold at every operating point.

Result A — lead-time (472 fires seen by both Dozier and polar):

Result B — multi-feature gate vs single p_subpixel (out-of-fold):

Is re-enabling Sentinel-2 SWIR worth it? Excellent coverage, ~4% active-fire marginal recall — a measured no

The framing. Sentinel-2 SWIR sits at zero in the scorecard (entry 0007). SWIR can see active flame at 20 m — so would re-enabling full CONUS S2 processing recover fires the current sensors miss, and is it worth the cost?

BLUF. Of 1,764 IRWIN fires in the S2 window, only 191 (10.8%) were missed by every current sensor. Over those missed fires, S2 scene availability is 100% — but an actual SWIR active-fire signal appears at only 4% of them (2/50, CI 1–14%), and already-ingested S2 added 0 unique fires. Extrapolated unique marginal recall ≈ 4% of the missed set — small — at a compute cost of ~1–2 TB/week and up to 5-day latency. Verdict: not worth full re-enablement for timely alerting.

Result.

Making standalone Dozier usable: a multi-feature gate (49% vs 12% precision) and an early-cue tier (median 4.3 h before polar) — shadow-only

The framing. The production pipeline hides standalone GOES ABI Dozier subpixel detections entirely — on their own they are ~0.05% wildfire precision (entry 0004/0007), so Dozier only counts as corroboration for a fire another sensor already carries. Two levers could change that without touching the live gate: (1) replace the single p_subpixel threshold with a multi-feature gate, and (2) surface a Dozier early-cue tier for detections that fire before polar confirmation. This entry measures both, shadow-only.

BLUF. On 100,000 random Dozier candidates (IRWIN/perimeter truth, 2.04% base, 5-fold spatial CV), a 13-feature classifier reaches 49.3% precision at the ~10% operating recall where the single p_subpixel threshold sits at only 11.6% — a +37.7 pp gain (PR-AUC 0.234 vs 0.051). It survives dropping the detection-density feature (39.6% precision, PR-AUC unchanged), so it is not a spatial-clustering artifact. Separately, Dozier detections that precede a polar detection are 38.4% IRWIN-confirmed — ~19× the 2.0% Dozier baseline — at a median 4.3 h lead. Both are GO candidates. Nothing is surfaced.

Piece 1 — multi-feature gate vs the single threshold.

Flip operating point (go/no-go). Cut over to shadow-score τ = 0.965 → 49.3% precision at 10% recall (the recall the current gate already runs at, so no recall is lost while precision ×4.3). For a stricter 40% surfacing bar the gate still returns 32.8% recall. At matched precision the single threshold collapses to 0.2% recall — the gate holds ~50× more. OOF ROC-AUC 0.869.

Piece 2 — Dozier early-cue tier.

Pre-flip check kills the flip: the multi-feature Dozier gate collapses out-of-time — we are NOT promoting it

The framing. Entry 0011 reported the 13-feature Dozier gate reaching ~49% precision vs 12% for the single p_subpixel threshold under 5-fold spatial cross-validation, and flagged it a GO candidate pending a fresh out-of-window confirmation. This entry is that confirmation. Before wiring anything into the live gate we ran the strongest temporal test the data allows — train on an earlier window, test on a strictly-later held-out window — with a hard decision rule: promote only if the multi-feature gate still clearly dominates out-of-time; otherwise stop.

BLUF. It does not dominate — it collapses. Trained on the earlier window and tested on strictly-later data, the multi-feature gate falls to near-chance ranking (holdout ROC-AUC 0.60–0.64) and is beaten by the trivial single p_subpixel threshold at every operating point, in both a count-based split and a positives-balanced split. The exact flip point τ=0.965 transfers to zero holdout recall. Decision: HOLD — do not flip. The 49% spatial-CV figure was time-correlated leakage. The production Dozier gate stays exactly as-is (standalone Dozier still hidden), and the shadow canary stays gated.

Result — holdout precision at the current gate's operating recall (~10%).

Precision across the recall curve (single vs multi-feature, holdout). Single wins at every point:

Does the Dozier early-cue tier survive out-of-time? Not validated — its 38% lives entirely in a not-yet-truth-matured window

The framing. The gate lever failed out-of-time (entry 0012). The other Dozier lever from entry 0011 is the early-cue tier: a Dozier detection that precedes a polar (VIIRS/MODIS) detection of the same fire within 5 km/+48 h, reported there at 38.4% IRWIN-confirmed precision (~19× the ~2% Dozier baseline) with a median 4.3 h lead. It is a deterministic rule, not a fitted model, so it cannot “overfit” in the ML sense — the honest test is temporal stability: measure precision + lead on an earlier vs a strictly-later, truth-matured window.

BLUF. A clean matured earlier-vs-later split turned out to be impossible, and that fact is the finding. Dozier detected_at is bimodal: ~82% is a 2-day historical backfill (2026-05-30/31), June 01–25 has zero Dozier rows, and live ingestion only resumed ~07-01. So the only fully ±14 d truth-settled data is the backfill; the recent live window is days old and its IRWIN truth is right-censored. On the settled-truth backfill the tier scores just 2.5% precision; the headline 38% survives only in the recent window — 77% of all early-cue detections come from that not-yet-matured window — where truth cannot yet be confirmed. Verdict: NOT validated out-of-time. The tier stays gated; re-test after ~2026-07-23, when the live window's ±14 d truth settles.

Result — the same tier, measured on each cohort.

*Cohort B truth is right-censored (detections 0–7 days old vs the ±14 d IRWIN matching window), so 39.5% is a lower bound that is not yet confirmable, not a validated number.

What this shows. Entry 0011's 38.4%/19×/4.3 h is faithfully reproduced as the blended figure (31.1%/18.3×/4.8 h) — but it is supplied almost entirely by cohort B: 2,037 of 2,637 early-cue detections (77%) are in the recent, unsettled-truth window. On the only settled-truth data (cohort A) the tier is near baseline (2.5% vs 1.0%), and its 24.6 h median “lead” is far longer than a polar revisit — consistent with coincidental later polar hits, not genuine early warning. Cohort B's tight 3.7 h leads look like a real early-warning signal, but its truth is not settled. Polar availability is comparable across the two windows (11,268 vs 10,594 dedup polar cells), so cohort A's low precision is not a missing-polar artifact.

Why cameras miss fires they can see — and a lever that lifts detection-in-coverage 37→49% at flat precision

The framing. Entry 0008 split the camera gap in two: coverage (infrastructure — is a fire near an online camera at all) and detection-in-coverage (algorithm — where a camera can see, how often it catches the fire), the latter only ~5–25%. Coverage is a build-out problem; detection-in-coverage is the algorithm lever. This entry asks why the in-coverage misses happen and tests the best fix.

BLUF. The dominant miss driver is range, not darkness: within 15 km cameras catch 47% of in-coverage fires, falling to 20% at 15–30 km and 4% at 30–40 km (undetectable at sensor resolution). The assumed night floor does not exist — night recall (45%) actually beats day (31%), almost certainly flame glow. The best lever is the simplest: lower the single-frame smoke trigger 0.50→0.30, which lifts detection-in-coverage 36.9%→49.4% (+12.5 pts, +34% relative) at essentially flat precision (58.1%→57.2%), monotonic across the sweep and stable in 5-fold spatial CV (0.50±0.11 vs 0.37±0.11); at a tighter 20 km viewshed it nearly doubles recall (17.1%→34.2%). Adding temporal persistence was a mistake — it hurt recall (to 24.7%). We flag the threshold lever as a shadow-first build candidate, gated until it passes the temporal-transport test that killed the Dozier gate (entry 0012).

Why the misses happen (in-coverage, R=40 km, 1,083 fires, current 0.5 trigger).

Range dominates (a clean 12× swing, 47%→4%); day/night runs the “wrong” way (glow); size is noisy at tiny n and not a usable driver. Where cameras do catch a fire, they flag smoke a median 14.7 h before the IRWIN discovery timestamp (within the ±24 h match window) — cameras are an early corroborator when they fire.

The lever — single-frame threshold sweep.

The physical floor (achievable headroom). Two hard limits, neither fixable by thresholding: (1) range — 30–40 km fires are ~undetectable (3.9%); (2) among near (<15 km) misses, 39% have a best smoke-probability of ~0 (the model never fired at all — out-of-field-of-view / occlusion / too faint; camera azimuth+FOV are recorded but unused, so a fire behind a fixed camera is invisible regardless of threshold). Only ~26% of near misses sit in the 0.30–0.50 band the lever actually recovers. So the lever captures the genuinely-recoverable slice; it does not (and cannot) fix the geometry floor.

The camera-threshold lever survives out-of-time (where the Dozier gate died) — validated, and built shadow-first

The framing. Entry 0014 found that lowering the single-frame smoke trigger 0.50→0.30 lifts detection-in-coverage 36.9%→49.4% at flat precision. Before building, we applied the exact discipline that killed the Dozier gate (entry 0012): does it hold out-of-time, and does it generalize across camera networks?

BLUF. It holds. Trained on 2026-07-01–04 and tested on the strictly-later 07-06–08, the 0.30 trigger still gains +8.3 pts of detection-in-coverage (34.8%→43.1%) at ~flat precision (56.1%→54.0%). It generalizes across every network — recall rises on all four vendors with precision essentially unchanged; on the purpose-built fire networks it is dramatic (HPWREN 45%→91% recall at 79% precision). This is the opposite of the Dozier gate, which inverted out-of-time. We therefore built it — shadow-first and gated: a hidden canary logs the would-surface detections; production is untouched until an explicit flip.

Temporal transport (train 07-01–04 → test 07-06–08, gap 07-05).

Out-of-time gain +8.3 pts recall, precision −2.1 pts (within the flat band). Verdict: HOLDS.

Per-network generalization (full window).

The 0.30 trigger lifts recall on every vendor with precision flat. Precision is network-driven (purpose-built fire cams 78–81% vs webcams 40–56%), not threshold-driven — so 0.30 generalizes as a global trigger, though surfacing tiers should still be per-network.

The operating point we would flip to, and its gate. A global single-frame trigger of 0.30 (replacing the per-network 0.55–0.85). Not flipped: it is built as a hidden shadow canary (cam_smoke_lowthresh_shadow) that logs every ≥0.30 detection, flags which are newly recovered vs the current per-network production thresholds, and records IRWIN proximity for a precision audit. On this run: 2,926 detections ≥0.30, of which 2,506 are newly recovered vs production, and 1,394 of those sit near a real IRWIN fire. The table is surface_allowed=0 with a DO_NOT_SURFACE marker and no serving-app read path; the production cam_smoke_daemon.py is byte-for-byte unchanged. Flip requires Mark's go plus a live precision audit of the newly-recovered set.

Production cutover: the 0.30 camera trigger goes live for two validated fire-cam networks (precision held)

The framing. Entry 0015 validated the lower 0.30 smoke trigger out-of-time and showed it generalizes across camera networks — strongest on the purpose-built fire cams. With that evidence, we flipped it to production, but only for the two networks that earned it, and reversibly.

What shipped. The single-frame smoke trigger dropped to 0.30 for HPWREN (pre-flip 0.40) and AlertCalifornia (pre-flip 0.50) — the two validated wildland fire-cam networks. Every other network stays at its current production threshold (Caltrans 0.85, Windy/default 0.7–0.75, …) because their precision floor is network-limited and unproven. The 2-cam-5 nm corroboration gate on the two flipped networks is unchanged — it remains a precision safeguard on top of the lower trigger. This is now user-facing for HPWREN + AlertCalifornia.

Precision held on recent live data (the go/no-go check). Measured on 2026-07-06–09, the flip raises volume 2–3× without degrading precision:

Precision is flat-to-improving (HPWREN 72→76%, AlertCalifornia 86→85%) while the surfaced-detection count roughly triples. No precision crater — the flip stands.

Reversibility & audit (built in). Per-network thresholds now live in a single-source config (cam_gate_config.py) rather than scattered across code + systemd env-overrides (which we consolidated and removed). Reverting either network is a one-line change (set HPWREN back to 0.40 / AlertCalifornia to 0.50) plus a service restart. Every surfaced detection now records the exact gate_threshold and a gate_version (pernet_flip0016) in its evidence, so which gate produced any detection is auditable after the fact. The production file was backed up before the edit; a fail-safe falls back to conservative built-in thresholds if the config is ever unreadable.

Live-surfacing verified. CAM-SMOKE detections flow evidence → fire_events → serving API/map as before: 153 CAM-SMOKE-bearing fire events in the last 72 h, 138 of them passing the serving-app surfacing filter (CAM-SMOKE is in the allowed source set, unlike standalone Dozier/ADP). Nothing else changed; the still-gated networks keep flowing at their old thresholds.

Mapping the recall ceiling: 90% of the “missed” IRWIN fires are a liveness artifact, not blindness — and the fusion gate misses zero

The framing. Entry 0007 reported the fused OLDGABE output at 72.4% recall against NIFC IRWIN (daily-scorer run 18, 24 h window). The natural next question is which ~28% of fires we miss, and why — because the lever depends entirely on the answer. A fire we never saw needs a new sensor or a lower detection floor; a fire one sensor saw but the fusion gate rejected needs a looser gate; a fire we detected yesterday but did not re-image in the last 24 h needs neither — it needs liveness. We took the exact run-18 window, reproduced the miss set, and classified every false-negative against a 30-day lookback.

BLUF. The “miss” is overwhelmingly a liveness artifact of the 24 h scoring window, not detection blindness. Of the 1,407 IRWIN wildfires scored as false-negatives in the window, 1,275 (90.6%) had a promoted, confirmed OLDGABE event within 10 nm during the last 30 days — we detected and confirmed these fires, we simply did not re-detect them in that particular 24 h. They carry 98.9% of the missed acreage and 100 of the 106 missed fires ≥100 acres. Only 132 (9.4%) are truly sub-floor (nothing of any status within 10 nm in 30 days) — small (94/132 are <1 acre), remote, Alaska-heavy, just 3,909 acres and 6 big fires. The “single-sensor-seen-but-fusion-rejected” bucket — the obvious “loosen the gate” hypothesis — is exactly zero. The dominant recall lever is an active-fire liveness/persistence pass, not a looser fusion gate and not a lower detection floor.

Result (reproduced run-18 window 2026-07-07→2026-07-08, 10 nm / 24 h vs IRWIN; 3,410 active IRWIN wildfires; 1,407 fire-level false-negatives classified against a 30-day promoted-event lookback):

Put differently: OLDGABE’s true detection recall — did we ever confirm this fire during its active life — is (3,410−132)/3,410 = 96.1%. The 24 h liveness recall published in entry 0007 (72.4%) understates that by ~24 points, and it is volatile precisely because it is a liveness metric: the last eight daily runs read 72.4 / 92.3 / 94.2 / 98.0 / 95.8 / 92.1 / 82.8 / 72.4%. A metric that swings 26 points day-to-day while the fire roster barely changes is measuring re-detection cadence, not coverage.

Temporal-transport kills the “geo lift”: the recovery+geo re-ranker’s de-confounded win was mostly in-sample optimism — it stays gated

BLUF. Entries 0001/0002 built a re-ranker to recover real fires that production tiers as CAUTION, and reported that adding geography (lat/lon) gave a large lift even under a reporting-bias-free neutral polar truth (in-sample neutral-truth AUC 0.94 vs 0.86 det-only; +0.081). That was the whole reason geography looked defensible rather than a reporting-bias artifact. We put it to the strict test the same gate that stopped the Dozier multi-feature gate and the early-cue tier: train on an earlier window, test on a strictly-later window, score against the neutral (independent VIIRS/MODIS) truth. The geo lift collapses out-of-time — from +0.081 in-sample to +0.019 — and recovery+geo no longer beats production on the neutral truth. The de-confounded advantage was substantially in-sample optimism. The canary stays gated.

The split. 214,614 CAUTION-tier events. Train on first_det in 06-27–07-02 (123,931 events), a 07-03 gap, test strictly-later 07-04–07-06 (38,549 events). Models train on the honest (IRWIN/perimeter/confirmed) truth; every scorer is then evaluated out-of-time against the NEUTRAL polar truth — an independent VIIRS/MODIS detection within 5 km in a subsequent overpass window (t0+3 h … t0+72 h), which is independent of HMS and of who-reports-what. Neutral-truth base rate on the holdout is only 3.3% (1,274 positives).

Out-of-time on the NEUTRAL (reporting-bias-free) truth — the gate.

Recovery+geo (0.856) does not beat production (0.861) on the neutral truth out-of-time, and its PR-AUC (0.177) is below det-only (0.201) — adding geography hurt ranking precision out-of-time. Every scorer, production included, scores 0 at the 80/90% precision operating points on this sparse 3.3%-base truth, so the comparison rests on AUC/PR-AUC discrimination, where the re-ranker has no out-of-time edge. Pure lat/lon still reaches AUC 0.723 alone — a geographic-memorization warning, not a skill signal.

For contrast — the same holdout scored against the HONEST (HMS-inclusive) truth. Here the re-ranker still looks excellent — which is exactly the trap the neutral truth exists to expose:

The de-confound, made explicit. Geography adds +0.077 AUC under the honest truth but only +0.019 AUC (and −0.024 PR-AUC) under the reporting-bias-free neutral truth, out-of-time. The gap between those two numbers is the reporting-bias artifact: most of what geography “knew” was where fires get reported (populated, IRWIN/HMS-dense geography [6]), not where independent polar sensors confirm them. Out-of-time, on truth that doesn’t care who reports, that knowledge evaporates.

An active-fire liveness pass for roster continuity — built shadow-first, and the two-condition gate throws out 99% of what a naive “keep-alive” would resurrect

BLUF. Entry 0017 established that 90.6% of “missed” IRWIN fires are a 24 h scoring-window liveness artifact: we detected and confirmed them, we simply did not re-image them inside the 24 h window. True detection recall is ~96.1% and the fusion gate misses zero, so the lever here is not new detection — it is operational roster continuity: keeping a known-active fire visible on the map between satellite passes instead of letting it blink off a naive 24 h-cadence roster. Because the failure mode of continuity is keeping a dead fire alive, the primary risk metric is added-FP, not recall. We built the pass as a hidden shadow canary and measured the tradeoff. The two-condition gate keeps 439 fires alive and refuses 42,286 — 99.0% of what a naive keep-everything pass would have resurrected.

The gate (both conditions required). For each confirmed OLDGABE fire, keep it “active” only when (a) a matching IRWIN incident within 10 nm is still listed active (no past FireOutDateTime) AND (b) a core-sensor touch (VIIRS/MODIS/GOES/HMS) landed within 10 nm inside the current liveness window (72 h). Condition (a) drops fires the incident record has declared out; condition (b) drops stale fires with no recent physical signal. Together they are the FP-safeguard.

The tradeoff on the current window (clock 2026-07-09 09:01 UTC; 44,553 confirmed fires, of which 1,828 were re-imaged within 24 h and 42,725 would blink off a naive 24 h roster):

What the gate filters, and why (the FP-safeguard, itemized).

Condition (b) is the load-bearing safeguard. Of the 7,044 not-re-seen fires that IRWIN still lists active, only 439 (6.2%) had a core-sensor touch inside the 72 h window; the other 6,605 (93.8%) were stale and correctly refused. This matters because IRWIN’s FireOutDateTime is sparsely and laggingly populated (zero of the population fires carried a past FireOut), so “still active” on its own is a weak signal — the recent-sensor-touch condition is what actually protects against IRWIN’s optimistic/lagging active listings. Among the 439 kept, none is IRWIN-declared-out and none is stale (0 measurable added-FP by construction).

The symmetric question to recall: of the fires we actually SHOW users, how many are false? — a precision / false-positive audit of the surfaced map

The framing. The program has spent most of its effort on the recall side — entry 0017 put true detection recall near 96.1% and showed the fusion gate misses zero real fires. The mirror-image question had never been measured: of the detections OLDGABE actually surfaces to users on the map and public API, what fraction are false positives — not real fires — and is there a clean lever that removes them without throwing away real fires?

BLUF. Measured against strict independent operational truth (NIFC IRWIN wildfire incidents ∪ WFIGS, matched within 10 nm and the incident’s active window), the surfaced map is 23.7% precise — but that number is a floor, not the false-positive rate. IRWIN lists only WF/WFU incidents; it does not list the large real population of sub-incident thermal (agricultural burns, pile burns, small fires) that a single VIIRS or MODIS pass correctly detects. When independent multi-sensor corroboration is added as a second truth, surfaced precision rises to 62.5%. The classical false-positive machinery already deployed (an OSM industrial/flare catalog, a persistence mask, a 0.10 confidence floor, and suppression of the noisy geostationary-only tier) has already removed the cleanly-separable FP: only 1 of 1,985 surfaced detections sits on the static FP catalog. And no operable threshold lever cuts the residual without an equal-or-greater recall cost — requiring corroboration on the probable tier removes 50% of the real IRWIN/WFIGS fires for zero precision gain. The one clean, defensible lever is small and geographic: a VIIRS-nightfire / oil-and-gas flaring-basin mask extension removes ~51 surfaced FP (2.6% of the map) at ~0 recall cost. We recommend it shadow-first — nothing here changes a shipped number.

The surfaced set (read from the serving code, not guessed). The live map/API path is /api/oldgabe/fires with OLDGABE_FUSED_MAP=1 (confirmed in the running process environment). It reads fused_fires with exactly this filter: last_seen ≥ now−24h AND tier != 'fp' AND confidence ≥ 0.10. That 0.10 floor is load-bearing — it drops the entire early tier (geostationary-only, conf 0.05) and every frp = 0 geostationary-persistent event (conf 0.08), i.e. the noisiest candidates never reach a user. Over the trailing 24 h window (clock 2026-07-09 ~09:30 UTC) the surfaced set is 1,985 fires: 458 confirmed, 1,527 probable.

Precision by tier — strict (operational) vs broad (+ independent multi-sensor corroboration).

The confirmed-tier broad figure (98.0%) is near-tautological — the confirmed tier is defined by ≥2 core sensors, so “is it multi-family corroborated” is almost the tier’s own definition. Its informative precision is the independent IRWIN/WFIGS match: 29.0%. The gap between 29.0% and 98.0% is the measure of how much real fire IRWIN simply does not carry as a listed incident.

Precision by source family (strict IRWIN∪WFIGS). The ranking is the opposite of sensor quality — the sharpest single-pass thermal sensors score lowest, because they correctly see real thermal that is not a NIFC wildfire incident:

What drives the false positives. The static FP catalog is not the answer — it has already done its job: 1 of 1,985 surfaced detections falls on an fp_polygons cell, because fusion applies the catalog before anything is written to the map. The residual imprecision lives in the 745 uncorroborated single-source detections (37.5% of the map: VIIRS-only 345, HMS-only 220, GOES-only 104, MODIS-only 67, and 9 single-source frp≥20 promotions). Crucially, these are not a clean FP bucket: their IRWIN-match rate (~24%) is the same as the map overall, i.e. “uncorroborated” does not predict “false.” The one clearly-false, geographically-concentrated driver we can name is oil-and-gas flaring: 51 uncorroborated surfaced detections sit inside known US flaring basins (33 VIIRS in the Permian, plus Eagle Ford, Anadarko, DJ-Niobrara), with representative low-FRP VIIRS hits at 31.886,−102.511, 32.723,−101.925, 32.188,−102.008 (Permian) — classic gas flares the OSM polygon set did not fully cover. Only 2 real IRWIN/WFIGS fires fall inside these basins at all.

The levers — measured precision gain vs recall cost, on neutral IRWIN∪WFIGS truth. Baseline surfaced real = 470 of 1,985 (23.7%). “Recall cost” = share of those 470 real fires a lever would drop:

The corroboration gate is a measured negative. L1 — the intuitive “require a second sensor” fix — lowers precision (23.7% → 23.0%) while discarding half the real fires, because the dropped single-family detections are exactly as real as the map average. FRP and source gates (L3–L5) buy only ~+4 pp of strict precision at a 13–27% recall cost — a poor trade for a recall-optimized product. The only lever that removes false positives at genuinely ~0 recall cost is the flaring-basin mask (L7): the 51 it targets are almost entirely industrial flares, and the 2 real in-basin fires are IRWIN/WFIGS-matched, so a truth-sparing mask never touches them. Its ceiling is small — about +2.6 pp of the surfaced map — but it is clean.

Executing the flaring lever: a VIIRS-nightfire persistent-source FP mask that passes an out-of-time transport gate — shadow-first

The framing. Entry 0020 audited the surfaced map and found exactly one clean, geographically-concentrated false-positive driver — oil-and-gas gas flaring — and set a hard precondition before it could ever change the map: build the fix as a persistent-source mask (not a per-detection threshold), and require it to pass a temporal-transport gate — construct the mask from an earlier window and prove it suppresses flare FP in a strictly later window without masking any real fire. This entry executes that build and reports the gate result.

BLUF. The mask is derived only from detections strictly before a 2026-07-03 cutoff: any fixed ~2 km cell whose VIIRS/MODIS detections recur across many distinct days (a persistence signature), fire-protection-guarded so a cell is dropped if it lies near any IRWIN incident (ever), near HMS smoke, near a WFIGS wildfire perimeter, or carries appreciable FRP. At the operating threshold (≥5 distinct days, matching the deployed polar-persistence daemon) this yields 660 mask cells, and they are what a flare inventory should be: 89.6% of their detections are at night (the VIIRS-Nightfire flaring signature) and 79 sit inside the Permian basin — the Permian concentration falls out of the persistence derivation, it was never coded in. On a strictly-later held-out window (2026-07-06 onward, with a 3-day gap so no source straddles the split) the mask suppresses 283 surfaced false-positive detections (8.3% of the 3,414-fire out-of-time surfaced set) while masking exactly 0 real IRWIN wildfire incidents and 0 WFIGS perimeters — at the 2 km suppression radius and at a conservative 5 km. All 283 suppressed detections are VIIRS-heavy and uncorroborated by IRWIN. The transport gate PASSES. The mask is written as a hidden shadow table (flaring_mask_shadow, surface_allowed=0); nothing here changes a shipped number.

Why the existing persistence machinery misses these. The deployed dynamic-persistence daemon suppresses a recurring cell only when it has zero polar-sensor hits — a single VIIRS pass is treated as confirmation of a real fire and keeps the cell live. Gas flares are precisely VIIRS-heavy, so they slip through that gate. This mask is the additive complement: it keys on VIIRS/MODIS multi-day recurrence at a fixed location with no incident and no smoke, which is the affirmative flare signature rather than the absence-of-polar signature. It is built as an extension of the same persistence machinery (same 0.02° grid, same 2 km mask radius, same idempotent-replace discipline), not a parallel system.

The transport gate — mask trained before 2026-07-03, evaluated strictly after 2026-07-06. “FP suppressed” = surfaced fused fires (the live serving filter: tier!='fp' AND confidence≥0.10) that fall inside the mask out-of-time; “real masked” = independent IRWIN wildfire fire-days / WFIGS perimeters in the later window that fall inside the mask — the number that must be zero:

The truth-sparing column is 0 at every threshold, at both radii — the more persistent the source, the more confident the flare call, and never once does the mask shadow a later real fire. The out-of-time test set carried 547 independent IRWIN wildfire fire-days and 19 WFIGS perimeters; the mask touched none of them. The fire-protection guards did real work: at the operating point they released 80 candidate cells for lying near an IRWIN incident, 193 near HMS smoke, and 19 for FRP ≥ 40 MW — those are the cells a naive persistence rule would have wrongly masked, and the guards remove them before the gate, which is why the gate comes back clean.

Replacing the 23.7% IRWIN floor with a measured number: a fire-vs-non-fire label corpus that decomposes surfaced precision into unlisted-real fire and genuine false positive

The framing. Entry 0020 left the program with a number it explicitly could not use: surfaced precision against IRWIN reads 23.7%, but that is a floor, not the false-positive rate — IRWIN matched the corroborated and uncorroborated surfaced buckets at the same ~24%, proving IRWIN cannot tell “not a fire” apart from “a real fire it never listed.” So the one number a detection program most needs about the map it ships — the fraction of surfaced detections that are genuinely not fire — was unmeasured. This entry measures it, by abandoning IRWIN-presence as the label and building an affirmative fire-vs-non-fire corpus over the entire surfaced set: positive evidence of fire, positive evidence of non-fire, and an explicitly adjudicated middle in between.

BLUF. Over the full trailing-window surfaced set (N = 1,985 fused fires; the live serving filter last_seen≥now−24h AND tier≠'fp' AND confidence≥0.10), positive evidence partitions the set into 1,363 positive-fire, 208 positive-non-fire, and a 414-detection unadjudicated middle; adjudicating that middle with independent heuristic evidence splits it 81 fire / 35 non-fire / 298 genuinely-ambiguous. Combining the buckets, the true surfaced false-positive rate is 19.7% (point estimate; Wilson-95 sampling interval [18.1%, 21.6%]; the wider [12.2%, 27.3%] epistemic bracket carries the ambiguous band honestly). It is strongly tier-stratified: the confirmed tier is 4.6% FP [3.0%, 6.9%] while the probable tier is 24.3% [22.2%, 26.5%]. It is almost entirely a single-family-VIIRS phenomenon: VIIRS-only detections are 49.4% FP, and every multi-family or geostationary combination is ≤16%. The headline the program had been missing — the decomposition of entry 0020’s 76.3-point precision gap: 56.6 points are unlisted-real fire, only 19.7 points are genuine false positive. Roughly three-quarters of the apparent “imprecision” was IRWIN incompleteness, not surfaced error. Fully read-only; no shipped number changes.

How the labels are assigned — positive evidence, not IRWIN-presence. Each surfaced fire is labelled by a precedence stack of affirmative signals, all independent of the confidence/tier that surfaced it: (1) positive fire if it matches a NIFC IRWIN/WFIGS wildfire incident (10 nm, active-window, the entry-0020 methodology); else (2) positive non-fire if it sits on the persistent-flaring mask (entry 0021’s flaring_mask_shadow) or a catalogued static industrial/persistence source (fp_polygons: OSM refineries, OSM gas flares, dozier/polar persistence); else (3) positive fire if ≥2 independent sensor families corroborate — where independence counts distinct physical platforms {VIIRS, MODIS, GOES-ABI, S2, S1} and deliberately excludes HMS, since NOAA’s Hazard Mapping System is an analyst product derived from those same satellites and cannot be a second independent vote; else (4) positive fire if VIIRS/MODIS detections recur across ≥2 distinct days at the location and it is not on any static-source catalog (multi-overpass persistence away from flares); else (5) unadjudicated middle.

Adjudicating the middle. The 414-detection middle is the honest hard part — neither positively fire nor positively non-fire. Each is adjudicated with the best independent evidence available short of a ground visit: proximity to (but outside the strict radius of) a catalogued industrial/persistence source → non-fire-leaning; multi-day polar persistence away from all catalogs, or an energetic single-day polar detection (FRP ≥ 10 MW) → fire-leaning; no polar corroboration and weak signal (FRP < 5 MW, i.e. an ephemeral geostationary/analyst-only hit) → non-fire-leaning; everything else → genuinely ambiguous. Result: 81 fire, 35 non-fire, 298 ambiguous. The large ambiguous residual is not hidden — it is carried straight into the interval below.

The estimate. With B = 208 positive-non-fire and the middle’s 35 non-fire + 298 ambiguous, the FP-count point estimate splits the ambiguous band 50/50: FP = B + 35 + 0.5·298 = 392 of 1,985. The epistemic bracket sends the ambiguous band entirely to fire (low) or entirely to non-fire (high); the Wilson interval is the binomial sampling error on the point count.

The FP is a single-sensor-family artifact. Split by the surfacing sensor mix, the false-positive burden is concentrated almost entirely in one place — lone-VIIRS detections, which is exactly the flaring class entry 0021 characterized. Every combination that adds a second independent family, or a geostationary confirmation, collapses to single digits or low teens:

The decomposition entry 0020 could not do. Entry 0020’s 23.7% IRWIN “precision” left a 76.3-point gap it could only call “unmeasured — some real, some false.” The corpus splits that gap for the first time:

So the “76.3% of surfaced detections IRWIN could not confirm” was ~74% unlisted-real fire and ~26% genuine false positive — the map is far more correct than the IRWIN floor implied, and its residual error is a nameable, already-mitigable class (persistent flaring, concentrated in lone-VIIRS detections).

Physical geo-priors don’t survive transport either: swapping memorized lat/lon for elevation, slope and ruggedness still fails to beat the production baseline out-of-time

BLUF. Entry 0018 killed the recovery re-ranker’s “geo lift” under temporal transport — raw lat/lon was memorizing the reporting surface, not fire geography — and named the constructive next step: replace raw coordinates with physical, non-memorizable geo-priors (terrain, fuel, climatology) and re-run the same gate; only promote if the physical prior holds out-of-time. We did exactly that. We engineered physical terrain covariates — elevation, slope, and terrain-ruggedness sampled from a coarse DEM — and re-ran the identical train-early / test-strictly-later gate against the reporting-bias-free neutral (independent VIIRS/MODIS) truth. Physical geo-priors do not survive transport either. On the neutral truth out-of-time, recovery+physical-geo reaches AUC 0.571 — nowhere near the production baseline (0.856), only marginally above recovery det-only (0.534, +0.037 AUC / ±0.00 PR-AUC), and below even the raw-lat/lon variant it was meant to replace (0.633). Fuel and climatology remain deferred (no box-local source), so this is an elevation/terrain-only physical test — but terrain is decisive on its own hypothesis, and the answer is no. The canary stays gated.

The split (identical to entry 0018). The same CAUTION candidate set, the same 19 detection-only base features, the same honest training label, the same neutral evaluation label, the same model. Train on first_det in 06-27–07-02 (123,931 events), a 07-03 gap, test strictly-later 07-04–07-06 (38,549 events). Models train on the honest (IRWIN/perimeter/confirmed) truth and are evaluated out-of-time against the NEUTRAL truth — an independent VIIRS/MODIS detection (belonging to a different canonical event) within 5 km in a subsequent overpass (t0+3 h…t0+72 h), gridded 0.05°×6 h over CONUS, HMS/IRWIN-independent. Neutral base rate on the holdout is 3.35% (1,292 positives) — reproducing entry 0018’s 3.3%. The only change vs entry 0018 is the geography channel: physical terrain covariates instead of raw coordinates.

Out-of-time on the NEUTRAL (reporting-bias-free) truth — the gate.

Physical geography fails the gate on both counts. Recovery+physical-geo (0.571) does not beat production (0.856) — it is 0.28 AUC and 0.16 PR-AUC below it — and it does not meaningfully beat recovery det-only (0.534): a +0.037 AUC nudge with PR-AUC flat (0.099 vs 0.100). It is also below the raw-lat/lon variant (0.633), so the physical prior we hoped would be more robust than memorized coordinates actually carries less out-of-time neutral signal than the coordinates. Physical-geo alone reaches only 0.583 — a weak geographic prior, well short of production. Every scorer scores 0 at the 80/90% precision operating points on this sparse 3.35%-base truth, so the comparison rests on AUC/PR-AUC discrimination, where physical geography adds no out-of-time edge. The single scorer that transfers is production — a plain FRP/family rule that tracks fire radiative power, the same physical quantity the polar sensors re-detect.

For contrast — the same holdout scored against the HONEST (HMS-inclusive) truth. Here, as in entry 0018, geography still looks attractive — the trap the neutral truth exists to expose:

The physical prior behaves exactly as designed — which is why it can’t rescue recovery. Raw lat/lon adds +0.068 AUC under the honest truth: a large lift, most of which is the reporting-bias artifact (it memorizes where fires get reported [6]). Physical terrain adds only +0.032 AUC under the honest truth — less than half the raw-coordinate lift — because it cannot memorize the reporting surface; it only knows physical terrain. That is the good news the constructive next step was chasing. But the same property means it brings almost no signal at all: its neutral-truth increment over det-only (+0.037) is as small as its honest-truth increment, and the absolute neutral AUC (0.571) sits far below the production baseline. Physical geography here is honest but inert: it does not encode the reporting surface, and it does not encode enough real fire-geography to lift a reporting-bias-free re-ranker out of the near-chance regime.

Bounding the last coverage lever: the sub-floor tail is real but low-yield — SWIR revisit is not the limiter, sub-pixel size is, and the recoverable upper bound is 48 fires (mostly larger remote Alaska burns)

BLUF. Entry 0017 showed that 90.6% of OLDGABE’s “missed” IRWIN fires are a 24 h liveness artifact, that the fusion gate rejects zero, and that only a small tail — 132 fires (9.4%) in that run — are truly sub-floor (nothing of any status within 10 nm in a 30-day lookback). It named the last open coverage thread: quantify that sub-floor tail against Landsat-9 / Sentinel-2 SWIR revisit cadence over remote Alaska to bound the only genuine coverage lever. We reproduced the sub-floor set (161 fires this re-run, up from 132 on the documented active-window drift) and bounded it. The revisit geometry is not the constraint — sub-pixel fire size is. Every one of the 161 is an open IRWIN record (null FireOutDateTime), so its active window runs discovery→now — days to weeks — and at Alaska-summer latitudes with S2A/S2B + L8/L9 sidelap each such window contains many daytime SWIR overpasses. Overpass availability is therefore not limiting (plausible-overpass upper bound = 161/161). What limits it is that 112 of 161 (69.6%) are sub-1-acre point records — at or below the 20–30 m S2/L9 SWIR sub-pixel floor no matter how often they are overflown — carrying just 13.8 acres between them. Gate detectability on a generous “≥1 acre sustained flaming” SWIR floor and the recoverable upper bound is 48 fires (29.8%), and because the acreage is so top-heavy those 48 carry 99.7% of the sub-floor acreage (4,626 of 4,640 ac). But that 48 is a hard upper bound before any cloud/illumination haircut; the robustly-sized core is just 24 fires ≥10 acres and 7 ≥100 acres, overwhelmingly larger remote Alaska burns. Against a roster already at ~95% true detection recall, a SWIR-cadence recovery path is a real but marginal Alaska-only, large-fire-only lever — exactly as entry 0017 suspected. It does not warrant a general recovery path.

The reproduced sub-floor set (run-18-equivalent window 2026-07-07→2026-07-08, 10 nm vs IRWIN WF/WFU; 3,446 active wildfires; 1,447 fire-level false-negatives; 1,285 liveness; 161 sub-floor — nothing of any status within 10 nm in the 30-day lookback; 4,640 total acres):

Geography — Alaska-heavy in acreage, CONUS-heavy in count. 34 of 161 fires (21.1%) are in Alaska but they carry 3,341 of 4,640 acres (72.0%); the other 127 (CONUS) are overwhelmingly the sub-1-acre point records (1,299 ac, 28.0%). By latitude: 83 below 45°N, 44 at 45–55°N, 34 at ≥60°N (all Alaska). Every one of the 161 is a genuine wildfire (IncidentTypeCategory=WF, kind FI) — no prescribed burns dilute the set. Discovery is 124 daytime / 37 night. Crucially, none of the 161 carries a FireOutDateTime: they are all open records, so the “is this a mop-up record with FireOut≈FireDiscovery” test entry 0017 anticipated is inapplicable here — the mop-up/record-only class instead shows up as the sub-1-acre point-record mass, which is the same orbit-undetectable population by a different signature.

The SWIR recoverability bound (Sentinel-2 MSI B11/B12 ~1.6/2.2 µm at 20 m; Landsat-9 OLI-2 B6/B7 ~1.6/2.2 µm at 30 m; combined S2A/S2B + L8/L9 daytime SWIR opportunity; upper bounds throughout):

Reading the bound honestly. The recoverable upper bound is 48 fires / 4,626 acres — and it looks acreage-rich only because seven ≥100-acre Alaska burns carry 88% of it. Strip to what a 20–30 m SWIR band would robustly resolve and the core is 24 fires ≥10 acres, mostly larger remote Alaska burns that no core polar/geostationary sensor happened to touch inside the window. Even that 24 is optimistic: it assumes a cloud-free daytime overpass coincides with active flaming, and Alaska/PNW summer cloud frequency, the sub-pixel marginality of the 1–10 acre band, and the requirement that any SWIR hit still clear the fusion gate and attribute correctly all push the realistic yield below the bound. For context, OLDGABE’s fusion stack already lists an S2-SWIR-Hotspot source (alongside S1-SAR-BurnScar); these 161 slipped past it, so the lever is really about densifying SWIR cadence (adding L9/L8 to the existing S2 feed, or reprocessing) rather than a missing sensor — which caps the upside at the 48-fire bound above.

Verdict — confirmed low-yield. The sub-floor tail is genuinely small: 161 of 3,446 active wildfires (4.7%) and 4,640 acres, against a roster whose true detection recall (ever confirmed during active life) is (3,446−161)/3,446 = 95.3%. Of that tail, 70% is orbit-undetectable sub-1-acre point records, the revisit geometry is favorable rather than limiting, and the recoverable upper bound is 48 fires — realistically a handful of larger Alaska burns. The coverage lever entry 0017 flagged is real but low-yield: it justifies at most a narrow, default-OFF, Alaska-focused, ≥10-acre SWIR-cadence shadow probe measured for recovered-recall vs added-FP — not a general recovery path, and a far smaller prize than the liveness/persistence pass (1,285 fires) that entry 0017 already identified as the dominant lever.

How fast does OLDGABE surface a real fire? — the first end-to-end detection-latency measurement, and why the “we detect it before it’s reported” number is mostly a lookback artifact

The framing. Twenty-four entries have measured OLDGABE’s recall, precision/FP, and coverage — but never latency: once a real fire starts, how quickly does OLDGABE actually surface it? A firefighter cares about speed as much as about whether a fire is found at all. This is the first entry to measure time-to-first-detection end-to-end, and the honest answer comes with a large caveat that turns out to be the finding.

BLUF. Against IRWIN’s reported discovery time as the t₀ proxy, OLDGABE’s defensible, artifact-free latency — its first independent post-discovery detection within 10 nm — has a median of +9.2 h, with 19.8% surfaced within 1 h, 40.4% within 6 h, and 68.9% within 24 h of the reported discovery (n = 2,932 fires). The tempting headline — “OLDGABE detects 71% of fires before they are reported (median −25 h)” — is not a defensible detection-skill claim: the median slides monotonically with the pre-discovery lookback window we allow (−25 h at a 48 h cap → −7 h at 24 h → +3.5 h at 6 h → +9.2 h at 0 h), which is the exact signature of a persistent-pixel spatial-association artifact — a 10 nm cell in fire country carries near-continuous GOES/HMS pixels that are not this ignition. On the fastest-path question: GOES-ABI wins the first-arrival race (it is the first sensor with any pixel in-cell for 76.2% of fires, on 5–15 min cadence), but its genuine post-report detection of a new fire is the slowest of the thermal sensors (median +13.8 h) because of coarse sensitivity; the polar/analyst products deliver tighter genuine confirmation (VIIRS +5.8 h, HMS +0.6 h). Cadence wins the first-pixel race; sensitivity wins the genuine-detection race.

Populations & t₀ (reproducible). Truth reuses the shipped scorer’s methodology (daemons/te_daily_scorer.py): IRWIN fires = fire_event_evidence where source='IRWIN', deduped on IrwinID/UniqueFireIdentifier, filtered to IncidentTypeCategory∈{WF,WFU}, 10 nm haversine match. t₀ = FireDiscoveryDateTime — explicitly a reported, human-entered discovery time and an ignition proxy, not true ignition; it can lag real ignition by hours, and OLDGABE can legitimately detect before it. OLDGABE’s own sensor evidence in this DB begins 2026-05-29, so the clean latency set is the 3,163 WF/WFU fires discovered inside coverage (2026-05-30→2026-07-09) — fires discovered earlier are excluded to avoid coverage-start censoring. Latency = (first OLDGABE evidence obs time for a promoted event within 10 nm) − FireDiscoveryDateTime, signed; negative = OLDGABE-first. Promoted = event status∈{confirmed,confirmed_growth,awaiting-confirmation,pre-ignition}.

Timezone verification (get this wrong and the whole metric is wrong). The host runs America/New_York (EDT, −04:00). detected_at is stored ISO-8601 with a Z suffix and the provider raw.detected field matches it verbatim — i.e. UTC; FireDiscoveryDateTime is epoch-milliseconds UTC. This analysis converts every timestamp with calendar.timegm (pure UTC), not the shipped scorer’s time.mktime (which interprets the string in host-local time). We verified the hazard numerically: for the same wall-clock string, timegm and mktime differ by exactly 4.0 h (the EDT offset) — so the local-time path would have skewed every latency by +4 h. Both fields are handled in UTC throughout.

Result 1 — the lookback-sensitivity table (this is the finding). Same fires, same 10 nm match, only the pre-discovery lookback cap varies. If the negative latencies were genuine early detections of the same ignition they would sit at stable, physically-plausible offsets; instead the median tracks the cap and the negative tail piles against whatever floor is imposed — the tell of spatial association, not detection:

At a 48 h cap the negative tail jams against the floor (p10 −45.9 h, p25 −41.8 h) — manufactured, not earned. Within a physically plausible ≤6 h pre-report window, ~36% of fires do surface before the human report, a real and useful signal but itself a mix of genuine early detection and residual association; we do not claim the ≥24 h figures.

Result 2 — artifact-free latency distribution (0 h floor, post-discovery).

Result 3 — by first-detecting sensor: cadence vs sensitivity, in real latency terms. First-arrival share is unambiguous (who has the earliest in-cell pixel); genuine post-report latency is the artifact-free per-family view (0 h floor):

Result 4 — by fire size (final acres) and region (0 h floor). Bigger fires are not systematically detected sooner: they are caught within the first hour more often (≥100 ac: 32.1% ≤1 h vs 16.0% for <1 ac — they are bright) but carry a longer median (≥100 ac +31.4 h; <1 ac +12.5 h; 10–100 ac +3.2 h), a reporting-promptness confound — large fires are reported by humans very fast, so OLDGABE’s independent detection more often trails a prompt report. Geographically, Alaska (+11.7 h) and CONUS (+9.2 h) are similar, but the West (+8.0 h, 72.9% ≤24 h) is far faster than the East (+42.4 h, 42.7% ≤24 h) — eastern fires are smaller, promptly human-reported, and often under canopy.

Survivorship pairing (do not read this as “OLDGABE finds everything in 9 h”). The latency set is conditioned on fires OLDGABE actually surfaced: of the 3,163 in-window fires, 2,988 (94.5%) had a promoted event within 10 nm at some point in their active window (2,932 with a post-discovery detection). That 94.5% is a generous ever-detected figure — any promoted evidence anywhere in the active window — and sits alongside, not in place of, the stricter operational recall and the sub-floor coverage tail already quantified in entries 0003 / 0007 / 0017 / 0024. Latency is measured on the fires we find; it says nothing about the ones we miss.

Verdict. The honest headline is the artifact-free one: median ~9 h to independent post-discovery detection, ~20% within an hour, ~40% within 6 h, ~69% within a day. The “detects 71% of fires before they are reported” framing is a lookback artifact and is retired here. The fastest first-alert path is a genuine cadence-vs-sensitivity split: GOES-ABI arrives first for three of four fires on cadence alone, but for a new small fire the sensitive polar overpass (VIIRS) and the analyst product (HMS) deliver the tighter genuine confirmation. Latency lever (research observation, not a flip): the data does not support a naive “GOES-first low-latency tier” — GOES owns first-arrival largely through cadence and persistent in-cell pixels, and its genuine new-fire detection is the slowest of the thermal sensors. The observations worth a default-OFF shadow probe are (1) an early-latency tier that pairs GOES cadence for breadth with a VIIRS-overpass / HMS weighting for genuine new-fire confirmation, measured for time-saved vs added-FP; and (2) a camera-first low-latency tier where camera coverage exists (CAM-SMOKE: +6.4 h median, 83% within 12 h), consistent with — and bounded by — the visibility limit of entry 0006. Neither changes a shipped number.

The liveness lever, backtested to a cutover decision — 38 day-slices say it buys ~7.6 recall points at ~20% precision, and the honest recommendation is NO-GO for surfacing

BLUF. Entry 0019 built the two-condition active-fire liveness gate as a hidden shadow and promised a ≥14-run forward validation with a recovered-continuity-vs-added-FP curve before any cutover. Rather than wait, we backtested the gate retrospectively across 38 daily slices (2026-06-01→2026-07-08), reconstructing for each day what a plain 24 h roster would surface vs. what production+liveness would surface, scored against a retrospective IRWIN active-roster truth. The result is decisive and it is a NO-GO. At the current operating point (10 nm, 72 h, both conditions) the lever recovers a real but modest +7.6 recall points (95% CI +4.8 to +10.4) — lifting the day-of roster recall from 23.7% to 31.3% — but it does so at a lever precision of only ~19.7%: it keeps roughly four IRWIN-uncorroborated fires for every one genuinely-active fire it rescues. The reason is structural and was hiding in the feed: IRWIN publishes no FireOutDateTime at all (0 of 992,511 rows), so condition (a) — “IRWIN still lists it active” — can never retire a dead fire, and the entire FP-safeguard rests on condition (b), the recent-sensor-touch window. Condition (b) is confirmed strongly load-bearing (dropping it inflates added-FP ~10×), but even with it the surfaced-roster precision caps near 20%. A continuity tier at 20% precision would inflate a firefighter-facing map ~5×; we do not flip it.

The backtest (retrospective, using full future knowledge). The single-snapshot daemon of entry 0019 called the one residual added-FP — an IRWIN-active fire that is actually out — “unmeasurable in one run.” The backtest measures exactly it. Because the feed never carries a FireOut timestamp, the real retrospective “out” signal is that IRWIN stops refreshing an incident once it is over. For each WF/WFU incident we take its full ingest span and define truth-active(T) := discovery−24 h ≤ T ≤ last-ingest+24 h (the 24 h pad mirrors the shipped scorer te_daily_scorer.py). At each slice T we reconstruct the deployed gate as-of T — keep a confirmed fire when (a) a WF/WFU IRWIN incident is present within 10 nm with first-ingest≤T (its deployed “still active” semantics, since no FireOut ever arrives) and (b) a core-sensor touch (VIIRS/MODIS/GOES/HMS) landed within the liveness window — and score each kept fire against truth-active(T). A confirmed OLDGABE fire is single-day in this DB (a multi-day fire spawns one confirmed event per day at the same location), so the lever’s reach is precisely the 24→72 h gap after a fire’s last imaging.

Operating point — recovered-recall vs added-FP, mean ± 95% CI over 38 slices (10 nm, 72 h, both conditions):

Pooled across the 38 slices the gate keeps 1,321 genuinely-active fires and 5,395 added-FPs. So the lever’s recall gain is real and its confidence interval clears zero comfortably — but ~80% of what it keeps is not corroborated as active by IRWIN at that moment. That is the FireOut-lag cost the earlier snapshot could not see, now quantified.

Sensitivity — and the proof that condition (b) is load-bearing. Same 38 slices; one knob varied at a time. Added-FP is the mean per-slice cost:

Reading the table: the 24 h window recovers nothing — by construction, a 24 h liveness window is the production roster, so the lever only ever acts in the 24→72 h gap. Widening the window buys recall almost linearly (48→72→96→168 h) at flat ~17–20% precision and steeply rising FP, so there is no window that makes this cheap; 48 h is the conservative knee (half the FP for two-thirds of the gain). Tightening the radius to 5 nm lowers precision (15.4%), so 10 nm is not the problem. The decisive row is the last one: dropping condition (b) multiplies added-FP roughly ten-fold (~142→~1,372 per slice; pooled 5,395→52,150) and drags precision to 14.3% — condition (b) removes ~90% of the FireOut-lag FP mass. This confirms entry 0019’s claim that the recent-sensor-touch condition, not the IRWIN listing, is what protects the roster — because IRWIN’s listing never expires in this feed.

The cutover PROPOSAL (mechanism described; recommendation is NO-GO — nothing is flipped). Were this ever promoted, the minimal reversible mechanism is: the serving query today is last_seen ≥ now−24h AND tier ≠ 'fp' AND confidence ≥ 0.10 over fused_fires. Add a default-OFF flag OLDGABE_LIVENESS_SURFACE (0 by default → query byte-for-byte unchanged). When set to 1, the time predicate becomes (last_seen ≥ now−24h) OR (fused_id ∈ keep_active set from the gated table, computed within the last 6 h), with those rows tagged a distinct tier='continuity' so the existing precision predicate (tier≠'fp', confidence≥0.10) and the aging logic still apply, the map/legend can style them apart from fresh confirmed detections, and they can never outlive the liveness window (aging still expires them at ≤72 h). The change is additive: no schema rollback, no data migration. The one-line reversal is OLDGABE_LIVENESS_SURFACE=0 — the serving query reverts to the exact current 24 h behavior and the continuity tier is simply never selected. Proposed operating point if pursued: the conservative 10 nm / 48 h / both-conditions knee (+4.6 recall pts at 20.4% precision, ~72 added-FP/slice). Measured cost at the current 72 h point: +7.6 recall pts for ~142 added-FP/slice at ~20% precision.

Go / no-go. NO-GO for public surfacing at any operating point measured. The binding fact is that this IRWIN feed carries no FireOut/containment timestamp, so the gate cannot retire dead fires and rides entirely on the recent-touch window; that caps surfaced-roster precision near 20% (added-FP ~80%). For a firefighter-facing roster whose entire value is precision (entries 0003 / 0007), a 5× roster inflation is the wrong trade. The recommendation is to keep the lever GATED as a shadow, do not flip. The only surfacing worth reconsidering later is a clearly-labeled, opt-in, default-OFF monitoring/continuity overlay kept strictly separate from the confirmed roster — and only after a genuine de-activation signal (a real FireOut/containment ingest, or a persistence/decay model) exists to power condition (a). Mark decides; this entry does not change any shipped number.

Temporal dynamics is the first added channel that alone nearly matches production out-of-time — but it can’t be promoted, because the part that transfers is GOES persistence, which production already encodes

BLUF. Entries 0018 and 0023 killed the recovery re-ranker’s geography channel — raw lat/lon (0018) and physical terrain (0023) both failed the temporal-transport gate on the reporting-bias-free neutral truth. The one feature channel never tried is temporal dynamics: the shape of a candidate’s own detection time-series (FRP trajectory, GOES multi-frame persistence, diurnal signature, multi-overpass growth, duration) — physical and non-memorizable, and exactly what a static FRP threshold cannot see. We engineered 20 temporal features strictly from each event’s pre-decision observations and ran the identical gate. Two findings, one negative and one genuinely new. (1) The promote question is still no: recovery+temporal reaches neutral AUC 0.559 — nowhere near production (0.856), and only +0.004 AUC / +0.004 PR-AUC above recovery det-only (0.555). Adding temporal shape to the recovery model buys nothing out-of-time. (2) But temporal shape is not inert like geography was: temporal-dynamics alone reaches neutral AUC 0.840 — the strongest single added channel we have measured (geo-only was 0.72 raw / 0.58 physical), nearly matching production, and unlike geography it does not collapse from honest to neutral truth (0.852 → 0.840, transferring almost intact). Permutation importance shows why it can’t be promoted: the transferable signal is GOES multi-frame persistence (tf_n_goes, dominating by 5–10×) — the exact quantity production’s tier_and_conf already uses. Temporal dynamics re-derives production’s own persistence lever; it does not exceed it. The canary stays gated.

The split (identical to entries 0018/0023). The same CAUTION candidate set, the same 19 detection-only base features, the same honest training label, the same neutral evaluation label, the same model. Train on first_det in 06-27–07-02 (123,931 events), a 07-03 gap, test strictly-later 07-04–07-06 (38,549 events). Models train on the honest (IRWIN/perimeter/confirmed) truth and are evaluated out-of-time against the NEUTRAL truth — an independent VIIRS/MODIS detection (belonging to a different canonical event) within 5 km in a subsequent overpass (t0+3 h…t0+72 h), gridded 0.05°×6 h over CONUS, HMS/IRWIN-independent. Neutral base rate on the holdout is 3.35% (1,292 positives) — reproducing entries 0018/0023. The only change vs entry 0023 is the added channel: temporal-dynamics features instead of geography.

The temporal features, and the leakage cutoff (stated explicitly). For each candidate, t0 = first_det = MIN(detected_at) over its non-leaky sources. Temporal features are built only from that event’s own non-leaky observations with detected_at < t0 + 3 h — the pre-decision accumulation window [t0, t0+3h), which ends exactly where the neutral evaluation window [t0+3h, t0+72h] opens. The two windows are temporally disjoint by construction, so no temporal feature can incorporate the subsequent-overpass detection that defines the neutral label — nor any other detection in that window — and the code asserts max(pre-obs) < t0+3h per event. This is a stricter non-leakage standard than the 19 base features receive (those aggregate the full event span, as in 0018/0023), so the temporal channel is held to a higher bar than the baseline it is added to. IRWIN/HMS/SAR/S2 confirmation channels stay excluded from features, as in 0018/0023. The 20 features cover: FRP trajectory (slope per hour, first/last/delta/range/std/max); GOES persistence (frame count, temporal span, fraction-of-expected-frames, max inter-frame gap); diurnal signature (day&night present, night ratio); multi-overpass growth (distinct overpasses, 2h-vs-1h detection growth); and duration/recency (obs count, span, inter-obs interval mean/std, spatial spread).

Out-of-time on the NEUTRAL (reporting-bias-free) truth — the gate.

The promote question fails on both counts — but the control tells a new story. Recovery+temporal (0.559) does not beat production (0.856): it is 0.30 AUC below. And it does not meaningfully beat recovery det-only (0.555): a +0.004 AUC / +0.004 PR-AUC nudge, inside the noise. So the tested scorer — adding temporal shape to the recovery re-ranker — is a clean negative, exactly like geography. What is different from 0018/0023 is the last row: temporal-dynamics alone reaches 0.840 AUC, only 0.016 short of production and dramatically above the geo-only controls (raw lat/lon 0.723; physical terrain 0.583). Temporal shape, unlike geography, carries real out-of-time neutral signal. Every scorer still scores 0 at the 80/90% precision operating points on this sparse 3.35%-base truth, so the discrimination comparison rests on AUC/PR-AUC.

The paradox, and its resolution: temporal-only 0.840 but recovery+temporal only 0.559. How can a channel that scores 0.840 alone add nothing (+0.004) when bolted onto det-only? Because every model is trained on the honest (HMS-inclusive, reporting-biased) label and then scored on the neutral truth. The detection features fit the biased honest label slightly better (det-only honest AUC 0.877 vs temporal-only 0.852) but their honest-fit ranking transfers poorly to the unbiased neutral truth (0.877 → 0.555, a 0.32 collapse). The temporal features fit honest truth marginally worse, but the honest-fit they find is dominated by GOES persistence — a physical quantity aligned with what the neutral polar truth actually rewards — so it transfers almost intact (0.852 → 0.840). When both feature blocks are present, the detection features win the internal honest-fit competition and pull the combined model back toward the non-transferring solution; the transferable persistence signal is masked. The clean transferable signal only surfaces when temporal features are used alone.

What carries the temporal signal — permutation importance on the neutral truth (the gate). Permuting each temporal feature in the combined model and measuring the drop in neutral AUC:

One feature — tf_n_goes, the count of GOES frames in the first 3 h — carries the temporal signal, by 5–10× over everything else. That is persistence: a real fire lights up many consecutive geostationary frames; a transient false alarm flickers a few. And it is precisely the quantity production’s tier_and_conf rule already uses (its n_goes ≥ 3 persistence clause). The elaborate FRP-trajectory, diurnal, and growth features — the parts a static threshold genuinely cannot see — add nothing or actively hurt out-of-time. Temporal dynamics, stripped to what transfers, is GOES persistence, and production is already a persistence rule.

For contrast — the same holdout scored against the HONEST (HMS-inclusive) truth.

On the honest truth, temporal dynamics adds a trivial +0.002 AUC to det-only — and here is the tell that separates it from geography: raw lat/lon added +0.068 AUC on honest truth (0018) by memorizing the reporting surface, then collapsed on neutral truth. Temporal dynamics adds almost nothing on honest truth (it cannot memorize the reporting surface either) yet nearly matches production on neutral truth — the opposite honest-vs-neutral fingerprint from geography. Temporal shape is honest and physically transferable; it just is not orthogonal to the production baseline.

Certifying the surfaced false-positive rate with independent physical evidence: a stratified, adjudicated fire-vs-non-fire corpus — 0022’s 19.7% holds at 20.7%, but only a third of detections can be physically certified and the rest still rest on heuristics

The framing. Entry 0022 gave the program its first true surfaced false-positive rate — 19.7% — but was honest that 298 of 1,985 detections (the “ambiguous middle”) were adjudicated only by light heuristics, so the number it could actually stand behind was the wide bracket [12.2%, 27.3%]. This entry attacks that residual with the strongest independent physical evidence we can marshal on-box: a stratified sample of the surfaced set, deliberately over-weighting where FP concentrates and uncertainty is highest, adjudicated against IRWIN/WFIGS incidents, ≥2-independent-sensor-family corroboration, Sentinel-2 SWIR / Sentinel-1 SAR burn evidence, a new independent thermal-facility registry (the WRI Global Power Plant Database, 3,839 US combustion plants — a source 0022 did not use), and a 30-day fixed-cell-vs-spreading persistence discriminator. The corpus is persisted as a hidden shadow table — the training seed for a learned false-positive filter.

BLUF. Over the trailing-window surfaced set (N = 1,476 fused fires today; live serving filter last_seen≥now−24h AND tier≠'fp' AND confidence≥0.10), a stratified sample of n = 454 re-weighted to the population by a Neyman estimator puts the true surfaced false-positive rate at 20.7% (95% sampling CI [16.9%, 24.4%]; ambiguous-band bracket [13.9%, 27.4%]). That is a clean independent confirmation of entry 0022’s 19.7% — the point moved +1.0 point, inside noise — and the residual ambiguous fraction fell from 0022’s 15.0% to 13.2%. The tier structure is unchanged and stark: confirmed tier is ~0% FP, probable tier is 26.9%. But the sharper, more honest finding is what happens when we trust only certified physical evidence and send every heuristic lean back to “ambiguous”: the estimate rises to 32.4% with a bracket of [2.2%, 62.6%], because only 33.5% of surfaced detections can be physically certified either way — the 20.7% point’s tightness leans on a heuristic multi-day-persistence “fire” class that is 40% of the sample and cannot be certified without imagery or a ground visit. The bottom line: 0022’s number is robust and now independently reproduced, and the new power-plant registry certified 21 non-fire detections 0022’s catalogs missed on the surfaced set — but the stronger evidence tightens the residual only modestly; closing it the rest of the way requires a human-reviewed gold subset, not more automation. Fully read-only on production; the corpus is a gated hidden shadow; no shipped number changes.

The sampling design. The surfaced population is partitioned into tier × sensor-family-mix strata, and the sample is allocated Neyman-style — heaviest where the false-positive burden and its variance are largest (from 0022: lone-VIIRS, HMS-only, and the oil-and-gas-basin slice). Small strata are near-censused; the large low-FP strata are sampled lightly. Every stratum is re-weighted to its true population size in the estimator, with a finite-population correction.

How each detection is adjudicated — a precedence stack of independent physical evidence. (1) Certified fire if within 10 nm of an IRWIN/WFIGS incident — and we verified the incident feed is entirely fire-type (93.7% WF wildfire, 4.5% RX prescribed burn, 1.8% CX complex; no non-fire all-hazards categories), so an incident match certifies a real thermal fire, not a false positive; else (2) certified non-fire if it sits on a static industrial catalog — fp_polygons (OSM refinery / gas-flare / persistence) or the new WRI power-plant registry (US combustion plants, 2 km); else (3) certified fire if ≥2 independent sensor families corroborate ({VIIRS, MODIS, GOES, S2, S1}, HMS excluded as analyst-derived); else (4) certified fire if independent S2-SWIR / S1-SAR burn/hotspot evidence sits within 10 nm; else (5) certified non-fire if the location shows fixed-cell persistence — ≥5 distinct detection-days over a 30-day window with <1.5 km spatial extent and never an incident (a stationary industrial source); else heuristic leans: fire if multi-day persistence shows spatial growth (≥3 km extent) or an energetic isolated polar detection (FRP ≥ 10 MW away from any catalog); non-fire if inside an O&G basin adjacent to a flare/refinery cluster, or a weak ephemeral hit with no polar corroboration in 30 days; else ambiguous.

The certified false-positive rate (stratified Neyman, re-weighted to population). A detection contributes 1 if non-fire, 0 if fire, ½ if ambiguous (point); the ambiguous band is swept entirely to fire (low) or non-fire (high) for the epistemic bracket; the CI is the stratified sampling error with finite-population correction.

The FP is a single-sensor-family artifact — reproducing 0022 independently. Split by surfacing sensor mix, the burden concentrates in lone-polar and analyst-only detections; every combination adding a second independent family collapses it:

The oil-and-gas-basin cross-cut (within 15 km of a flare/refinery cluster; n = 62 sampled) reads 15.3% non-fire — lower than lone-VIIRS because many basin detections sit on the industrial catalog and are certified non-fire outright, while genuine fires also occur there. The residual error still lives exactly where 0022 said it did: a probable-tier, lone-polar or analyst-only detection in an oil-and-gas basin.

The honest part — how much of this is actually certified. Two estimators of the same set, one trusting the heuristic leans, one refusing them:

The wide certified-only bracket is the point of the entry: the sharp 20.7% depends on a large heuristic fire class (multi-day persistence with growth, 40% of the sample) that we cannot physically certify without imagery or ground truth. Because that class leans toward fire, if it is wrong it can only push true FP up — which is why the honest point sits near the top of 0022’s bracket and the certified-only estimate is higher, not lower.

The 0020/0022 decomposition, reproduced out-of-sample.

Two independent adjudications, a day apart, on different sampled sets, land within ~1–6 points on every component — the decomposition is stable, not an artifact of either method.

A power-plant / industrial-combustion false-positive mask, built pre-cutoff and transport-gated on a held-out later window — masks 68 combustion plants (7 net-new beyond the refinery/flare masks) and hides zero real later fires at 1.5 km, but a landfill fire 1.99 km from a waste-to-energy plant shows why 2.0 km is one incident too wide — 0029

The framing. Entry 0028 certified the surfaced non-fire detections against an independent thermal-facility registry — the WRI Global Power Plant Database — and found that the certified-non-fire detections it could name all landed on combustion power plants: an industrial-heat false-positive class that the shipped gas-flaring/refinery masks (osm_refinery, osm_gas_flare, flaring_nightfire_persistence) do not cover. Coal / gas / oil / biomass / waste-to-energy plants run persistent hot stacks and flares that VIIRS 375 m and GOES sub-pixel routinely flag as fire. This entry builds a mask for them with exactly the discipline the flaring mask shipped under — a small static polygon per plant, gated on evidence that OLDGABE actually generates persistent detections there — and, because a real wildfire can start next to a power plant, subjects it to a strict temporal-transport gate: the mask is built from data strictly before a cutoff and evaluated on the held-out later window for the one number that matters, how many real later fires it would hide.

BLUF. From 3,719 US combustion plants (WRI GPPD, fuel ∈ {Coal, Gas, Oil, Petcoke, Cogeneration, Biomass, Waste}; solar/wind/hydro/nuclear/storage excluded as non-combustion), intersecting each with OLDGABE’s own persistent detections over the pre-cutoff window (train < 2026-06-21; VIIRS/MODIS/GOES thermal evidence), requiring ≥5 distinct detection-days in a 1.5 km cell that is single-sensor-family, spatially static (≤1.5 km extent), and ≥5 km from any pre-cutoff IRWIN/WFIGS incident, yields a mask of 68 plants. Only 7 are net-new beyond the existing fp_polygons masks (the other 61 already sit under a refinery/flare polygon); the 7 are landfill-gas / waste-to-energy and standalone gas+cogen sites the OSM catalogs miss. On the held-out later window the mask would tier 8,497 net-new persistent thermal detections (152 net-new production-surfaced events) as fp — and it hides zero real IRWIN/WFIGS fires: the nearest actual held-out incident to any masked plant is 1.99 km, outside the 1.5 km radius. Widen the radius to 2.0 km and that single incident — a genuine landfill fire 1.99 km from the Kiefer Landfill waste-to-energy site — falls inside, so 2.0 km masks one real fire and 1.5 km masks none. 1.5 km is the safe radius. The mask is persisted as a hidden shadow table (surface_allowed=0, DO_NOT_SURFACE, not read by serving); production fusion, tiers, fused_fires and the map are untouched. No cutover.

A necessary correction to the truth signal. The obvious way to ask “is there a real fire at this plant later?” — look at fire_events near the plant with status='confirmed' or a non-null irwin_id — is wrong here, and the reason is itself the finding. At the flagged plants the production pipeline auto-confirms the plant’s own persistent signature and cross-links it to a distant IRWIN incident id (the same id repeated day after day). Plains End (a gas peaker in Colorado) generated 21 “confirmed” test-window events (confidence up to 0.9) carrying a borrowed IRWIN id whose actual incident sits 12.7 km away; Hidden Hollow (waste-to-energy, Idaho) did the same with an incident 10.4 km away. Those labels are not evidence of a fire at the plant — they are the industrial false positive we are targeting, wearing a borrowed badge. The truth-sparing test therefore uses actual IRWIN incident coordinates (and WFIGS points), never the fire_events cross-link. Measured against real incident geometry, every masked plant’s nearest held-out fire is ≥1.99 km.

The transport gate. Build strictly on pre-cutoff data; evaluate on the held-out later window. “Net-new” counts only detections/plants not already covered by an existing fp_polygons polygon (no double-counting the flare/refinery masks). “Real incidents masked” = masked plants with an actual post-cutoff IRWIN/WFIGS incident point inside the radius — the pass/fail. “Growth-caveat” = masked plants whose own detections spread >1.5 km across the later window (disclosed, not a fire unless an incident corroborates).

The single real-fire mask at 2.0 km is instructive: it is a landfill fire 1.99 km from a waste-to-energy plant — the exact case where a combustion facility and a genuine fire coexist. The 1.5 km radius spares it at every persistence threshold. The growth-caveat plants at 1.5 km/≥5 (5 of them: Delray, ExxonMobil Baytown, Gregory, Port Arthur, Valero Corpus Christi) are all Gulf-Coast refinery/gas complexes whose multi-unit footprint spans several VIIRS pixels — none has a real incident within range, and all but one already sit under osm_refinery, so they are excluded from the net-new set anyway. The 7 net-new plants (Bowerman Power LFG, Delta Energy Center, Higgins Generating Station, Lake Charles cogen, Pine Tree Acres landfill-gas, TVA Southaven combined-cycle) are compact (≤0.72 km train extent), single-family, 5–11 persistent days, and their nearest real held-out incident is ≥4.3 km or none.

A short-horizon fire-spread nowcast is input-feasible but shows no skill yet — every physical driver (RAWS wind, ETOPO slope, dead-fuel moisture) is on-box and queryable at active fires, yet a wind-driven downwind+upslope growth model scores at the 90° random-direction null (mean 88.6°, mean-resultant R=0.02, Rayleigh p=0.88) while the fire’s own motion predicts its direction (R=0.44); the two blockers are a live-only wind snapshot and 0.01°-quantized growth truth — 0030

The framing. OLDGABE detects fires. The re-ranker work (entries 0018/0023) showed that fuel/terrain/climatology priors do not help the detection decision — but those same covariates are the physically correct inputs for a different problem: spread. A 1–6 h direction-and-growth nowcast would turn “there is a fire here” into decision-support — “it is running NE at roughly X.” This entry does the honest feasibility-plus-prototype work: inventory the real inputs on-box, build the simplest defensible physical model (wind-driven elliptical growth with an upslope bias), and validate it against observed fire growth with a baseline so the skill number means something. The result is a clean negative for the simple current-wind model, with the two fixable reasons named.

BLUF. Every physical driver a spread nowcast needs is already on-box and queryable at fire locations: 2,078 RAWS stations (all with wind speed/direction and dead-fuel-moisture, refreshed hourly), a coarse-DEM terrain sample (ETOPO2022 60″, median slope 17.9 m/km at the fire set), and the fused active-fire track itself. So the inputs are feasible. But the simple physical model has no measured skill. On 250 recently-active confirmed/probable fused fires (each with an ≥8-detection, ≥3 h multi-sensor track), a downwind spread bearing (wind-direction + 180°) lands at the 90° random-direction null — mean angular error 88.6°, mean-resultant length R = 0.02, Rayleigh p = 0.88 (no concentration in either wind convention, since R is sign-independent). Adding an upslope bias makes it worse (93.2°). A wind×dryness rate index does not predict observed footprint growth (Spearman −0.20 vs. a predict-median baseline of 0). Yet the target is genuinely predictable: a persistence baseline — the fire’s own earlier centroid motion — predicts its later direction well (mean error 55.5°, R = 0.44, Rayleigh p < 10⁻³⁰⁰). The two blockers are diagnosed: (1) wind is a live snapshot only — there is no time-archived wind co-registered to historical detections, and the median fire is 27 km from its nearest RAWS, so the wind vector is temporally and spatially mismatched to the growth window; (2) the only on-box growth truth is detection-footprint expansion at 0.01° (≈1.1 km) member-coordinate quantization, GOES-dominated, with median observed displacement 0.44 km — below one pixel, so the growth vector is mostly noise. Verdict: feasible in inputs, not-yet-skillful in output. Nothing was surfaced; no production table was touched.

The input inventory. The point of a feasibility pass is to state plainly what exists, what is a proxy, and what is deferred.

The prototype. Deliberately the simplest defensible physical cut, not an ML model. Direction = a weighted vector sum of a downwind unit vector (bearing = RAWS wind-direction + 180°, weighted by wind speed) and an upslope unit vector (bearing of steepest DEM ascent, weighted by slope magnitude, capped at the wind weight) — fire runs downwind and uphill. Rate = an interpretable wind-driven, dryness-scaled index, ROS ∝ wind_speed × exp(−fuel_moisture/10), in the Rothermel/empirical-grassfire spirit [21][22]. Observed growth per fire = split the fire’s own detection track into early/late thirds by time, take each third’s centroid; the early→late bearing is the observed spread direction and the centroid displacement over elapsed hours is the observed rate; the mean detection-to-centroid radius (late minus early) is a footprint-growth proxy. Skill is scored with circular statistics (mean angular error, mean-resultant length R, Rayleigh test) against a 90° uniform-direction null and against a persistence baseline [16][23].

‡ On the 17 resolved-displacement fires the wind vector shows moderate concentration (R=0.50) but at a ~176° offset — the observed footprint centroid moves opposite to downwind. This is almost certainly a truth artifact, not physics: coarse GOES footprints are first detected at the hot leading edge and then “fill in” toward the older/interior pixels, dragging the centroid backward. It is a reason to distrust the footprint-growth proxy, not evidence the model is anti-skillful. On the robust n=250 sample the wind axis carries essentially no directional information at all (R=0.02).

Rate skill. On the resolvable subset (n=17) the wind×dryness ROS index vs. observed centroid-displacement rate gives Spearman −0.20 (Pearson −0.11); wind-speed alone −0.21; ROS index vs. footprint-radius growth −0.21. All are at or below zero — no positive skill over the predict-median baseline (which is 0 by construction). With this truth and this sample the rate model is indistinguishable from no information.

Can a disciplined GOES-first provisional-alert tier buy real lead time at acceptable precision? — measured across 481k early GOES-first candidates and out-of-time: no. Every gate that adds discipline lowers precision, and the confirmable slice is ~13% before any gate — a MEASURED NEGATIVE that upholds entry 0025’s caution

The framing. Entry 0025 found that GOES-ABI wins the first-arrival race for ~76% of fires on cadence alone, yet is the slowest thermal sensor for genuine new-fire detection, and it retired the “we detect before it’s reported” headline as a lookback artifact. It left one question open and worth answering honestly: is there a disciplined GOES-first provisional tier — persistence, FRP, and not-on-a-known-false-source — that recovers a real latency gain at a precision a clearly-labeled advisory surface could tolerate? This entry defines that candidate population, measures the full precision-vs-lead-time curve across gates, and validates the best operating point out-of-time. The answer is a clean negative, and it sharpens why.

BLUF. The GOES-first candidate population — canonical events where a GOES-ABI detection is strictly the first sensor at the location, i.e. everything a provisional fast-alert could draw from before any polar/HMS/IRWIN confirmation — is enormous and mostly false: 480,953 events over 39 days (~12,300/day), of which only 12.9% are ever corroborated as a real fire (12.4% by an IRWIN/WFIGS fire within 10 nm; a vanishing 0.5% by a later co-registered VIIRS/MODIS overpass in the same canonical event). So a raw GOES-first alert is wrong ~6.8 times out of every 7.8, at 12,000+ alerts/day. The decisive result is that no disciplined gate rescues it: requiring GOES multi-frame persistence (≥3, ≥5 frames), an FRP floor (≥10, ≥30 MW), excluding the flaring/refinery and static-thermal masks, or any combination — every gate holds or lowers precision (the full curve sits in an 8–13% band) while shedding recall and lead time. Persistence and FRP go the wrong way (they select persistent warm surfaces, not small new fires), and the masks barely move the needle because the false mass is diffuse coarse-pixel noise across fire country, not concentrated on known industrial points (only 0.6% of candidates sit on a flaring/refinery polygon). Out-of-time, the gates transfer as non-skill: the apparent early→late precision jump (9.9%→17.1% raw) is a seasonal fire-density base-rate effect, not gate discrimination, and the single point that clears ~29% precision on the later window (FRP≥30) does so at ~158 alerts/day, 5% recall, and only ~4.8 h of lead. Verdict: a GOES-first provisional tier is too false-alarm-heavy to surface, even gated. We recommend no shadow build; entry 0025’s caution stands, now quantified end-to-end. Nothing was surfaced and no production table was touched.

The candidate population (reproducible). Working unit is the canonical event (the shipped corroboration/ground-truth key). For every canonical event we take, per sensor family, the first detection time; a GOES-first candidate is one where the GOES-ABI family (Dozier-subpixel + ABI-ADP-smoke + raw ABI, G18/G19) is strictly earliest — nothing polar, HMS, or IRWIN arrived at or before it — so by construction it is sub-confirmation at alert time and is exactly the pool a provisional alert would fire on. Truth reuses the shipped scorer (te_daily_scorer.py): a candidate is genuinely confirmed if an IRWIN/WFIGS WF/WFU fire lies within 10 nm (independent), or a later VIIRS/MODIS detection co-registers into the same canonical event (≥2 independent families). Lead time is measured against an actual confirmation event — the first co-registered polar detection minus the alert time — not against IRWIN’s reported discovery time, precisely to avoid the reporting-time artifact entry 0025 diagnosed. The candidate window is t₀ ∈ 2026-05-31 → 2026-07-09 (a 2-day tail is trimmed so late candidates have time to confirm; right-censoring caveat below).

Read this first: the confirmable fraction is ~13% before any gate is applied, and it is almost entirely IRWIN-proximity (12.4%), not independent sensor agreement (0.5%). A 10 nm IRWIN match in fire country is itself the spatial-association signal entry 0025 warned about — so 12.4% is an upper bound on the fraction that are genuinely this new fire, not a floor. The clean lead-time number (+7.9 h vs an actual polar overpass) is real but exists for only the 0.5% of candidates that ever get a co-registered polar detection; it is not a property of the population.

Result 1 — the precision-vs-lead-time curve (this is the finding). Same candidate set; one gate at a time and in combination. Precision = fraction of kept candidates that are genuinely confirmed; recall = fraction of the eventually-confirmed set retained; lead = median hours the alert precedes the first co-registered polar detection (measured only where such a detection exists):

The curve never leaves the 8–13% precision band. The best precision in the whole table is the raw population (12.9%); every added constraint makes precision worse, not better, and buys nothing but lost recall and shrinking lead. That inversion is the mechanism: GOES multi-frame persistence and a high FRP select for bright, durable hot signals — which in the GOES-first pool are disproportionately persistent warm surfaces and residual industrial/agricultural burn, not the small, brief new ignitions IRWIN logs. The two dedicated masks remove real false sources but almost no volume (recall stays 89–99%), because the false-alarm mass is diffuse coarse-pixel noise spread across the fire-prone West, not concentrated on the 0.6% of points that are flares/refineries. There is no operating point that is simultaneously earlier and precise enough.

Result 2 — transport gate (out-of-time; entries 0018/0021/0027 discipline). Build/choose the gate on an earlier window (t₀ < 2026-06-23, n=280,880, 27,763 confirmed) and measure it on a strictly-later window (t₀ ≥ 2026-06-23, n=200,073, 34,114 confirmed). If a gate carried real discrimination it would hold its precision advantage out-of-time; instead precision tracks the ambient fire density of the window:

The precision on the late window is higher for every gate — including the raw no-gate population — by roughly the same factor. That is a base-rate effect: the later window is deeper into fire season, so any 10 nm cell is more likely to sit near an active IRWIN fire, lifting the proximity-match rate regardless of what the gate does. It is not gate skill, and it does not transport as a precision guarantee an operator could rely on. The one point that reaches ~29% out-of-time (FRP≥30) trades it for ~158 alerts/day, ~5% recall of confirmed fires, and ~4.8 h of lead — i.e. it surfaces one-twentieth of the real fires, still gets it wrong ~7 times in 10, and does so barely earlier than simply waiting. No operating point clears a defensible advisory bar.

The precision bar, stated explicitly. The confirmed roster runs at ~90%+ precision (entries 0003/0007); a provisional/advisory tier can tolerate less, but it must beat surfacing nothing — call the generous advisory bar ~50% (roughly one real fire per false alert). The full curve is reported above precisely so the verdict does not hinge on where that bar sits: nothing measured — in-sample or out-of-time, raw or gated — approaches 50%. The population tops out at ~13% in-sample and ~17% on the season-dense later window; the only points reaching ~29% do so at negligible volume and single-digit recall. Wherever a reasonable advisory bar is drawn between the confirmed roster and a coin flip, a GOES-first provisional tier is below it.

Adding Landsat-8/9 OLI SWIR to densify coverage: the truly-blind tail is 16 fires and 100% Alaska, all outside the deployed S2-SWIR footprint — so L8/L9 buys zero as a cadence densifier and at most 3–4 large Alaska fires as a spatial gap-fill. Confirmed REDUNDANT-over-CONUS / narrow-Alaska-only

BLUF. Entry 0024 named the one genuine remaining coverage lever — the sub-floor tail — and noted OLDGABE already ingests a Sentinel-2 SWIR-hotspot source (S2-SWIR-Hotspot), so the only open question is whether densifying that cadence by adding Landsat-8/9 OLI SWIR (B6 ~1.6 µm / B7 ~2.2 µm, 30 m; L8+L9 phased for ~8-day repeat, tighter at high latitude via swath sidelap) recovers any net-new fire. We reproduced the sub-floor set against the full any-status event net and the answer is a clean no, with one narrow Alaska caveat. Two facts decide it. (1) The truly-blind tail — IRWIN-active wildfires with no OLDGABE event of any status (candidate, masked-fp, confirmed, …) within 10 nm over their entire active window — is just 16 fires, and every one is in Alaska (stable at 16 across four consecutive scoring windows 07-08→07-11). Not one is in the Lower 48: OLDGABE’s detection net blankets CONUS so densely that zero CONUS wildfires are event-free, so a densifier has nothing to recover there. (2) The deployed S2-SWIR daemon is hard-coded to a CONUS bounding box (24–50°N, 125–66°W) — it is not pointed at Alaska at all — so all 16 blind fires sit outside the existing S2 footprint. That reframes the lever entirely: L8/L9 would add no temporal cadence anywhere S2 already looks (S2’s 5-day repeat is denser than L8/L9’s 8-day at every latitude, and the open sub-floor windows span weeks, so S2 already delivers many overpasses per window), and the only coverage it adds is spatial — Alaska, where the real gap is that the S2 daemon’s box stops at 50°N. Of the 16 Alaska fires, 12 (75%) are sub-1-acre point records (1.8 ac total) — hopeless at a 30 m SWIR sub-pixel floor no matter how often overflown — leaving a net-new recoverable upper bound of just 4 fires ≥1 acre (505 ac), a robust core of 3 fires ≥10 acres, dominated by a single 409-acre Alaska burn. Verdict: L8/L9 SWIR densification is REDUNDANT over CONUS and a narrow, Alaska-only, large-fire-only spatial gap-fill worth at most a default-OFF shadow — and even that gap is better closed by pointing SWIR ingestion at Alaska (L8/L9 or extending the S2 box north) than by “more cadence.” Nothing surfaced, no production table touched.

Feasibility — the lightest real path to L8/L9 SWIR active fire. Unlike VIIRS/MODIS there is no operational Landsat active-fire feed (no FIRMS-style Landsat hotspot product); hotspots must be derived from OLI SWIR scenes using an established algorithm — the Schroeder et al. (2016) Landsat-8 OLI active-fire algorithm [12] or the Kumar–Roy (2018) reflectance-based OLI fire test [25] (B6/B7 thresholds plus a folding/contextual test). The good news is the ingestion mechanism already exists in-shape: the deployed S2 daemon reads B11/B12 COGs over an open STAC/COG path (earth-search.aws.element84.com, streaming only the SWIR bytes it needs via rasterio’s HTTP driver) — the identical pattern applies to Landsat Collection-2 (usgs-landsat / Element84 landsat-c2-l2 on AWS), reading OLI B6/B7 COGs and applying the same B12–B11-style threshold logic. Realistic ingestion cost over CONUS+AK is modest: L8+L9 acquire on the order of a few hundred daytime WRS-2 scenes/day combined; with the strided SWIR-window reads the S2 daemon already uses, that is low-single-digit GB/day, not a scene-download pipeline. The binding caveat is latency: Landsat Collection-2 L2 (surface-reflectance/temperature) publishes at ~12–24 h+ — so, exactly like S2, this is a coverage/confirmation lever, not a fast-alert lever (consistent with entry 0024’s framing that SWIR is burn-scar/coverage, not active alerting).

The reproduced sub-floor set (scoring window 2026-07-10→2026-07-11, entry-0024-equivalent; 10 nm vs IRWIN WF/WFU; 3,698 active wildfires; 16 truly-blind fires — no OLDGABE event of any status within 10 nm/active-window over a 30-day lookback; the count is invariant — 16 — across all four windows tested):

The whole tail is Alaska. All 16 fires carry POOState=US-AK; zero are inside the deployed S2-SWIR CONUS box; every one is an open IRWIN record (null FireOutDateTime), so its active window runs discovery→now (days to weeks). The mechanism is structural: OLDGABE’s candidate/detection net (GOES-ABI + HMS + polar) is so dense over the fire-prone West that no CONUS wildfire ends up event-free within 10 nm over a multi-week window — the blind tail is precisely the region the net under-covers, which is Alaska.

The net-new coverage delta over the deployed S2 baseline (deployed S2-SWIR-Hotspot = B11/B12 20 m, ~5-day A+B, CONUS box 24–50°N only; candidate add = L8/L9 OLI B6/B7 30 m, ~8-day; upper bounds throughout, from revisit geometry not imagery):

Reading the delta honestly. The net-new recoverable upper bound is 4 fires / 505 acres, and it is acreage-rich only because a single 409-acre Alaska burn dominates it. Strip to what a 30 m OLI SWIR band would robustly resolve and the core is 3 fires ≥10 acres. Two independent reasons make even that 3–4 optimistic. First, it assumes a cloud-free daytime overpass coincides with active flaming — Alaska summer cloud frequency, solar-zenith, and the overpass/flaming coincidence all push realistic yield below the bound. Second, and decisively for the densification question: L8/L9 adds no temporal reach that S2 lacks. S2’s two-satellite 5-day repeat is denser than L8/L9’s two-satellite 8-day repeat at every latitude (both tighten poleward via sidelap, but S2 stays ahead), and the sub-floor windows are weeks long, so wherever S2 can look it already delivers multiple overpasses. L8/L9’s only contribution over the deployed baseline is spatial — Alaska, which the S2 daemon’s box simply excludes. So “densify cadence with L8/L9” is the wrong framing for the win that exists; the win, such as it is, is “cover Alaska at all,” achievable equally by extending the existing S2 box north.

Verdict — confirmed low-yield / redundant-as-densifier. Against a roster whose generous ever-detected recall is (3,698−16)/3,698 = 99.6% within the active window, the truly-blind tail is 16 fires, 75% of them orbit-undetectable sub-1-acre point records, entirely in Alaska and entirely outside the deployed S2 footprint. Adding Landsat-8/9 OLI SWIR as a cadence densifier of the existing (CONUS) S2 feed recovers zero fires. The only real prize is a narrow, Alaska-only, large-fire-only spatial gap of 3–4 fires — and it is a coverage-footprint fix, not a cadence fix. If pursued at all it justifies at most a default-OFF, Alaska-focused, ≥10-acre SWIR shadow probe measured for recovered-recall vs added-FP — whether that SWIR comes from L8/L9 or from extending the S2 box north is an implementation detail. This is exactly the “real but marginal” ceiling entry 0024 anticipated, now localized to Alaska and quantified to single digits.

A learned “real-bogus” FP filter on the lone-VIIRS slice re-derives the shipped masks — the spectral/footprint/diurnal signal beyond the masks carries only 6% of the discriminative power and cannot flag off-registry false positives out-of-time at usable precision

BLUF. OLDGABE’s surfaced false positives concentrate in the probable-tier lone-VIIRS slice (entries 0022/0028: ~20% surfaced FP, ~27% in the probable tier). We already ship static/persistence FP masks — refinery/gas-flare polygons, VIIRS-Nightfire flaring, and Dozier/polar persistence cells (all in fp_polygons) — plus a power-plant registry cutover-candidate. The question for this wave, framed skeptically: can a learned classifier separate genuine early fire from persistent non-fire in that slice using signal beyond what the masks already encode? The trap (identical to entry 0027’s temporal re-ranker, which just re-derived production’s n_goes≥3): if the classifier’s power is entirely mask/persistence/registry features, it is redundant with what ships. We built two feature sets — a mask/persistence baseline (on-polygon, distance-to-mask, fixed-cell persistence, flare-cluster density, registry proximity, peak FRP) and a non-mask candidate (VIIRS I4/I5 brightness temperature, I4−I5 thermal anomaly, pixel footprint scan×track, day/night ratio, FRP distribution shape, saturation fraction) — trained class-balanced HistGradientBoosting on each and the union, and ran a strict train-earlier / test-later transport gate. The result is a clean negative. Out-of-time, mask-only reaches PR-AUC 0.984; non-mask reaches only 0.869; the union (0.978) is no better than mask-only. Permutation importance on the union puts 94% of the discriminative signal on mask/persistence/registry features (top five all mask: flare-cluster density, distance-to-mask, persistence-days, persistence-extent, registry distance) and only 6% on the entire non-mask block, the best of which (mean I5 background temperature, +0.004) sits an order of magnitude below any mask feature. The decisive test — can the non-mask model catch false positives the shipped polygons miss (off-polygon detections), out-of-time, at usable precision with ~0 real-fire cost? — fails: on the 830 off-polygon test detections the non-mask model at its precision-0.90 training threshold catches 286 FP but flags 207 real fires to do it (precision 0.58), while the union catches the same 332 FP as mask-only alone (119 vs 116 fires flagged) — adding the non-mask channel buys nothing. The masks already capture the separable signal; a learned real-bogus filter re-derives them and adds no promotable, orthogonal value. Fully read-only on production; no shadow build, no shipped number changes.

The population and the labels. The FP-concentrated slice is the lone-VIIRS caution population: event_corroboration_tier where tier='CAUTION' and families∈{VIIRS, VIIRS+HMS} (one independent family, HMS not counted) — 11,572 events, each with its VIIRS member detections (1,760,347 detections parsed for spectral/morphology features). Because 454 hand-adjudicated rows (entry 0028) are too few to train + transport-split, we expand with the same entry-0028 strong-evidence rules applied programmatically: FIRE = IRWIN/WFIGS incident ≤10 nm, or multi-day polar persistence with growth (≥2 distinct days & ≥3 km extent), or energetic FRP≥10 MW away (>10 km) from any static source; NON_FIRE = on a shipped fp_polygons mask, or fixed-cell persistence (≥5 distinct days in a ~2 km cell, <1.5 km extent, no incident), or inside an O&G flare/refinery cluster; everything else = AMBIGUOUS (held out). Labels: 1,566 FIRE / 4,233 NON_FIRE / 5,773 AMBIGUOUS. Supervised set = 5,799.

The transport gate. Split on each event’s first detection time at the 60th percentile (2026-07-05T11:41Z): train = 3,457 (2,536 NON_FIRE / 921 FIRE), test strictly later = 2,342 (1,697 NON_FIRE / 645 FIRE). Positive class = NON_FIRE (i.e. the model predicts “this is a false positive”); test base FP rate 72.5%. Metrics are precision/PR-AUC of FP-identification on the held-out-time set.

Out-of-time FP-identification — the gate.

Mask-only nearly ceilings; the union does not beat it (0.978 vs 0.984, inside noise and marginally lower); non-mask alone is a clear 0.115 PR-AUC below mask-only. Adding the spectral/footprint channel to the mask baseline moves nothing out-of-time.

Where the signal lives — permutation importance on the union (out-of-time PR-AUC drop).

The top five features are all mask/persistence/registry; the entire non-mask block — brightness temperature, thermal anomaly, footprint, diurnal, FRP-shape, saturation, the physics a static polygon genuinely cannot see — contributes 6% of the discriminative power, its best member (mean I5 background temperature) an order of magnitude below any mask feature. Just as entry 0027’s temporal channel collapsed onto GOES n_goes, the real-bogus spectral channel collapses onto the persistence/registry cuts the masks already encode.

The decisive test — can non-mask catch FP the shipped polygons miss? Shipped-mask coverage = on a fp_polygons polygon (on_static=1). The only place a learned filter could add value is off polygon. We fixed each model’s threshold on the training fold at precision ≥ 0.90 and applied it to the 830 off-polygon test detections (332 NON_FIRE / 498 FIRE):

Off-polygon, the non-mask model collapses to 0.58 precision — it flags 207 real fires to catch 286 false positives, an unusable recall cost. The union catches the same 332 true-FP as mask-only (flagging 119 vs 116 real fires) — identical to mask-only within noise. The non-mask channel adds zero incremental off-polygon FP-catch. What does carry the off-polygon catch is persist_days — freshly-computed fixed-cell persistence that is not yet a fp_polygons entry — but that is a mask-family (persistence) feature, i.e. the honest incremental lever is “extend the persistence polygon set to freshly-detected persistent cells,” not “learn a non-mask spectral classifier.”

A finer fire-growth-truth dataset (full-precision VIIRS 375 m footprints + IRWIN acreage series, replacing the 0.01°-quantized member tokens of entry 0030) shows the persistence spread signal is real and stronger, not a truth artifact: mean-resultant R rises 0.44→0.54 on the same all-fires test and 0.63→0.72–0.84 on a now-robust resolvable subset (n=70–117, up from n=13–17); the current-wind model still scores at the random-direction null and the earlier 176° “spreads-upwind” offset vanishes — confirming it was a coarse-GOES footprint-fill artifact — 0034

The framing. Entry 0030 prototyped a short-horizon fire-spread nowcast and found a clean split: the simple physical wind model had no measured skill (R=0.02) while a persistence baseline — the fire’s own earlier motion predicting its later direction — scored R=0.44. But 0030 could not say how much of that 0.44 was real, because its growth truth was detection-footprint centroids read from fused member tokens quantized to 0.01° (≈1.1 km), GOES-dominated, with a median observed displacement of only 0.44 km — below one pixel. On the tiny resolvable subset (n=17) the wind vector even pointed ∼176° opposite to observed motion, which 0030 diagnosed (but could not prove) as a footprint-fill artifact. This entry removes that bottleneck: it rebuilds the growth truth from the raw per-detection evidence at full precision and re-measures every 0030 statistic against it.

BLUF. The raw evidence table carries the finer truth 0030 lacked. Rebuilding growth tracks from full-precision VIIRS active-fire pixels (375 m native, ≈5-decimal coordinates — roughly 3× finer than the 1.1 km fused rounding) and, independently, from the IRWIN IncidentSize acreage time-series, yields a 971-row growth-truth dataset (313 VIIRS footprint tracks, 14 MODIS fallback, 644 IRWIN acreage series). On this finer truth the persistence signal strengthens: on the same all-fires direction test the mean-resultant length rises to R=0.54 (n=150, mean angular error 48.2°) from 0030’s 0.44, and on a resolvable-displacement subset that is now robust — n=70 at ≥0.75 km, n=117 total, versus 0030’s n=13–17 — it reaches R=0.72 (≥0.75 km) and R=0.84 (≥1.1 km, mean error 22.8°, 86% within 45°). The current-wind model still lands at the random-direction null (R≈0.14–0.22, mean error 96–100°), and critically the 176° reversal of 0030 is gone (mean error ≈97°, i.e. simply uninformative rather than anti-correlated) — direct evidence the reversal was a coarse-footprint artifact, exactly as 0030 suspected. On rate, the wind×dryness index still shows no skill (Spearman −0.12), but two independent growth-rate truths agree: VIIRS footprint growth-rate vs. IRWIN acres/hour correlate at Spearman 0.56 (n=63 spatially-linked fires) — the observed rate is real even though the physical predictor does not capture it. Verdict: the persistence-spread signal is real and now well-resolved; a default-OFF persistence advisory is defensibly buildable. Wind still adds nothing without a time-archive. Nothing was surfaced; the only write is a hidden, gated shadow dataset.

Growth-truth source inventory. The point of this pass was to find the finest time-resolved growth record available per fire and state the precision of each candidate honestly.

Method (mirror of entry 0030, finer truth). Anchor on the same corpus — fused_fires with tier∈{confirmed,probable}, n_frames≥8, track span ≥3 h — but instead of parsing the 0.01° member tokens, spatially join the raw full-precision detections (fire_event_evidence: VIIRS-N20/N21/NPP preferred, MODIS fallback) within 7 km of the fused centroid and within the fire’s time window, then recompute the growth track from those coordinates. Split each fire’s detections into early/mid/late thirds by time; the early→late centroid bearing is observed direction and the displacement / elapsed hours is observed rate; the persistence baseline is the early→mid bearing. Direction skill uses circular statistics (mean angular error, mean-resultant length R, Rayleigh test) against a 90° uniform null [23]; the wind model is the RAWS downwind bearing (wind_dir_deg + 180°). Independently, the IRWIN acreage series gives a per-incident acres/hour rate; the two rate truths are cross-checked on fires within 10 km of an IRWIN WF incident. Every read is mode=ro on production.

Comparison to entry 0030 (0.01°-quantized truth): persistence all was R=0.44 (n=165, mean 55.5°); persistence resolvable was R=0.63 (n=13); the wind model was R=0.02 all / R=0.50 at a spurious 176° offset on n=17. On finer truth persistence rises at every cut, the resolvable subset grows ∼5× in size, and the wind model’s R stays at the null with the 176° offset replaced by an ordinary ≈97° (uninformative) error — i.e. the reversal was the coarse-footprint fill artifact, not physics. Median observed displacement rises from 0.44 km (quantized) to 0.84 km (finer), which is why the resolvable subset is no longer starved.

Rate skill. The wind×dryness ROS index still does not predict observed footprint growth-rate on the resolvable subset (Spearman −0.12, n=117; wind-speed-alone +0.07) — no skill over the predict-median baseline, confirming entry 0030. But the observed rate itself is real: on the 63 footprint fires that spatially link (≤10 km) to an IRWIN WF incident, the VIIRS footprint growth-rate and the IRWIN acres/hour series correlate at Spearman 0.56 — two physically independent growth-rate measurements agreeing. The gap is therefore in the predictor (live, spatially-mismatched wind), not in the growth target.

Re-tested as a genuine no-overlap forecast (predictor uses only the fire’s track up to a decision time T, validated strictly on growth after T), the persistence spread signal of entry 0034 does not survive: honest direction skill collapses from R=0.72–0.84 (which shared its start point with the target) to mean-resultant R≈0.18–0.22, mean angular error ≈106°, and — decisively — the concentration sits at a ≈177° offset: at VIIRS-overpass cadence the footprint centroid tends to regress opposite its recent motion, so naive persistence points a firefighter roughly the wrong way. Predicted growth-rate also loses to a climatological-mean baseline in every horizon band. Verdict: not surface-worthy; nothing flipped — 0035

The framing. Entry 0034 rebuilt the fire-growth truth at full VIIRS 375 m precision and showed a persistence direction signal of mean-resultant R=0.72–0.84 on resolvable fires — but stated, in its own honest-bound #1, that the persistence baseline shares its start point with the observed vector (both anchored on the early centroid), so part of that R is construction overlap, not forecast skill. An advisory that helps a firefighter is a forecast: at decision time T you may use only what the fire has done up to T, and you are graded on what it does after T. This entry does exactly that — it removes the overlap entirely and measures the honest forecast skill entry 0034 recommended as the next step.

BLUF. The overlap was doing the work. Splitting each reconstructed VIIRS/MODIS footprint track at a decision time T — predictor window = the overpasses at times ≤T (first ≈60%), validation = each later overpass, with no shared displacement segment between them — the persistence direction forecast falls to mean-resultant R≈0.18–0.22 (all 796 forecast pairs / 232 fires), mean angular error ≈106°, only 17% within 45° (a random null gives 25%) and 36% within 90° (null 50%). A forecast is skilful only when R is high and the mean offset is near 0°; here the offset sits at ≈177° — the modest concentration that exists is concentrated on the wrong answer. Mechanistically this is blob mean-reversion: a fire filling in around its ignition footprint has a centroid that random-walks and regresses, so successive VIIRS-overpass displacements anti-correlate. A last-step (“same as the immediately prior step”) baseline is no better — higher R (0.31) but the same ≈178° offset, i.e. equally anti-persistent. On rate, the persistence prediction never beats the trivial “assume the median growth rate” forecast: climatological-mean MAE is lower in every horizon band (e.g. 6h: 0.022 vs 0.047 km/h). The transport cut (fit the one free parameter — predictor fraction — on earlier fires, test strictly-later fires) is stable but transports a null: later-half R=0.22 at the same ≈179° offset. One honest positive is retained: the footprint growth-rate magnitude is a real measurement (VIIRS footprint ROS vs. independent IRWIN acres/hour, Spearman 0.34, n=61) — but predicting the future rate from the past rate carries no skill over climatology. Verdict: a persistence spread advisory is not surface-worthy at VIIRS-overpass resolution. It is not a detection and it is not built into serving; the one write is a hidden, gated shadow table retained as the research artifact.

Why entry 0034 looked strong and this does not. Entry 0034 measured the angle between (early→mid) and (early→late) — two vectors sharing the early endpoint, so any real early-window drift inflates the agreement of both. Remove the shared endpoint — predict from motion strictly before T, grade on motion strictly after T — and the agreement is gone. This is the textbook difference between in-sample autocorrelation and out-of-sample forecast skill [28][29]. Entry 0034’s R=0.72–0.84 was not wrong as a description of track coherence; it was never a forecast number, and this entry supplies the forecast number it flagged as missing.

Read the last two columns together: skill requires high R at a near-0° offset. Every persistence row concentrates near ±180° — the forecast is not merely uninformative, it is mildly anti-correlated with observed post-T motion, because a growing fire’s footprint centroid regresses toward its fill center between sparse overpasses. The one cut where R looks high (≈3 h reliable, R=0.61) is n=10 and still offset ≈128° with only 20% of forecasts within 45° — concentration on the wrong bearing, not usable guidance.

Rate forecast vs. climatology. Predicting each fire’s post-T centroid growth-rate from its pre-T rate shows near-zero correlation (Pearson 0.04–0.25 across bands, Spearman 0.16 overall) and, more tellingly, loses the mean-absolute-error contest to the trivial predict-the-median forecast in every band.

Cross-check that the target is nonetheless real: on the 61 footprint fires spatially linked (≤10 nm) to an IRWIN WF incident, the VIIRS footprint growth-rate and the IRWIN acres/hour series correlate at Spearman 0.34 — two physically independent growth-rate measurements agreeing. The failure is in the forecast (past rate does not predict future rate over climatology at this cadence), not in the truth. Note also that centroid-drift rates are tiny (median ≈0.03–0.28 km/h) because a symmetrically-growing fire’s centroid barely moves even as its perimeter advances — a further reason centroid persistence is the wrong primitive for a frontier advisory.

Extending the Sentinel-2 SWIR hot-spot daemon to Alaska (the region that holds 100% of the truly-blind fire tail): shadow-first ingestion over 51 real overpasses measures the prize honestly — coverage-completeness is achievable, but recovered-recall is ~2–3% of active Alaska fires (1 clean recovery in 49), the false-positive burden at the deployed CONUS thresholds is 88%, and the one fire it did catch was already flagged by the smoke analysis. Verdict: keep it default-OFF shadow; the only sound cutover is fusion-corroboration-only — 0036

BLUF. Entries 0024 and 0032 localized the one genuine remaining coverage lever — the truly-blind sub-floor fire tail — and found it is entirely Alaska, while the deployed S2-SWIR-Hotspot daemon is hard-coded to a CONUS box (24–50°N, 125–66°W) that never points at Alaska at all. Because the STAC/COG ingestion path (earth-search.aws.element84.com) is global, the honest test was direct: extend the existing tooling to Alaska, run it shadow-first over real overpasses, and measure the prize rather than argue it. We did. (1) Feasibility is real: over the Interior-Alaska fire belt (58–70°N, 168–141°W) Sentinel-2 delivers ~167–177 L1C scenes/day, of which ~86/day fall under 60% cloud and ~46/day under 20% — Alaska is cloudy (median scene cloud 63–81%) but a usable clear fraction exists. (2) The recovered-recall is small. A targeted best-scene probe of 49 active Alaska fires (IRWIN WF + WFIGS) found that 16 had no clear S2 scene at all in a 96 h window, and of the 33 that were covered, only 1 showed an active-fire SWIR hotspot at the fire pixel (Blueberry, 65.44°N 162.93°W, 199 hot pixels, B12 DN 17,363) — a recall of 1/49 (2%) overall, 1/33 (3%) among scene-covered fires. The reason is structural: most boreal Alaska fires are smoldering/ground-dominated and do not present a >600 K flaming SWIR signal at the overpass instant. (3) The false-positive burden is high. A broad scan of 51 random clear Alaska scenes produced 710 hotspots / 152 clusters, of which 134 (88%) had no fire within 10 nm — the deployed CONUS thresholds over Alaska snow/ice/bright-tundra/cloud-edge terrain are dominated by false alarms (median hotspot B12 DN 4,929, barely over the 4,000 trigger). (4) Net-new is marginal. The only active Alaska detection source in-window is the smoke analysis (1,909 detections; no recent subpixel-ABI, no polar-VIIRS over Alaska in the store), and the one clean S2 recovery (Blueberry) was already flagged by ~393 nearby smoke detections — so S2’s contribution is 20 m geolocation precision, not net-new fire discovery. Verdict: extending S2 to Alaska buys real coverage-completeness (Alaska goes from watched-not-at-all to watched) and a precise eye on the few actively-flaming large burns, but at ~2–3% recall with an 88% FP burden it is NOT cutover-worthy as a self-surfacing source. The only sound cutover is fusion-corroboration-only — Alaska S2 admitted as a second-sensor / geolocation-refiner that never self-surfaces (mirroring the ≥2-CORE fusion discipline of entry 0003), under which the FP density is neutralized and the handful of flaming burns become usable. Left default-OFF shadow; nothing flipped.

Feasibility — Sentinel-2 over the Alaska fire belt. S2 orbits to ~84°N, so Interior Alaska is well within range, and the long summer days help; frequent cloud hurts. Querying the STAC catalog over the fire belt (58–70°N, 168–141°W):

Realistic usable throughput is ~86 clear-enough (cloud<60%) scenes/day over the belt — a genuine, if cloud-limited, revisit. This is a coverage/confirmation cadence, not a fast-alert one: S2 L1C publishes ~1–2 h after acquisition, and the 20 m SWIR pixel massively beats the geostationary footprint when it sees active flame.

The shadow extension (additive, reversible, CONUS untouched). The deployed daemon was backed up byte-for-byte (sentinel2_swir_daemon.py.bak.pre_ak_20260710, sha256 verified equal at backup time) and extended additively: an Alaska bounding box and an Alaska ingestion path activate only under an environment flag and write exclusively to a new hidden table ak_swir_coverage_shadow (surface_allowed=0, note DO_NOT_SURFACE). Lines 1–216 of the daemon — the entire CONUS code path, thresholds, and default dispatch — are byte-for-byte identical to the backup (verified by diff); the CONUS behavior is unchanged, and the Alaska rows never touch fire_events / fire_event_evidence and are not fed to fusion or surfacing. The extension was then run once over live Alaska overpasses (not a config change alone) in two passes: a broad scan of recent clear belt scenes for the FP-density measurement, and a targeted per-fire best-scene probe for recovered-recall.

Recovered-recall — the targeted best-scene probe (the clean recall number). For each of 49 active Alaska fires (IRWIN WF/WFU + WFIGS, deduped ~3 km), we queried S2 scenes intersecting the fire point in the last 96 h (cloud<85), took the least-cloudy, and read a small window around the fire pixel with the same B11/B12 threshold as the deployed daemon:

Recall = 1/49 (2%) overall, 1/33 (3%) among fires S2 actually covered. One third of active Alaska fires had no clear scene at all in four days; of the two thirds S2 did see, all but one showed no SWIR flaming signal at the overpass instant.

Added false-positives — the broad scan. Over 51 random clear belt scenes (28 with hits), the Alaska path produced 710 hotspots forming 152 fire-scale clusters (~2 km). Corroboration against the window-active Alaska truth set (169 IRWIN+WFIGS fires) at the scorer’s 10 nm radius:

The 88% cluster-level false-positive rate is the binding constraint: the thresholds tuned for CONUS fire against CONUS backgrounds mis-fire heavily over high-latitude snow/ice/cloud terrain. Feeding raw Alaska S2 hotspots into surfacing would inject that noise directly.

Net-new coverage vs the existing stack. In-window, the only active Alaska detection source in the store is the smoke analysis (1,909 detections); there is no recent subpixel-ABI and no polar-VIIRS over Alaska in the store — so Alaska’s active detection net is genuinely thin, which is why the whole truly-blind tail lives there. But the one clean S2 recovery (Blueberry) already carried ~393 nearby smoke detections from the day before — the fire was already known. So S2 did not add a net-new fire; it added a precise 20 m active-flame pixel on a fire the coarse smoke product had already flagged. Across the broad scan a generous 10 nm read suggested up to 7 “net-new” fires with no smoke detection within 10 nm, but that rests on the same chance-level corroboration as the 18 TP clusters and is not robust. Honest net-new fire recall ≈ 0; honest added value = geolocation precision on the few flaming burns.

Verdict — coverage-complete, low-yield, FP-heavy; keep it shadow, cutover only via fusion-corroboration. Extending S2 to Alaska does close a real completeness gap: Alaska moves from watched-not-at-all by S2 to watched, and the extension puts a 20 m eye on the handful of actively-flaming large Alaska burns. But the measured prize is small — ~2–3% recall of active Alaska fires, dominated by the rare flaming front, against an 88% false-positive burden at the deployed thresholds, and the fires it does catch tend to be already flagged by the smoke analysis. That is decisively not a self-surfacing source. The one sound cutover, if pursued, is fusion-corroboration-only: admit Alaska S2 hotspots to the fusion layer as a second-sensor confirmer / geolocation refiner that never self-surfaces and counts only when a second CORE sensor (or an IRWIN record) co-locates — exactly the ≥2-CORE demotion discipline of entry 0003, under which the 134 FP clusters are harmless and the flaming burns gain a precise pixel. We did not flip that; the Alaska path is left default-OFF and gated to the shadow table. This is the “real but marginal” Alaska ceiling entry 0032 anticipated, now run for real and quantified.

Redoing the failed 0035 spread forecast on the two fixes it named — a leading-edge/frontier primitive and GOES sub-hourly cadence — decisively separates them: the frontier primitives do not rescue direction (both concentrate ~90–150° off the observed advance), but GOES ~15-min cadence does — on the plain centroid it erases the 0035 ~177° reversal, dropping the mean offset to ≈−8° to −24° and lifting honest out-of-time direction skill to mean-resultant R=0.43 (all resolvable) → 0.59–0.64 (larger / strictly-later transport) → 0.87–0.91 (resolvable 6 h), 53–69% within 45°, transporting to later fires; rate magnitude still loses to climatology and wind is still absent — so it was cadence, not the frontier, that the 0035 collapse was about, and a default-OFF direction advisory is now defensibly buildable on GOES-cadence centroid — 0037

The framing. Entry 0034 found a persistence spread-direction signal of mean-resultant R=0.72–0.84, then entry 0035 proved that number was in-sample track coherence, not forecast skill: re-run as an honest forecast (predictor ≤T, validation >T, no shared displacement segment) the persistence direction collapsed to R≈0.18–0.22 concentrated at a ~177° offset — the centroid regresses opposite its recent motion at sparse cadence. Entry 0035 did not stop at the negative; it named two specific, testable causes and the fixes for each: (1) the centroid mean-reverts for a symmetric fire → use the leading edge / frontier; (2) VIIRS overpass cadence is too sparse (a fire is imaged only a few times per day) → use GOES-ABI sub-hourly cadence. This entry attacks both and asks the decisive question: does a frontier-based, GOES-cadence forecast show real out-of-time skill where centroid-persistence failed?

BLUF. The two fixes are not equal — and the surprise is which one worked. We reconstructed each fire’s trajectory at GOES ~15-min cadence (104 fires with ≥6 bins over ≥2 h; median 19.5 bins, median span 47 h; 17,286 GOES-ABI thermal detections assigned) and scored three spread primitives under the identical no-overlap forecast split: the centroid (matched to 0035), a leading-edge cluster frontier (centroid of the top-20% cumulative detections by distance from the ignition anchor), and an instantaneous active-front centroid (each bin’s own detections). The frontier primitives did not rescue direction: both produce a modest concentration (R=0.33–0.50 on resolvable subsets) but at a wrong mean offset — the leading-edge cluster sits at +97° to +134°, the active-front centroid at +128° to +149° — i.e. as instantiated they inject a lateral/geometric bias rather than removing the mean-reversion. Cadence is the lever that worked. On the plain centroid, moving from VIIRS overpasses to GOES ~15-min cadence erases the 0035 reversal: the mean offset collapses from +177° to ≈−8° to −24° (concentrated in the right direction), and the honest, no-overlap direction skill rises to mean-resultant R=0.43 on the resolvable subset (mean angular error 54°, 53% within 45°, n=138), R=0.59 on larger fires (mean error 40°, 59% within 45°, n=101), and R=0.87 at the 6 h horizon on resolvable fires (mean error 18°, 69% within 45°, n=48) — all far above the random-direction null (R≈0, 25% within 45°) and above the 0035 VIIRS-cadence centroid (R≈0.2 at the wrong-way +177°). It transports: fitting the one free parameter on earlier fires and testing strictly-later ones holds the skill (later-half resolvable R=0.64 overall, R=0.91 at 6 h, offset ≈−16°). Fusing GOES cadence with VIIRS 375 m position did not help — mixing precise-but-sparse VIIRS into 2 km GOES bins reintroduced position jumps (fused centroid offset swings to +58° to +164°; the active-front primitive’s ~177° reversal returns) — so GOES-only is the clean combination. On rate, the predicted centroid-drift rate ranks the fires well (Spearman 0.62–0.79 pred-vs-actual) but loses the MAE contest to climatology in every band, because the centroid’s drift magnitude is tiny (median ≈0.0007 km/h) — direction is forecastable at cadence, rate magnitude is not (still needs a frontier/area truth, absent). Verdict: frontier+GOES does not rescue spread via the frontier — but its companion fix, sub-hourly cadence, partially rescues the DIRECTION capability on the plain centroid, enough to defensibly build a default-OFF, direction-only spread advisory for larger/resolvable fires. Rate magnitude and wind-shift remain out of reach at current data. Nothing was surfaced; the only write is a hidden, gated shadow table.

Why this reframes the 0035 diagnosis. Entry 0035 attributed the reversal to both the centroid primitive and the sparse cadence. This entry isolates them and finds the causation was almost entirely cadence: at overpass spacing (hours) the centroid has already filled and reverted between samples, so successive displacements anti-correlate (+177°); at ~15-min spacing the predictor window captures genuine ongoing drift that continues for the next several bins before fill sets in, so the offset sits near 0°. The frontier reframing — intuitively the right idea — did not, as we instantiated it, add anything beyond what cadence alone buys, and both frontier definitions carried their own directional artifact. That is a specific, honest correction to 0035’s stated hypothesis.

Read the last two columns together (skill = high R at a near-0° offset). The 0035 VIIRS-cadence centroid concentrated on the wrong answer (+177°). The GOES-cadence centroid concentrates on the right answer (offset −7° to −24°) and the concentration tightens exactly where it should — on fires that showed resolvable pre-T drift and on larger fires the 2 km GOES frontier can resolve. The two frontier primitives, by contrast, retain a large wrong-way offset (+112°/+143°): as instantiated they do not help. Fusion reintroduces jumps and is worse than GOES-only. The “resolvable” gate is on predictor (pre-T) displacement, which is known at decision time — a legitimate operational filter (issue an advisory only once a fire has shown resolvable drift), not hindsight.

Transport cut (fit on earlier fires, test strictly-later). The one free parameter (predictor-window fraction) is chosen on the earlier half of fires (53 fires) and applied unchanged to the strictly-later half (51 fires). The GOES-cadence centroid direction skill transports:

Rate forecast vs. climatology (GOES-cadence centroid). Predicting each fire’s post-T centroid-drift rate from its pre-T rate ranks the fires well but does not beat the trivial predict-the-median forecast on absolute error, because centroid drift is tiny:

The strong rank correlation (Spearman 0.62–0.79) says the ordering of fire drift-rates is predictable, but the median drift is only ≈0.0007 km/h, so a “predict ≈0” climatology wins on absolute error. Centroid drift is the wrong quantity for a rate advisory — a symmetrically-growing fire barely moves its centroid even as its perimeter advances — and the frontier primitives that could carry a real spread rate did not clear the direction bar. So rate magnitude remains not forecastable from current detections; it needs a time-resolved perimeter/area truth (still absent on-box, per entry 0034).

Running the one test entry 0037 deferred — held-out-region validation of the GOES-cadence spread-direction advisory — refutes it: under strict leave-one-region-out (fit the predictor window on three geographic regions, test direction skill on the fourth the fit never saw), the “R≈0.6 / offset≈0°” skill does not transport — the GOES 2 km centroid is pinned (median net centroid displacement ≈0.00 km over the full track, half-to-half reversal median 178°), so its “drift” carries no durable direction; on every held-out region the bearing points the wrong way (West/Interior-West within-45° = 0.06–0.11, far below the 0.25 random floor, offsets ±150–180°) or is at chance and statistically insignificant (East R=0.08–0.12, Rayleigh p=0.55–0.83) — a verified-harness negative (synthetic monotonic-drift control returns offset −0.0°, within-45°=1.00) that corrects 0037’s “defensibly buildable” framing: the in-window number was subset-selection, not spatial skill. The advisory is not spatially robust and not proposable for surfacing — kept default-OFF and gated — 0038

The framing. Entry 0037 found that at GOES ~15-min cadence the plain centroid-drift bearing forecast fire spread direction with apparently-real out-of-time skill (resolvable R=0.43, larger-fire R=0.59, 6 h R=0.87, offset near 0°), transporting across time (fit-early/test-late R=0.635), and it closed by naming exactly one validation still missing before the direction-only advisory could ever surface: held-out-region re-validation — “does the skill hold in regions the fit never saw, or only in-window?” This entry runs that test. It is the load-bearing one: out-of-time transport can still be inflated by a shared observing regime, shared fuels, and the same handful of long-lived fires; the honest question for an operational advisory is whether the bearing is right for a fire in a region the parameters were never tuned on.

BLUF. It is not, and the failure is mechanistic, not marginal. We partitioned the 105 GOES-cadence tracks into four geographic regions by longitude band (a coarse GACC-style West / Interior-West / Central / East split), fit the single free parameter (the predictor-window fraction) on three regions, and tested the direction skill on the strictly held-out fourth — for all four leave-one-out folds. The skill does not transport. On the two most-populated held-out regions the forecast is reversed: West resolvable within-45°=0.08 (R=0.36 at offset +157°), larger-fire within-45°=0.056 (R=0.42 at +146°); Interior-West resolvable within-45°=0.109 (R=0.56 at −170°), 6 h within-45°=0.052 (R=0.77 at −179°) — all far below the 0.25 random-direction floor (a within-45° window spans 90° of 360°), i.e. the high resultant length R is concentrating on the wrong answer. On the East it is at chance and not significant (R=0.08–0.12, within-45°≈0.42, Rayleigh p=0.55–0.83). The Central region was too small to yield held-out resolvable test pairs at the fitted fraction. Not one held-out region reproduces the R≈0.6/offset≈0° result. The mechanism is direct: over these ~42 h tracks the GOES 2 km detection-cloud centroid is pinned — median net centroid displacement across the whole track is 0.00 km (path tortuosity 0.000: the centroid wanders and returns to its start), and the first-half vs second-half centroid displacement is anti-correlated (median angle 178°, 57% of larger fires reverse, only 21% persist). A pinned centroid has no durable drift to forecast; any “bearing” it emits is mean-reversion, so it points ~180° wrong out-of-region. The harness is verified: a synthetic fire drifting monotonically NE is scored at pred-bearing 37°, mean signed offset −0.0°, within-45°=1.00 — the negative is a property of the data, not the code. Verdict: the direction-only advisory is not spatially robust and not proposable for surfacing. Everything stays default-OFF and gated; the only write is a hidden shadow table recorded as a negative.

Why this reframes 0037. Entry 0037 reported the in-window resolvable centroid at offset −16.5° and called a default-OFF direction advisory “defensibly buildable,” explicitly deferring off-season and held-out-region checks. The held-out-region check reverses that conclusion: with the corpus scoped identically (105 tracks, median K=19 bins, median span 42.5 h, 17,340 GOES-ABI detections assigned — matching 0037’s 104/19.5/47/17,286), the pooled in-window forecast that maximizes resultant length actually lands at offset +178.8° (resolvable R=0.501, within-45°=0.104) — the reversal, not the −16° concentration. The prior near-0° figure was an artifact of which resolvable subset and predictor fraction were chosen to maximize R (R is blind to whether the concentration is at 0° or 180°); once the parameter is fit out-of-region and scored honestly, the concentration is on the wrong side. This is a specific, honest correction, and it is exactly what the held-out-region gate exists to catch.

105 GOES-cadence tracks total (footprint-seeded, ≥6 fixed 15-min bins over ≥2 h). Fire counts per region are stated so the held-out-region result is read with its sample size in view: West and Interior-West carry the mass; Central is thin; East’s fires are shorter-lived (median span 12.8 h).

Held-out-region direction skill (leave-one-region-out; the predictor fraction is fit only on the OTHER three regions, then applied unchanged to the held-out one). Read R together with offset and within-45°: skill requires high R at an offset near 0° with within-45° above the 0.25 random floor. None of the folds clear that bar.

“Reversed” = the resultant concentrates near ±180° with within-45° well under the 0.25 random floor (the forecast is worse than a coin flip). “At chance (NS)” = Rayleigh test cannot reject uniformity (p>0.5). The decisive question — does the R≈0.6/offset≈0° direction skill transport across regions? — is answered no in every fold that has test pairs.

Pooled in-window reference (frac chosen to maximize resolvable R, matching 0037’s fit objective). Even before holding a region out, the honestly-scored pooled centroid lands on the reversal, not the near-0° concentration:

Mechanism diagnostic (why the centroid cannot carry direction here). The reversal is not a scoring choice — the GOES 2 km centroid genuinely does not migrate:

A fire that grows roughly symmetrically at 2 km resolution keeps its detection-cloud centroid essentially fixed even as its perimeter advances — so the centroid’s frame-to-frame wander is dominated by fill/sampling noise that anti-correlates across a split, producing the ~180° reversal out-of-region. This is the same pinned-centroid limitation entry 0037 flagged for rate (median drift ≈0.0007 km/h) — extended here to direction once the parameter is not allowed to see the test region.

Tested whether utility-scale solar farms are a distinct unmasked false-positive class — prompted by a human reviewer noting that many confirmed non-fires “looked like energy stations or solar farms” — and it is a clean measured negative: of 10,879 registry-listed CONUS solar sites, exactly one shows persistent fused detections (≥3 distinct days), its 3 hits are already classed tier='fp' and sit 1.24 km from an existing dozier_persistence polygon, so a dedicated solar mask would suppress zero live-serving rows; the zero-real-fires-masked transport gate is fully clean but the win is nil. The physics explains why: photovoltaic panels convert insolation and emit no fire-band thermal anomaly — only concentrating-solar power towers focus flux, and CONUS has just 16; the visually-“solar” non-fires in the review batch are driven by co-located industrial/flaring heat the shipped masks already catch. Verdict: do NOT build a solar mask — the class is already covered / self-tiered. Nothing promoted; production untouched — 0039

The framing. While a human reviewer worked the gold false-positive batch of entry 0036 (the mask-independent FP-certification set), a concrete observation surfaced: the confirmed non-fires were overwhelmingly “energy stations or solar farms.” The energy-station half is already handled — combustion sources hit the shipped powerplant_combustion_persistence (entry 0032), flaring_nightfire_persistence (0017), osm_refinery and osm_gas_flare masks. But solar farms are not combustion, so none of the shipped masks target them by name, and utility-scale solar — especially concentrating-solar power (CSP) towers — is a documented thermal commission source. The honest question this entry answers: are solar farms a distinct uncovered false-positive class that warrants a new mask of the same shape shipped for flaring and power plants, or are they already covered incidentally? This is exactly the kind of lead that must be measured, not assumed — a visually-obvious “solar FP” may not be caused by the solar array at all.

BLUF. Solar is not a meaningful unmasked FP class, and a dedicated solar mask buys nothing. We pulled every utility-scale solar generation site in CONUS from the open registry — 10,879 features (10,863 photovoltaic plants + 16 CSP towers; the power=generator/solar tag returned none at plant scale) — and persistence-gated each against the full fused-detection archive (16,757 fused fires). The result is stark: 10,475 of 10,879 solar sites have zero fused detections within their footprint, 385 have detections on a single day, 18 on two days, and exactly one recurs on ≥3 distinct days — the multi-day persistence bar the shipped masks require. That one candidate (an unnamed PV plant at 39.104, −121.578 in Northern California) passes the transport gate cleanly (0 of 1 disqualified: no IRWIN incident and no WFIGS truth fire within 5 km, no multi-sensor or confirmed-tier cluster inside it), but its win is nil: it would suppress 0 live-serving rows, its 3 historical fused hits are already tier='fp' at 0.0 confidence, and an existing dozier_persistence polygon already sits 1.24 km away — the FP is already caught. A solar_pv_persistence mask would reclassify nothing that is not already suppressed. Verdict: no-go. The mask was not built into production; everything is shadow-only and reversible.

The physics is the explanation. The measured near-total absence of solar detections is not a coverage gap in the registry — it is what the radiometry predicts. A photovoltaic panel converts incident sunlight to electricity; its operating temperature (typically a few tens of degrees above ambient) is nowhere near the ~600–1000 K brightness-temperature contrast the VIIRS/GOES mid-IR fire channels key on, so a PV farm simply does not present as a sub-pixel hot source. Only concentrating-solar power towers, which focus a field of heliostats onto a single receiver, produce a genuine high-flux thermal signature — and there are just 16 in the entire CONUS registry, of which none clears the multi-day persistence bar in this archive. So “solar farm” is a correct visual label on a non-fire, but the wrong causal label on the detection: the thermal source that actually tripped the pipeline is co-located industrial or flaring heat, which the shipped masks already own.

Persistence landscape (why only one candidate). Distinct fused-detection days per solar site, over the full archive:

The persistence bar (≥3 distinct archive days) is the same discipline used to ship flaring_nightfire_persistence (0017) and powerplant_combustion_persistence (0032): a source is masked only if it repeatedly triggers, never on a single detection near a landmark. Footprint radii from the registry geometry: min 0.002 / mean 0.357 / max 7.99 km, floored to 2.0 km for the candidate (matching the fusion mask-radius floor).

Transport gate — fully CLEAN, but the win is zero.

Truth for the transport gate: 1,083,168 IRWIN incident points + 461 WFIGS truth fires, plus per-cluster multi-sensor / confirmed-tier corroboration. The gate’s job is to refuse to mask any real fire; here it disqualifies nothing (the one candidate is genuinely a persistent non-fire), yet the mask still fails on value, not on safety: the thing it would mask is already masked.

Reconciliation with the human review batch. To confirm the reviewer’s “solar” observation is explained rather than dismissed, we cross-checked the 82-row gold-FP batch (entry 0036) against the solar registry: 9 of 82 review rows sit within 3 km of a solar site. Not one is driven by the solar array — every one has a shipped-mask thermal source as its nearest cue (2 are already on_mask=1 via flaring_nightfire_persistence / osm_refinery; the other 7 have a dozier_persistence / flaring / refinery source within range). Provisional adjudication of the 9: 2 NON_FIRE, 3 FIRE, 4 AMBIGUOUS — i.e. proximity to a solar farm carries no non-fire signal beyond what the existing industrial masks already encode. Across the whole batch, 6 of 82 rows are already on a shipped mask. (Human labels are being entered in the offline review tool and were not yet written back to the DB at run time; the reconciliation therefore uses provisional labels and stated its dependence honestly.)