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).