2024 / HELENE + MILTONOriginal missing
Florida's critical storms
We have the official outcome records but not the original independently dated Poseidon forecast files. Later reconstruction is not a 2024 operational win.
CPX-046 / FORECAST ACCOUNTABILITY
Prediction is a claim.
Evidence is the proof.
No retrospective relabeling. No erased misses. No claimed early-warning advantage without a preserved forecast that an outside party could have seen before the event.
Published evidence register · Updated November 13, 2026 · Research-stage, not official warnings
01 / SOURCE HISTORY
Our past. With its limitations intact.
Forecast verification requires the actual prediction, when it became externally visible, what was knowable then, and a common scoring rule. These cases reflect what we have recovered, not what we wish we could prove.
2025 / ATLANTICOriginal missing
No invented forecast history
Official season outcome: 13 named storms · 5 hurricanes · 4 major hurricanes. The original Poseidon 2025 forecasts have not been recovered with a verified pre-event timestamp, so no accuracy score is claimed.
2026 / OCT 7 WORKING RECORDInternally sealed
Real forecasts. Honest scoring.
Archived seasonal forecast: 10 named · 1 hurricane · 0 majors. As of November 13 the NHC reports 10 named · 1 hurricane · 1 major. The first two fields match provisionally; the major-hurricane field misses by one.
Archived Isaias outlook: peak 85–90 kt, Category 2. Observed peak: 105 kt, Category 3. This was an intensity underestimate; it cannot be presented as a perfect call.
The internal SHA-256 seal is preserved. The original October 7 publication time has not been externally authenticated; this is not an independently certified advance forecast.
2026 / OCT 10 REVISIONAnalyst update
A revision is not a prior prediction.
The October 10 analyst scenario included one major hurricane only after Isaias became a major hurricane. That field was already known, so it does not count as an early correct prediction.
Research archive preserved; original production engine/weights not independently recovered.
02 / FORWARD PROOF, 2027–2030
Pre-register the method. Then face the future.
The official 1851–2025 NOAA Atlantic HURDAT2 record has now been downloaded, SHA-256 pinned and independently rerun through our isolated research build. This is a historical truth archive, not proof of old Poseidon-issued forecasts. The next genuinely forward predictions must carry independent pre-event receipts.
OFFICIAL NOAA / 1851–2025Source verified
175 real historical seasons
NOAA/NHC original archive SHA-256: 75202780ce56c8458f16fbef5fb8a026b99b00878dbb95c458211b3e6ef125b4
Reconciled official season counts for 1851, 2024 and 2025; source revision September 23, 2026. Historical replay is retrospective, not pre-event forecasting skill.
FIXED CLIMATOLOGY vs CSU / 2004–2023No proven win
Deficit stays in the record
Real NOAA source replay: prior-30-year climatology 5.14 named storms MAE; archived CSU April forecasts 4.90 on 20 matched seasons. Paired 95% bootstrap improvement interval: −1.47 to +1.01 storms.
CSU source transcription and issue vintages still require independent audit. This is a simple climatological model, not the proposed multisensor Poseidon engine; no NOAA/CSU superiority claim.
03 / NO-HINDSIGHT CONTROL
From now on, the original goes first.
Each prediction must cross the issuance gate before the outcome is measured. Neither a backfilled date nor an internal hash by itself establishes when a forecast was made.
01
Freeze the actual call
Publish the prediction, issue time, horizon, affected geography, expected outcomes, underlying information cutoff, model/version hash, and every probability or stated range.
02
Obtain outside time evidence
Publish the exact file hash with an independently dated receipt from a public deployment, provider-signed log or trusted timestamp service. Preserve the original unchanged.
03
Score all outcomes
After verification, grade every valid forecast and non-event, including missed events and false alarms, against identical-timestamp NHC/NWS baselines. Show sample sizes and error distributions.
04
Append, never rewrite
New information generates a new forecast version with a new publication receipt. It never changes the old forecast or its score.
Evidence levels: A — independent pre-event witness; B — recovered internal seal, chronology not independently attested; C — post-event analysis; U — original missing. Only independently witnessed, properly paired A-level records qualify for proven forward-skill claims.