The Gate Held. The Model Did Not.
Out-of-sample validation of the EWS-U3 early-warning model, executed 28 September 2026 · MINERVA M-3 (Ratings & Calibration). Four tests pre-registered and frozen before the run; data hashes committed 26 Sep; the FORECAST years sealed from every fitting step. The model did not clear its gate, and the failure publishes in full — that was the commitment, made before anyone knew it would be needed.
Statement of Pre-Registration
Every test below was fixed in the frozen protocol (EWS-U3-OTH-PROTOCOL v1.0) before the model saw a single out-of-sample row. The harness fits once on the training window and never reads the forecast years; the frozen artifacts were hash-committed on 26 September, two days before the run. The run got to edit nothing but the numbers.
Frozen artifacts (sha256, first 16): proxy scores v1.2 78a39832dcc12260 · headline matrices c464564c · 663b4aa4 · 08eb2d60 · c121e85b · labels v3 ec20bf1b19e5a60e · harness v2 b1771c8b6a344d51 · freeze manifest da84a012a68d1993 (26 Sep, 87 files — verified intact)
The Frozen Tests — and What the Run Returned
| Test | Frozen criterion | Observed | Verdict |
|---|---|---|---|
| T1 — Transition discrimination | Transition-event F1 ≥ 0.35 and Brier win vs. the Markov baseline | Best transition F1 0.228 (raw-feature RF); Markov 0.000. Brier: best model 0.210 vs. Markov 0.187 — no model beat the baseline. | FAIL |
| T2 — Spectral contribution | Spectral features ≥ +1pp on the scoring rule, or the layer is demoted | Spectral Δ = −3.93pp; the PGE proxy column also negative at −1.9pp (the C-M5-1 arm, reported pass or fail as committed). | Spectral DEMOTED to interpretive frame |
| T3 — Beats naïve | +5pp over the best naïve baseline on transition F1 | Not reached — best transition F1 0.228 against the frozen floor. | FAIL |
| T4 — Calibration | Calibration bins within 0.15 tolerance | Worst bin gap 0.141 — inside tolerance; the model was under-confident in all four bins (empirical accuracy exceeded stated confidence throughout). | PASS |
Internal texture, declared and non-certificational: on the training-window holdout the sticky-label pathology was visible (Markov accuracy 0.923 at transition-F1 0.000) and the full model's transition-F1 printed below the T1 threshold. The signal was there before the run; we ran the gate anyway, because a gate you only hold when you expect to pass is not a gate.
The Pre-Run Gates
| Gate | What it proves | Result |
|---|---|---|
| 1 — Time-safety scan | No future information in any artifact; FORECAST seal intact (report fa74d1f34ede6393) | PASS |
| 2 — Adjudication | Label reproducibility κ = 0.942 against the 0.75 gate (70 disagreements, every one resolved by rule citation) | PASS |
| 3 — Provenance | 34/34 series carry documented sources | PASS |
| 4 — Harness integrity | Synthetic pathology panel reproduces all four certified findings under the final code (report 4f47abab5cba14ba) | PASS |
The failure is the model's, not the machinery's: all four integrity gates passed. The instrument measured honestly and reported a negative. That distinction is the entire asset.
The κ Caveat — Unchanged by the Verdict
The κ = 0.942 measures reproducibility of the frozen labeling text under two documented independent readings by one operator — a reproducibility ceiling, not human dual-adjudication. Human dual-adjudication remains pre-registered as a verification amendment: sample, blind sheets, and log ready unchanged. A failed run does not retire the caveat; if anything it sharpens the question of whether the labels, not the model, are the thin layer.
What Failure Costs — On the Record
The model's early-warning tiers are demoted: what was planned as a certified instrument reverts to what it actually is — an unproven signal, and it will be described that way everywhere it appears. The Early Warning Feed's 28 Feb 2027 launch window converts to a written soft-landing finding unless a re-run earns the gate. What does not move: the Calibration Ledger, the corrections register, and every scored claim published to date — those were never the model's claims. They were ours, and they stand. Specific consequences, executed under the pre-committed soft-landing (protocol §8 / manuscript §5.3): the non-commutative dynamics are retired; the Hilbert-space formalism is abandoned; the theory continues as a classical network-dynamics + constructal-thermodynamics framework, retaining the graph Laplacian as classical wave dynamics. The spectral layer is demoted to an interpretive frame (T2), and the PGE proxy column — which also failed to earn its place (−1.9pp, C-M5-1) — is retired from the scoring stack. The rebuild/re-run path: the raw-feature arm was the best performer at this gate, exactly as the internal RC-24 ablation predicted (+0.32pp for spectral; raw alone ahead) — the rebuild starts classical, on the raw arm, with the label-stickiness problem named as the first engineering target. The pre-registered human dual-adjudication amendment (§3) proceeds unchanged — a failed run sharpens, rather than retires, the question of whether the labels are the thin layer. Any re-run faces the full gate cycle again — five gates, a fresh freeze, a new appointment. No date is promised here; a date would be a guess, and this page does not guess.
The Honest Paragraph
What failed first was the thing the model was built to do: see transitions. Out of sample, over the horizon, the sticky-label crutch removed, no arm reached the threshold — the best transition F1 printed 0.228 against a frozen 0.35, and no model beat the dumbest baseline on the scoring rule. What we will change: the formalism (the quantum-analogy machinery retires, as pre-committed — it was always declared analogy, and the gate priced it), and the feature stack (every addition beyond the raw features made the model worse; the build follows the evidence back to the simple arm). What we will not touch: the Calibration Ledger, the corrections register, the four integrity gates that measured this failure honestly, the labels' audit trail, and the habit that built all five. One result deserves its own sentence: the model passed calibration — under-confident in every bin. The instrument knew what it did not know; it simply did not know enough. We publish the failure because that was the deal, made before anyone knew it would be needed — and because a gate you only hold when you expect to pass is not a gate. The negative is the instrument working. Veritas per ledgere.