Issue #414 — interval evidence disposition

Disposition: blocked; no interval claim

No interval scores, calibration fits, coverage estimates, widths, proper scores, or calibration parameters were computed. There is no evidence-based diagnosis that the production intervals are miscalibrated: the prerequisite #411 eligible, matured forecast lineage is absent, and the #409 canonical corpus is blocked. This is an evidence-readiness disposition, not evidence for or against interval calibration or probabilistic skill.

Point forecasts and their production behavior are unchanged. No interval adjustment, point forecast change, performance metric, or model winner is proposed or selected. Every interval metric is unavailable (null), rather than zero.

Evidence captured from current #411 report

The sole source for the snapshot facts and evidence hashes below is ISSUE_411_PARITY_BASELINES_BLOCKED.json. That report records its latest verified prospective snapshot as of 2026-09-27T11:00:00Z:

• 48 timestamp-eligible forecasts; all 48 are legacy/unversioned and excluded.
• 0 confirmatory forecasts, 0 expected pairs, and 0 matured pairs.
• The latest provenance-complete origin, 2026-09-27T11:00:00Z, was not mature for the cohort at that snapshot.
• Prospective and historical-canonical tracks are blocked and their metrics are null. The canonical corpus gate is blocked; the frozen prospective cohort is not ready.
• The recorded blockers include versioned-provenance mismatch and no confirmatory versioned forecasts in the frozen cohort.

Hashes are reproduced exactly from the #411 report; no new artifact, database, code, corpus, or cohort hash is asserted here:

#411 evidence	SHA-256
#411 audit code	79a95fc3debd4134ab68cfc55795007a340f15a6eff2ef7212a875a9846c28fc
Snapshot database	b102f14db535264ecc2474a3b338c929abea4c39616f233a5701611c9300f57f
Last verified frozen-cohort content hash in snapshot	462287eb433a27c16e4f25360d9ee40e5dda2fabdced13e5390859a055f7bbb8
Frozen #410 protocol file	1b8c6705d53882a66c0ab1277cf2a63c845d00c2c7d0cc978c0c75d8121efd18
Frozen #410 protocol content	4da936d505df9a15cbd073ee85ee0e4ed2db483e6058f6b81ce163f4724bb4f0
#397 corpus audit	e2bf0c671a5a4327c6852d02614a7d2e403ed12f8fe7c4362336c5e87d44a0c7

The #411 report has no cohort path/hash or ledger path/hash for this snapshot; none is inferred. These hashes identify evidence referenced by #411 and do not turn blocked or missing evidence into eligible observations.

Why no interval claim is possible

Interval diagnosis requires observations with versioned, provenance-complete production forecast lineage and exact, matured outcomes. The current #411 snapshot supplies no confirmatory forecast origins or eligible pairs. The timestamp-eligible legacy rows cannot establish which forecast configuration and interval-generation policy produced them. The newest complete origin has not matured. Separately, the #409 corpus is blocked, so it cannot supply a canonical historical evaluation. There is consequently no admissible sample on which to estimate calibration, evaluate proper scores, or diagnose miscalibration. Null metrics mean not evaluated, not a failed or successful interval result.

Exact re-open gates

Re-open interval evaluation only after all of these gates are met and the evidence is frozen before outcome-score access:

1. Eligible matured lineage (#411): produce immutable, versioned production-policy forecast records with policy/configuration identity, point-in-time provenance, exact origin/target/horizon keys, recorded interval quantiles, and exact matured outcomes for the same vintage. The frozen cohort must report nonzero confirmatory origins and matured eligible pairs; legacy/unversioned rows remain excluded. Re-run #411 and preserve its report and source evidence hashes.
2. Canonical corpus (#409): resolve its documented readiness requirements and obtain an eligible point-in-time corpus audit with complete exact target and warm-up coverage, source/vintage provenance, and no unresolved gaps, duplicates, or errors. Keep historical-canonical and prospective evidence distinct.
3. Frozen interval estimands and quantiles: before scoring, freeze the interval-producing policy and quantile levels. For the current #410 protocol, these are cumulative target log-return q10, q50, and q90; L=q10 and U=q90, with finite ordered bounds. Do not fit or retune quantiles on evaluation/test outcomes. Any different quantiles or estimand require a separately versioned, hashed protocol frozen before score access.
4. Prior-only calibration: if calibration is proposed, fit its parameters using only data available before each evaluation origin (nested, walk-forward/prior-only fitting); record fit windows, input lineage, and parameter versions. Do not use evaluation, holdout, or D3 outcomes to fit, select, or revise calibration.
5. Frozen scoring protocol: register and hash the protocol, forecast candidates, exact pairing/outcome vintage, cutoffs, strata, complete test family, and analysis procedure before outcome-score access. Under frozen #410 v11-prospective-1.0.1, score exact matured pairs using per-quantile pinball loss and the primary 80% Winkler score, alongside 80% coverage and interval width; coverage alone is not a proper-score or skill claim. Preserve attempted/failed/abstained/missing denominators and apply its paired 24-origin moving-block bootstrap and family-wise Holm adjustment to the entire registered family. Any changed method requires a new protocol version and digest before a new independent evaluation.
6. Disjoint evaluation and D3: reserve the chronological test/holdout disjoint from all calibration fitting and model/quantile selection. Keep the prospective D3 phase separate and untouched by training, calibration, or protocol changes; the frozen protocol requires at least 30 calendar days and 200 mature exact pairs per horizon for D3. Do not reuse test or D3 outcomes for calibration or claim confirmation after tuning.

Passing readiness gates authorizes a separately frozen evaluation; it does not pre-judge its metrics, establish miscalibration, or select a winner.