lusoris · Jun 13, 2026 · Jun 13, 2026
@@ -0,0 +1,10 @@
+- **`fr_regressor_v2_ensemble_v1_seed{0..4}` registry rows: restored
+  dropped `license` / `license_url` / `sigstore_bundle` metadata and
+  made the smoke/production state honest.** PR #865 regenerated the five
+  ensemble ONNX files in smoke mode to fix a `codec_vocab` 14→6 input-dim
+  mismatch, but also dropped the per-row license metadata and left the
+  rows labelled neither production nor consistently smoke. The license
+  fields are restored, `smoke: true` now truthfully describes the shipped
+  weights, and the production flip is deferred to the one-shot post-RC
+  retrain (ADR-1105) with a strict-xfail tracking marker on
+  `test_fr_regressor_v2_ensemble_seed_rows_are_production`.
@@ -0,0 +1,135 @@
+<!-- markdownlint-disable MD013 MD060 -->
+# ADR-1105: `fr_regressor_v2_ensemble` production flip deferred to the one-shot post-RC retrain
+
+- **Status**: Accepted
+- **Date**: 2026-06-13
+- **Deciders**: Lusoris
+- **Tags**: `ai`, `models`, `rc`, `docs`
+
+## Context
+
+[ADR-0321](0321-fr-regressor-v2-ensemble-full-prod-flip.md) flipped the
+five `fr_regressor_v2_ensemble_v1_seed{0..4}` rows in
+`model/tiny/registry.json` from smoke to production: it shipped real
+LOSO-validated ONNX weights (gate verdict `PROMOTE`, mean PLCC ≈ 0.997),
+per-seed sidecars, and `smoke: false`. Those weights were trained with a
+codec one-hot of width 14 (`codec_vocab` = 12 codec entries + 2 norm
+dims).
+
+`codec_vocab` was subsequently trimmed to 6 (`x264`, `x265`,
+`libsvtav1`, `libvvenc`, `libvpx-vp9`, `unknown`; see
+`model/tiny/fr_regressor_v2_ensemble_v1.json`). This made the
+production ONNX input dimension stale: a model expecting `[batch, 14]`
+can no longer be fed the current `[batch, 6]` codec one-hot, which
+surfaced as an `eval_probabilistic_proxy.py --smoke` load failure.
+
+PR #865 fixed the load path by regenerating the five ONNX files at the
+correct `[batch, 6]` width, but it did so with the trainer's `--smoke`
+mode — one epoch on a synthetic corpus, i.e. throwaway placeholder
+weights, not a production fit. The registry was correspondingly set to
+`smoke: true` with new sha256 values. Two side effects of that PR were
+not intended:
+
+1. The `license`, `license_url`, and `sigstore_bundle` fields were
+   dropped from the five rows. This is a pure regression — every
+   registry entry (smoke or not) must carry license metadata, and the
+   `test_every_entry_has_license_metadata` invariant enforces it. The
+   fields are restored unconditionally in this PR.
+2. The five rows now claim neither production nor a consistent provenance
+   record: the on-disk ONNX are smoke (new sha), while the retained
+   per-seed sidecars still describe the older `[batch, 14]` production
+   weights (old sha, `PROMOTE`). The
+   `test_fr_regressor_v2_ensemble_seed_rows_are_production` invariant
+   (added by ADR-0321) consequently fails on `smoke is False`.
+
+Producing real production weights at `codec_vocab = 6` requires
+re-running `export_ensemble_v2_seeds.py` against the corpus — a full
+retrain/re-export. The operator has locked all model retraining into a
+single one-shot step to be run only after the toolchain reaches RC and
+the feature numbers are frozen, explicitly to avoid retraining models
+repeatedly. The ensemble is in scope for that one-shot retrain. Doing a
+piecemeal ensemble-only retrain now would contradict that decision and
+risk shifting numbers that the one-shot run is meant to freeze.
+
+## Decision
+
+For the release candidate, ship the ensemble seed rows honestly as smoke
+placeholders and defer the production flip to the locked one-shot retrain:
+
+1. Restore `license`, `license_url`, and `sigstore_bundle` on all five
+   rows (the #865 regression). Keep `smoke: true` and the regenerated
+   `[batch, 6]` sha256 values, which match the ONNX actually shipped.
+   Update each row's `notes` to state plainly that these are smoke
+   placeholders pending the one-shot production re-export.
+2. Keep the `test_fr_regressor_v2_ensemble_seed_rows_are_production`
+   assertions verbatim (they remain the target production contract), but
+   mark the test `@pytest.mark.xfail(strict=True)` with a reason citing
+   this ADR. `strict=True` means the test fails the suite the moment the
+   one-shot retrain lands real weights (`smoke: false` + a sidecar whose
+   sha256 matches the shipped ONNX), forcing removal of the marker — so
+   the deferral cannot silently outlive its cause.
+3. When the one-shot retrain runs, it must re-run
+   `export_ensemble_v2_seeds.py` so the ONNX bytes and sidecars
+   regenerate together at `codec_vocab = 6`, then flip `smoke: false` and
+   remove the xfail marker — exactly the workflow ADR-0321's follow-ups
+   already mandate (hand-flipping rows remains forbidden).
+
+This does not modify any Netflix golden-data assertion (CLAUDE.md §8);
+`model_registry_schema_test.py` is a fork-local file.
+
+## Alternatives considered
+
+- **Retrain the five ensemble seeds now (production flip immediately).**
+  This is the eventual correct end state but contradicts the locked
+  one-shot retrain decision (no piecemeal retraining before the
+  toolchain is RC-frozen) and would risk moving numbers the one-shot run
+  is meant to freeze. Rejected for RC; it is precisely what the one-shot
+  retrain will do.
+- **Revert the ONNX to the old `[batch, 14]` production weights.**
+  Restores `smoke: false` consistency but re-breaks the load path under
+  the current `codec_vocab = 6`, reintroducing the
+  `eval_probabilistic_proxy` failure #865 fixed. Rejected.
+- **Leave the test failing as a known local red.** The schema test is
+  non-gating in CI (it runs under the `|| true` block in
+  `tests-and-quality-gates.yml`), so this would not break master. But a
+  bare red is noise that can mask a future regression in the same test
+  and carries no self-healing signal. The strict-xfail marker is the
+  honest, self-documenting, auto-alerting representation. Rejected in
+  favour of strict xfail.
+- **Delete the stale per-seed sidecars.** They describe the older
+  production weights, not the shipped smoke ONNX. Keeping them preserves
+  the genuine `PROMOTE` provenance and the sha the one-shot retrain will
+  supersede; the only consumer is the now-xfailed production test.
+  Rejected (kept) to retain provenance.
+
+## Consequences
+
+- **Positive**: The RC registry is internally consistent and honest —
+  every entry has license metadata, and `smoke: true` truthfully
+  describes the shipped weights. The deferral is tracked by a strict
+  marker that fails loudly when resolved.
+- **Negative**: The probabilistic ensemble head ships at smoke quality in
+  the RC. Any consumer that loads it gets placeholder predictions until
+  the one-shot retrain. This is documented in the model card and state.md.
+- **Neutral / follow-ups**: The one-shot retrain must (a) re-export the
+  five seeds via `export_ensemble_v2_seeds.py` at `codec_vocab = 6`,
+  (b) flip `smoke: false`, (c) remove the xfail marker in
+  `model_registry_schema_test.py`. Tracked in `docs/state.md` and the
+  retrain plan.
+
+## Supply-chain impact
+
+- **New dependencies**: none.
+- **Removed dependencies**: none.
+- **Build-time fetches**: none.
+
+## References
+
+- Parent / superseded-context: [ADR-0321](0321-fr-regressor-v2-ensemble-full-prod-flip.md),
+  [ADR-0303](0303-fr-regressor-v2-ensemble-prod-flip.md),
+  [ADR-0309](0309-fr-regressor-v2-ensemble-real-corpus-retrain.md).
+- Regression source: PR #865 (`8b7ae731a`) — dropped license metadata and
+  regenerated ONNX in `--smoke` mode.
+- `req` — operator direction: retrain all models exactly once, after the
+  toolchain reaches RC and feature numbers are frozen, to avoid repeating
+  the retrain; the ensemble is in scope for that one-shot run.
@@ -914,3 +914,4 @@ ADRs may exist there for local session continuity, but the tracked
 | [ADR-1102](1102-phase4b9-container-only-publishing.md) | Phase 4b.9 container-only canonical artifact policy: the vmaf-dev-mcp container is the exclusive source for release binaries, published container images, and CI benchmark/snapshot artifacts. Host-side builds are diagnostic-only (IDE/clangd, debugger, sanitizer sweeps). Extends ADR-0496 to the publishing surface and adds docs/development/publishing.md. | Accepted | 2026-06-08 | container, build, release, publish, phase4b, docs-policy, fork-local |
 | [ADR-1103](1103-hip-vif-mirror2-boundary.md) | Fix integer_vif_hip boundary condition: clamp_i → mirror2_i. The HIP integer VIF kernel used clamp boundary mode instead of mirror2 (the CPU reference), causing parity failures at the frame boundary for non-trivial content. Supersedes the ADR-0566 claim that clamp matches CPU within places=4. | Accepted | 2026-06-13 | hip, vif, parity, boundary, correctness, fork-local |
 | [ADR-1104](1104-float-vif-avx512-golden-regression-fix.md) | Remove AVX-512 dispatch from float VIF convolution to restore Netflix golden scores. ADR-0504 added AVX-512 float convolution dispatch but the wider FMA partial-sum tree produces different rounding than AVX2, causing the Netflix golden VMAFEXEC_score assertion (76.66740433333332, places=4) to fail on AVX-512 CPUs. Fix: remove HAVE_AVX512 dispatch blocks from vif_filter1d_s/sq_s/xy_s; float VIF uses AVX2 path matching upstream Netflix/vmaf. All 271 golden tests pass after fix. | Accepted | 2026-06-13 | simd, correctness, float-vif, bug-fix, fork-local |
+| [ADR-1105](1105-ensemble-v2-prod-flip-deferred-oneshot-retrain.md) | Defer the `fr_regressor_v2_ensemble` production flip (ADR-0321) to the locked one-shot post-RC retrain. The ADR-0321 production weights were LOSO-validated at codec_vocab=14; the vocab was trimmed to 6, so #865 regenerated the ONNX in smoke mode (smoke=true) to keep the load path correct. For RC: restore the license/license_url/sigstore_bundle fields #865 dropped, keep smoke=true (matches shipped weights), and mark `test_fr_regressor_v2_ensemble_seed_rows_are_production` xfail(strict=True) so it auto-fails the moment the one-shot retrain lands real weights. No Netflix golden assertion touched. | Accepted | 2026-06-13 | ai, models, rc, docs, fork-local |
@@ -8,27 +8,37 @@ form _"give me the CRF where the **lower** bound of the 95 % interval is
 still ≥ 92"_ — driving the new `vmaf-tune --quality-confidence` flag
 (planned, see [ADR-0237](../../adr/0237-quality-aware-encode-automation.md)).
 
-> **Status — production (per-seed `smoke: false`).** As of 2026-05-06
-> the five `fr_regressor_v2_ensemble_v1_seed{0..4}` rows in
-> `model/tiny/registry.json` carry **production** ONNX weights:
-> trained on the 5,640-row Phase A canonical-6 corpus (9 Netflix
-> sources × `h264_nvenc`) after the 9-fold LOSO ship-gate cleared
-> with mean PLCC 0.997 (spread 0.001) per
-> `runs/ensemble_v2_real/PROMOTE.json`. Each non-smoke ONNX has a
-> matching sidecar `fr_regressor_v2_ensemble_v1_seed{N}.json` with
-> the canonical encoder-vocab-v2 + codec-block layout + training
-> recipe + per-seed gate evidence. Fresh seed exports also include ADR-0661
-> `run_provenance` with the corpus, PROMOTE verdict, argv, output seed
-> sidecars, and optional registry target. The shared
-> `fr_regressor_v2_ensemble_v1.json` manifest is unchanged. See
+> **Status — smoke placeholder for the RC; production flip deferred to
+> the one-shot post-RC retrain ([ADR-1105](../../adr/1105-ensemble-v2-prod-flip-deferred-oneshot-retrain.md)).**
+> The five `fr_regressor_v2_ensemble_v1_seed{0..4}` rows in
+> `model/tiny/registry.json` currently carry `smoke: true`. The
+> ADR-0321 production flip (2026-05-06) shipped LOSO-validated weights
+> (mean PLCC 0.997, spread 0.001 per `runs/ensemble_v2_real/PROMOTE.json`)
+> trained against a codec one-hot of width 14. `codec_vocab` was later
+> trimmed to 6, which made those weights' input dimension stale; PR #865
+> regenerated the on-disk ONNX at the correct width but in `--smoke`
+> mode (1 epoch, synthetic corpus) to keep the load path working. Those
+> smoke weights are placeholders, not a production fit. Re-establishing
+> production at `codec_vocab=6` requires re-running
+> `export_ensemble_v2_seeds.py`, which is part of the locked one-shot
+> post-RC retrain (the ensemble is in scope). Until then,
+> `test_fr_regressor_v2_ensemble_seed_rows_are_production` is marked
+> `xfail(strict=True)`; it auto-fails the suite the moment the retrain
+> lands real weights (`smoke: false` + matching sidecar sha), forcing
+> removal of the marker. The per-seed sidecars
+> (`fr_regressor_v2_ensemble_v1_seed{N}.json`) still describe the older
+> production weights and retain their PROMOTE provenance; the one-shot
+> retrain regenerates ONNX + sidecars together. See
 > [ADR-0303](../../adr/0303-fr-regressor-v2-ensemble-prod-flip.md)
 > (gate definition),
 > [ADR-0309](../../adr/0309-fr-regressor-v2-ensemble-real-corpus-retrain.md)
 > (separate-PR rule),
 > [ADR-0319](../../adr/0319-ensemble-loso-trainer-real-impl.md) (LOSO
-> trainer), and
+> trainer),
 > [ADR-0321](../../adr/0321-fr-regressor-v2-ensemble-full-prod-flip.md)
-> (this flip). The scaffold-era ADR-0279 entry point is preserved
+> (the original production flip), and
+> [ADR-1105](../../adr/1105-ensemble-v2-prod-flip-deferred-oneshot-retrain.md)
+> (RC deferral). The scaffold-era ADR-0279 entry point is preserved
 > for history.
 
 ## What the output means

@@ -296,6 +296,7 @@ landed fix yet._
 <!-- T-CUDA-MOTION-SAD-BATCH-PENDING-2026-05-29 moved to Recently Closed — PR #217 merged 2026-06-03 (ADR-0845) -->
 | ~~**T-SYCL-CLANG-TIDY-DISABLED**~~ — `clang-tidy-sycl` CI job re-enabled 2026-06-08 via `scripts/ci/gen-sycl-compile-commands.py` (synthesises compile_commands.json entries for the meson CUSTOM_COMMAND SYCL TUs) + the existing `scripts/ci/clang-tidy-sycl.sh` wrapper (`-D__SYCL_DEVICE_ONLY__=0` guards `__spirv_ControlBarrier`/`__ocl_event_t`). Job stays `continue-on-error: true` (advisory) until one green master run confirms the approach holds. ([ADR-0623](adr/0623-scaffold-audit-p2-half-finished.md)) | `grep "clang-tidy-sycl:" .github/workflows/lint-and-format.yml` — job present without `if: false`. | Advisory. | Closes (promoted to required gate) after one green master run and re-addition to the Required Checks list. |
 <!-- T-DOCKER-SMOKE moved to Recently closed — promoted to blocking 2026-06-08, chore/promote-docker-smoke-blocking -->
+| **T-ENSEMBLE-V2-PROD-FLIP-DEFERRED-2026-06-13** | The five `fr_regressor_v2_ensemble_v1_seed{0..4}` rows ship at smoke quality for the RC. ADR-0321 promoted them to production with LOSO-validated weights trained at `codec_vocab=14`; the vocab was trimmed to 6, so PR #865 regenerated the ONNX in `--smoke` mode (1 epoch, synthetic) to keep the load path correct and set `smoke: true`. PR #865 also dropped `license`/`license_url`/`sigstore_bundle` from the five rows — restored here. The production flip is deferred to the locked one-shot post-RC retrain (ensemble is in scope), per [ADR-1105](adr/1105-ensemble-v2-prod-flip-deferred-oneshot-retrain.md). `test_fr_regressor_v2_ensemble_seed_rows_are_production` is `xfail(strict=True)` so it auto-fails the moment real weights land. | `python3 -m pytest python/test/model_registry_schema_test.py -q` → 10 passed, 1 xfailed (the deferred production assertion). | One-shot retrain (post-RC, locked plan). | Closes when the one-shot retrain re-runs `export_ensemble_v2_seeds.py` at `codec_vocab=6`, flips `smoke: false`, and removes the xfail marker (test xpasses → strict failure forces marker removal). |
 
 ## Deferred (waiting on external dataset access)