[IanMayo]

Summary

Two time-late studies, each registered before its fresh seeds and decided by a five-reviewer panel. The owner merges: the PR adds a row to SRD §5 and lines to §9.

Spec 015: repair and re-registration (test seed 337)

Repairs the three faults spec 014's panel found (the evidence's duration volume, the edge rule, the drifting reference) and decides the same quantities again. Nothing trained.

• Held-course false manoeuvres 0.53 (0.80 for spec 014's search on the same periods); headroom with a perfect evidence 1.17 [1.10, 1.25], lower limit 1.099: inconclusive, and by the registered rule no GPU budget.
• Panel (experiments/20261113-tl3/review.md): numbers reproduce; the headroom depends most on how the reference is fitted (1.11 to 1.19); the evidence repair is not finished at the edges of the prior's support.

Spec 016, study L: classical controls and a learned model (test seed 341)

Owner's decision 2026-11-13 (option B): build the classical controls and train a learned whole-period model in parallel under a plan registered first, scored only once the controls exist.

• Controls (nothing trained): wild-bearing editing, an effective bearing sd from the estimated error correlation, a broadened prior pooled with Annex B's, a cautious declaration rule, a hedged track (timelate/hedge.py, corr.py, run4.py, study4.py).
• Model: a bidirectional transformer over the whole period, three seeds of 12.5 million simulated periods on a T4 each ($14; November's Modal total about $41 of the $50 budget), its reported track the arc model fitted at the manoeuvres it declares (regress/period_data.py, timelate/learned.py, tma2-tlpredict).
• Registration experiments/20261113-tl4-decision/ (plan, mix, training plans, code and criteria bound by digest before any training run or fresh seed; two addenda). The scorer refuses a learned track without the controls' committed validation record.

Registered outcome (experiments/20261114-tl4/, 2,860 periods, scored once): learning is shown on no split and there is harm on Annex B data, so by the rule the classical controls stand and the learned track stops.

Split	Declared control ÷ learned (three seeds)
Shifted prior	0.82, 0.80, 0.82
Shifted prior 2	0.87, 0.87, 0.94
Correlated errors	0.83, 0.83, 0.84
Outliers	0.78, 0.77, 0.81
Bias (absent from training)	0.98, 1.00, 0.97
Annex B	0.82, 0.81, 0.82
Held course	1.06, 1.06, 1.06

The classical control against the standing search: 1.68 with 2% outliers, 2.44 with 5%, 1.33 with correlated errors, 1.70 on quiet targets; nothing under bias or a shifted prior.

Panel (experiments/20261114-tl4/review.md, five approve with findings; all 42 ratios reproduce):

• The 20% is a tail: the median per-period ratio is 1.00 on every split. The whole loss is where the model declares fewer manoeuvres than the control; its threshold (0.5) was never tuned while the control's settings were. The registered comparison measured the model's detector, so "this model with this read-out stops".
• The model's own output is 0.71 of the control on clean data, and level under the shifted prior in range only.
• The mismatch conditions lie inside the model's training ranges: it lost in-distribution.
• The control is not free (it calls more one-manoeuvre targets quiet) and knows the simulator's error family.
• Not delivered: a working prior-mismatch diagnostic, a Huber cost, a whitened fit, the export check on trained models, IMM/EKF tracks beside, the weights' digests in the test record, unit tests for the runner and scorer.

Also in this PR

• GPU launcher back to module-scope functions (the loop-built form had never run and failed at container start-up); the workflow bounds its wait.
• Time-late workflow can fetch trained checkpoints and score them beside the controls.

Follow-on (owner's choice)

Panel's recommendation: adopt the classical control as the standing time-late system and study, classically, whether the shifted-prior gap is identifiability; only if it is not, try the trained models as proposers inside the search (CPU only).

Posts

Spec 015: kickoff, write-up and slides dated 2026-11-13 (time-late-repair). Spec 016: kickoff 2026-11-13, write-up 2026-11-14-time-late-learned.md, slides 2026-11-14-time-late-learned-slides.md (with track plots).

Test plan

• uv run ruff check ., ruff format --check ., uv run mypy
• pytest tests/unit/test_timelate4.py tests/unit/test_period_model.py tests/unit/test_timelate3.py
• Site builds
• CI green on the head commit

🤖 Generated with Claude Code

https://claude.ai/code/session_019erPajyAwQ6mBQ2Xz4xNHs