[IanMayo]

Summary

Spec 017's test on seed 344 (scored once, plan tl-f.json as registered), the five-reviewer panel, and the two posts. PR 34 merged the study's code and registration. A registered study, not a gate; nothing trained, nothing spent.

Record (experiments/20261114-tl5/; shifted split, 451 scored periods):

• 0.37 of true manoeuvres add less than 20 to the χ² when left out (Annex B: 0.16).
• Control ÷ reference told every manoeuvre: 2.27 [2.04, 2.55].
• Not being told the ones under 20 costs 1.58; control ÷ that reference 1.43 [1.28, 1.61].
• Attainable share 0.65 / 0.44 / 0.14 at thresholds 10 / 20 / 40.
• Control ÷ the search under the targets' own prior: 1.17 [1.06, 1.29].
• By the logic fixed in the decision record, a proposer study is not recommended (one of two conditions held).

Panel, round 1 (review.md): five approve with findings; every number reproduces. The reading is narrowed:

• The pooled split is two kinds of period: where a manoeuvre is hidden the control is 1.16 times the reference not told it; where none is, 2.65. Under a third of periods carry 83% of the attainable gap.
• The logic's second condition counted manoeuvres, not cost, and turns on the match tolerance: the "no" is the rule's output, not the evidence's.
• The attainable part is at least half a prior problem (of 138 detectable manoeuvres the control misses, the right prior finds 66).
• The detectability measure is not additive and the references are not bounds.
• Corrections to the first record are listed at the end of the README (the tuning/test "disagreement", the IMM/EKF ratios, shifted 2).

Also: write-up site/blog/2026-11-14-shifted-gap.md, slides post and deck, figures with sidecars, scripts tl5_{supplement,tracks,charts}.py, AGENTS.md (results, gotchas, next fresh seed 345), tasks ticked.

Recommended follow-on (owner's choice): a manoeuvre prior that adapts to the period, classical and on CPU, registered against the control with the true-prior search as its ceiling; no GPU proposer now.

Not delivered: the split under each way of fitting the reference; course and speed in the estimands; a joint detectability; unit tests for detect.py and run5.py.

No SRD, annex or constitution change in this PR.

Test plan

• uv run pytest -q tests/unit/test_timelate5.py (tables regenerate from the summary)
• uv run ruff check . and ruff format --check .
• Site build with the new posts
• CI green

🤖 Generated with Claude Code

https://claude.ai/code/session_019erPajyAwQ6mBQ2Xz4xNHs