[IanMayo]

Summary

Iteration 3 (second zig, both turns across the line of sight) stops at validation under its pre-registered rule (spec 007 FR-006). Seed 204's test split is not scored and stays reserved.

σ = 1°, seed-204 validation	Range, zig+300 s	Course, zig+300 s	Speed, zig+600 s
LSTM (iteration 3)	8.8%	6.19°	1.10 kt [1.06, 1.16]
BiLSTM (iteration 3, reported)	7.2%	4.36°	1.02 kt [0.98, 1.05]
2c models (out of distribution)	11.4% / 17.7%	6.53° / 7.72°	2.01 / 1.93 kt
classical	3.4%	1.43°	0.34 kt
naive	40.2%	89.7°	4.23 kt

• Operative condition fails: the LSTM's speed interval needed a lower bound under 1 kt.
• Sanity condition holds: iteration-3 LSTM / 2c LSTM, paired speed ratio 0.55 [0.52, 0.59]. This is a weak reference, as recorded before training.
• Gap to classical widened: LSTM/classical speed ratio 3.29 [3.11, 3.54], against 2.4× in 2c. The forecast (1.1–1.3 kt) held.
• BiLSTM: it finished by early stopping after seven chained jobs over three dispatches (35 runner-hours).

Contents

• experiments/20261021-i3a-val/: CI run 39604573996 output (models, metrics, paired.json) and validation.md (the rule, tables, training, reading).
• Site:
  • The iteration-3 results page and sidebar entry.
  • A blog post.
  • The viewer catalogue gains the iteration-3 pair. prebuild.mjs accepts a validation experiment without verdict.json.
• Docs: AGENTS.md gate status, plus a note that an incomplete model needs a re-dispatch. Spec 007 status and tasks: T012 not run, T013 done.

No agent panel: this is the mechanical application of a rule committed before training, not a gate decision. The next Part B lever is the owner's choice.

Constitution check

• Data types defined only in schema/tma.yaml (I): no schema change.
• Deterministic; golden files unchanged (II).
• Pure sim/features; one CLI per stage; Keras-equivalent layers (III): no code change outside the site prebuild.
• Every model result sits beside naive and classical on the same rows (IV).
• Respects the current gate (V): the test split was not scored, per FR-006.

Site pages

• Glossary entry: no new term.
• Data dictionary: regenerated automatically.
• Algorithm note: no new algorithm.
• Blog post: site/blog/2026-10-22-iteration-3.md.
• Not applicable.

Checks

• pnpm lint && pnpm typecheck && pnpm test && pnpm schema:check && pnpm site:build: all pass (325 tests).

---

Generated by Claude Code

[github-actions]

Site preview: https://debrief.github.io/Gitma/pr/53/ (updates on every push; removed when the PR closes).