[IanMayo]

Summary

Iteration 3 setup (spec 007). The owner's realism check and lever choice are recorded as decision 22.

The change is to the problem: ownship gets a second zig, and both turns cross the line of sight. The first zig keeps its timing. The models keep the 2c configuration and train from scratch on a fresh dataset, seed 204.

Evidence behind the choice (research R1, R2)

• R1 (seed-203 gate predictions, observational slices):
  • Shorter first legs are worse for every estimator, so an earlier zig was not chosen.
  • Turning towards the contact helps only when the turn crosses the line of sight.
  • Larger turns help every estimator.
  • No slice brings the models within 2× of classical.
• R2 (classical forecast on the new ranges): speed at zig + 600 s halves, 0.67 → 0.36 kt, and at zig + 300 s 1.13 → 0.56 kt.
  • The cost is large turns: median 143° against 61°.
  • Windows run longer, because second turns that do not fit are redrawn.
• Forecast for the models, recorded before training: 1.1–1.3 kt, probably still above 1 kt.

What this PR contains

• Generator 0.3.0 (packages/sim):
  • zig_rule: across_los puts the target on the other bow at a drawn offset, by the shorter turn.
  • zig2_delay_s adds a second zig, giving three legs.
  • The new draws come after every 0.2.0 draw, so earlier ranges are byte-identical.
  • Tested by hand-worked known answers, a 200-scenario rule check, a new golden file, and the unchanged 0.2.0 goldens.
• Schema 0.9.1 adds ranges, parameters and metadata for the second zig, and ChangeDecision.plan / test_seed.
  • 0.1.0 and 0.2.0 data stay readable.
  • Data and features goldens change only in their version strings.
• G0:
  • The independent-geometry check gains a hand-computed leg-3 case.
  • With across-LOS data, it recomputes every turn of every scenario with literal atan2.
  • Every turn is also checked kinematically (arc chord, heading bracket, turn rate, summed angle).
  • G0 passed on 0.3.0 (panel round 3, no rejections: experiments/20261017-g0-i3-r3/); the G0 guard accepts only it.
• Slices: zig bands extend to 180°.
• tma-g2:
  • From seed 204, --split test needs a decision bound to the plan and test seed (FR-009).
  • paired.json: paired bootstrap speed ratios against the other phase and classical (FR-006 validation rule).
  • The other phase is scored with its own normalisation when trained on other data.
• Plan configs/iter3/a:
  • Seed 204 digests are pinned.
  • The model configs are identical to 2c's.
  • Seed 204 is reserved for iteration 3.
• Viewer:
  • It draws the second turn and leg 3 and shows the second zig.
  • The seed-204 validation split is in the catalogue before any training, with a featured bookmark.
• Docs:
  • Spec 007: spec, research R1–R4, tasks.
  • Annex A §A.7 record and D4.
  • SRD §3 row and decision 22.
  • AGENTS.md, glossary ("across the line of sight"), generator note.

Status

• Decision record committed before training: experiments/20261017-i3-decision/.
• Training is running on train/i3a (run 39024150564). The validation rule, any test scoring and the G2 panel follow in a later PR.

Constitution check

• Data types defined only in schema/tma.yaml; pnpm schema:gen run (I): schema 0.9.1 regenerated, and schema:check passes.
• Deterministic: seeded PRNG only; golden files updated deliberately, with the reason stated (II):
  • The new draws come after the existing ones.
  • The 0.2.0 sim goldens are unchanged.
  • The data and features goldens change only in generator_version and schema_version.
  • There is a new three-leg golden.
• sim/features stay pure functions; one CLI per stage; Keras-equivalent layers only (III): simulateTruth(parameters) reproduces the tracks, because the realised turns are stored.
• Any reported result sits beside naive and classical numbers on the same test set (IV): no model result yet, and the R2 forecast is classical only.
• Respects the current gate: no training before G0, no scope beyond SRD §1 (V): the 0.3.0 G0 passed its panel before the decision record and training.

Site pages

• Glossary entry (site/docs/glossary.md) for any new term: across the line of sight, leg, zig.
• Data dictionary: regenerated automatically from the schema (nothing to edit).
• Algorithm note (site/docs/algorithms/) for any new or changed algorithm: the generator's across-LOS rule and second zig.
• Blog post (site/blog/) for any new result or gate decision: comes with the iteration-3 result.
• Not applicable: this PR adds none of the above.

Checks

• pnpm lint && pnpm typecheck && pnpm test && pnpm schema:check && pnpm site:build

🤖 Generated with Claude Code

https://claude.ai/code/session_01Gv4979qzsScV5ohT5RNUoG

[github-actions]

Site preview: https://debrief.github.io/Gitma/pr/52/ (updates on every push; removed when the PR closes).