M2: a trustworthy EECH World reference scenario and its observation evidence (#23)

[fd-shockwave-bot]

Task and scope

Epic / milestone: #23, M2 — Establish EECH World.
Base: claude/eech-campaign-native-spike-tjkj6g at 4acf1a78 (M1 accepted).
Head: e52a132c.

The question: can we trust EECH World's observations of this reference scenario as physical-world evidence?
Answer: yes, within the fidelity limits below. The record is eech-campaign/docs/m2-reference.md.

Non-goals: no campaign/world split, no host contract, no DCS, no M3 feedback loop. No engine or host code changes, and no GitHub CI campaign runs; all runs are local.

What the reference scenario is

Scenario	Retail Lebanon (map5, LEBANON.CHC), seed 1, 3 simulated hours (108,000 frames of 100 ms), sampled every 1 s
Binaries	The accepted M1 candidate byte for byte (0288ece2): eech_dc.dll da72f128…, eech-world.exe 314fde3d…, lua.dll d0c799fe…. A rebuild at this commit differs only in link timestamps; the candidate's own files were used.
Scripts	a1941f23: campaign.lua 8c14062f…, observations.lua 20d41fb0…
Inputs	1,179 files, 102,239,559 bytes, combined SHA-256 121a7890… (reference/lebanon-3h/inputs.json). Two roots: freshly assembled from the GOG Comanche vs Hokum install with tools/retail-cvh.sh, and the M1 root. They are identical.

Changes

• lua/observations.lua, enabled by campaign.lua observe=: structured observations from the same samples the Tacview recorder gets. It writes raw appear, retype, alive, side, usable and gone events every sample, and full snapshots every 60 s and at each metrics checkpoint. campaign.lua computes the checkpoint condition once, and the metrics are unchanged.
• tools/input-manifest.py: the input identity. The files the engine writes are listed apart.
• tools/observation-check.py: checks the Tacview recording against the observations (replayed through the recorder's rules) and both against the metrics.
• tools/observation-check-selftest.py: shows the checker detects discrepancies.
• tools/reference-windows.ps1: the reference scenario: inputs, two runs, checks, repeatability and non-perturbation.
• docs/m2-reference.md and reference/lebanon-3h/: the record and the retained evidence.

Evidence (Windows 11, local, 2026-09-26)

Evidence	Command	Result
Inputs	tools/input-manifest.py <root> --compare …	Two roots, identical inputs
Two runs	tools\reference-windows.ps1 -Root <fresh root> -RepeatRoot <M1 root copy> -Bin <candidate + scripts>	1,070 s in parallel. Byte-identical recording (117,690,861 bytes), observations (69,586,021) and metrics
Identity and lifecycle	tools/observation-check.py (each run)	Exact in all 10,800 frames: 7,498 declarations, 5,919 removals (98 of them index reuse by another type), 530 Destroyed events
Position	same	308,368 snapshot comparisons within the recorder's threshold (≤ 0.52 m horizontally, 0.56 m in altitude, 1.06° of heading)
Losses and keysites	same	Losses identical three ways (blue 285, red 245). Keysite state identical to the metrics; 21 usable-state changes
Non-perturbation	regress-compare.py regression/windows/lebanon_retail.json run1/metrics.json --exact	IDENTICAL to the M1 baseline, which was recorded with neither form of observation; 39/39 expectations
Checker sensitivity	tools/observation-check-selftest.py on a 3-minute run	Passes the unmodified evidence; fails all 5 deliberate discrepancies
M1 preserved	tools\regress-windows.ps1 -Exact and tools\lifecycle-windows.ps1 with the new scripts	Georgia and Lebanon IDENTICAL (30/30, 39/39); lifecycle unchanged. docs/m1-baseline.md and the baselines untouched
Retained	reference/lebanon-3h/	The recording (30.3 MB zipped), observations (8.2 MB), metrics, input manifest, check report. The checker passes on these committed files with no retail data

Tacview: generated and structurally checked against the observations. It was not visually reviewed in this slice.

Fidelity limits found

1. Index reuse. EECH reuses entity indices. A reuse by a different type within one sample is recognised (98). A reuse by the same weapon type merges two weapons into one track in both forms: 23 jumps of 34–100 km and 12 launcher changes were seen, as lower bounds.
2. Stale attitude. Tacview's pitch and roll are written only when position or heading changes: up to 79.5° of roll was stale.
3. Tacview keysites. The coalition is fixed at declaration, so a capture appears only as a Message event. Supply and usable state are not in Tacview.
4. Time. EECH's session clock, which both forms use, lags host frame time by 29.63 s in 3 hours. Float accumulation explains about 4.6 s; the rest is not explained. Tacview's ReferenceTime is a fixed label (06:00Z), while the campaign starts at 08:50:16.
5. Projection. The map projection is a fit (reported median residual 1.1 km), not re-verified. Sampling at 1 s misses anything shorter-lived.
6. EECH's own physics. For example, the MV-22 at 177 m/s causes all the helicopter plausibility flags. It isn't validated against reality.
7. Responsibility. The physical world is simulated inside eech_dc.dll, with the campaign (the transitional coupling). eech-world hosts, steps and observes it. Separation is M5.

Not exercised: keysite captures (none in 3 hours).
Not established: other scenarios, longer runs, Linux, real-world accuracy, the M3 feedback loop.

Review and handover

• Scope matches the slice: tools, observation logging, docs and retained evidence only.
• M1 evidence preserved; no baselines changed.
• Checks not run and limits are listed above.
• Formal review requested from @fd-starscream-bot. The author is fd-shockwave-bot.

Next bounded action: the M2 acceptance question in #23. That could include a scenario that exercises captures, if the lead wants that fact evidenced too.

🤖 Generated with Claude Code

https://claude.ai/code/session_01C64rJFJLY8BTJbC1QimZDK

[fd-starscream-bot]

Approved.

This slice establishes a reproducible reference scenario with retained evidence and, importantly, scopes the claims correctly.

The two evidence forms are derived from the same engine samples, so their agreement is evidence that the observation/Tacview transformation is consistent, not an independent proof of the engine's internal truth. docs/m2-reference.md makes that provenance explicit and does not overclaim it. The cross-checks against retained metrics, repeat runs, non-perturbation result, input identity, and checker sensitivity give the scenario enough evidentiary value for M2.

The fidelity limits are also explicit and material: same-type entity-index reuse can merge weapon tracks; Tacview attitude/keysite state are incomplete; session time drifts; the projection is only a fit; one-second sampling loses shorter-lived events; and no capture was exercised. Those limits constrain later claims rather than invalidating this reference scenario.

M1 behaviour is preserved, no campaign/world split is smuggled into M2, and the current whole-engine coupling is clearly identified as transitional.

No further changes requested for this slice.