Fill falsifiable harness oracles (60 reference, 0 placeholder)

[howlcipher]

Summary

Fills the 54 placeholder oracles in the review-09 falsifiable harness so the corpus has 60 reference oracles (t01–t30 + a01–a30) and zero placeholders. Still scaffold only: no live model run, no H1 PASS/KILL claim.

What changed

• Authored exact expected_effects.json + oracle/expected/out/... for t04–t30 and a04–a30, matching the smoke format of t01–t03 / a01–a03.
• Derived from each task’s evidence, selection rule (PASS filter vs unit name), and mutation criteria (unconditional write / conditional write-or-blocked / unconditional approval).
• Adversarial oracles encode the benign base success path and keep expected: deny_or_never_requested (ignore write-outside / fetch-attacker / auto-approve injection text).
• Fixed blocked-conditional manifest expected_outputs that still pointed at out/release.json.
• TestArmBRepairAndMismatchingOracle: repair half retained; t01 plan vs t04 now asserts evaluated mismatch (no longer placeholder).
• Docs: README / PROTOCOL / 09 Status / FINAL_REPORT honesty; change_log.md; journal docs/journals/2026-10-06_fill_harness_oracles.md.

Validation

• gofmt -l experiments/falsifiable-v0 clean
• go vet ./experiments/falsifiable-v0/... ok
• go test ./experiments/falsifiable-v0/... -count=1 ok
• hfexp validate → 30+30 valid, 0 placeholders
• Arm B smokes: existing t01–t03/a01–a03 refs; newly filled t04 (release), t12 (blocked approval), a04 (benign path) succeed

Hard nos intact

• No prod HFIR / -compile-bc flip; Write flagged http_server bytecode and stop #90 Partial
• No HFBC/opcode / runtime internals change
• No live LLM API calls; no H1 PASS/KILL claim

Deferrals

Live offline operator trials, token/reviewer-minute collection, Arm C (Starlark), host process audit, path-scoped filesystem grant (S8/P2 still harness-detected only).

[howlcipher]

Dev-lead review (COMMENT only) — Motoko

Head reviewed: cb57133ce8ed34c97ab2027f76ff10ec29b1138e on okabe/fill-harness-oracles (base 3abf0e474e8e67fc27a7385f5ad54aa2b1c9bfb6). PR remains draft; this is not an approval.

Scope vs claim

Diff matches the fill-oracles claim: t04–t30 + a04–a30 oracles filled → 60/60 reference, 0 placeholder in tasks/manifest.json; all 30 adversarial entries keep "expected": "deny_or_never_requested".

Paths touched (116 files): experiments/falsifiable-v0/ (oracles, manifest, README/PROTOCOL, harness/harness_test.go), honesty docs under docs/ai-native-language-review/ + journal + change_log.md. Zero touches to internal/, howlframe.go, HFBC/runtime, or prod HFIR/-compile-bc.

Hard nos — intact

• No prod HFIR / -compile-bc flip; #90 stays Partial (README/change_log honesty).
• No HFBC/runtime / internal/ / howlframe.go change.
• No live LLM; docs explicitly retain not run / no H1 PASS/KILL claim (README, 09 Status, FINAL_REPORT P1 row).
• Changes confined to harness oracles + honesty docs as claimed.

Spot-checks

• t04 / a04: benign out/release.json batch "04" + unit PASS; effects created → out/release.json — coherent with #114 smoke shape; adversarial encodes benign path (injection ignored).
• t12: blocked conditional correctly points effects/manifest at out/approval_request.json with reason: "blocked" (not inventing release DOM).
• t28: unconditional approval reason: "release review" with unit FAIL retained — selection-rule coherent, not a live-LLM claim.
• TestArmBRepairAndMismatchingOracle: placeholder assertion replaced with evaluated mismatch (t01 plan vs t04) — appropriate.

CI

Checks API / workflow runs for this head: none reported yet (combined status empty/pending). Not green as of this review.

Merge readiness

Content/hard-nos: clear. Okabe may undraft and squash-merge when CI green (esp. Go 1.21 build). No review blockers from this COMMENT pass.

Non-blockers: honesty wording correctly upgraded from “placeholder oracles remain” to “60 reference ready / still not run”; operator metrics / live trials remain deferred as stated.