f3e3f0f9fe
Umbrella ADR for Phase 5 of ADR-0114. Decomposes the eval-lane work
into eight sub-phases that ship under their own ADRs:
5.1 ADR-0119.1 — sealed-holdout encryption (one lane, fab_control)
discharges Obligation #1 for that lane (delegated)
5.2 ADR-0119.2 — CORE-original GSM8K-style corpus (200 cases)
delegated; preserves real-GSM8K-test integrity
5.3 ADR-0119.3 — lane runner (parser → solver → verifier → realizer)
correct/wrong/refused triple; wrong==0 gate
5.4 ADR-0119.4 — frontier-baseline comparison (citation only)
discharges Obligation #7
5.5 ADR-0119.5 — adversarial generation; misparse rate zero
discharges Obligation #8
5.6 ADR-0119.6 — depth-curve measurement harness
discharges measurement-half of Obligation #6;
threshold ε lives in ADR-0120
5.7 ADR-0119.7 — sealed GSM8K test as the holdout
discharges Obligation #1 for the lane that gates 0120
5.8 ADR-0119.8 — overall lane gate composing 5.1..5.7
registers new `gsm8k_capability_shape` for ADR-0120
ADR-0114a obligation roll-up after Phase 5 completes:
- 6 of 10 obligations already discharged on main (#2/#3/#4/#5/#9/#10)
- Phase 5 closes #1 (per-lane), #6 (harness), #7, #8
- #6 threshold and the `expert` row gate live in ADR-0120
No code lands with this ADR; pure roadmap. Each sub-phase ships
independently. Parallel work enabled:
- 5.1 → Gemini-class brief
- 5.2 → Codex-class brief
- 5.3 → me
Critical invariant: `evals/gsm8k_math/dev` and `public` splits are
CORE-original; the real GSM8K test set enters only via the encrypted
holdout under 5.7. Test-set integrity preserved.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
4 lines
194 B
Text
4 lines
194 B
Text
# Public keys for encrypted eval holdouts (ADR-0105 / ADR-0119.1).
|
|
# Add recipients as new lanes are sealed.
|
|
|
|
fabrication_control: age1rzajyarfvq73vws6prehpauyfxpa4uywf3k4kw4cvtas6hfmkcmq4ha2e3
|