core/docs/adr/ADR-0111-physics-expert-demo-promotion.md

ADR-0111 — physics Expert-Demo Promotion

Status: Accepted Date: 2026-05-22 Author: CORE agents + reviewers Depends on: ADR-0091, ADR-0092, ADR-0100, ADR-0106, ADR-0109, ADR-0110

---

Context

ADR-0110 landed the first reviewer-signed expert-demo promotion (mathematics_logic) and, in doing so, validated the ADR-0106 + ADR-0109 contract end-to-end against a real domain. ADR-0110 §Consequences explicitly noted that the bridges it landed (dev-mode holdout fallback, fabrication_control holdout authoring, by_class fold in reporting) make future expert-demo promotions for the remaining ratified domains "feasible without new contract work — they need only their own lane results to materialise and their own signed claims."

ADR-0111 is the second worked promotion. It exercises the contract a second time on a distinct domain to confirm the ADR-0110 framing — no contract change, no infrastructure bridge — and to retire the "this only works for math" objection.

---

Evidence

The physics domain attaches three lanes via the en_physics_v1 pack manifest (provenance: adr-0100:reviewed:2026-05-21):

Lane	Shape	Public	Holdout
foundational_physics_ood	accuracy_shape	accuracy=1.0 (117/117)	accuracy=1.0 (39/39)
inference_closure	inference_shape	all_pass_rate=1.0, replay_determinism=1.0, overall_pass=True (20 cases)	same (12 cases)
fabrication_control	refusal_shape	by-class refusals 3/3/3, fabricated=0 across all classes	same (9 holdout cases)

All thresholds documented in ADR-0109 §2 are met on both public and holdout splits for every attached lane.

inference_closure and fabrication_control results are shared infrastructure across all domains that attach them; the digest distinguishes physics from math via domain_id + evidence_revision, not via lane-result divergence. (See ADR-0106 §1 — the bundle incorporates domain_id precisely so a shared lane result can support multiple signed claims without digest collision.)

Infrastructure preconditions (already landed)

ADR-0110 landed:

1. Plaintext holdout dev-mode fallback files for inference_closure and elementary_mathematics_ood.
2. Nine fabrication_control holdout cases across the three refusal classes.
3. by_class top-level → metrics fold in core/capability/reporting.py:_fetch_lane_split.

ADR-0111 needed only one additional small bridge:

4. Plaintext holdout dev-mode fallback for foundational_physics_ood. Copied evals/foundational_physics_ood/holdouts/v1/cases.jsonl to cases_plaintext.jsonl so holdout_runner._decrypt_holdout resolves the file without CORE_HOLDOUT_KEY (matches the ADR-0105 dev-mode convention exactly).

This is not a contract change. ADR-0106 + ADR-0109 contract bodies remain untouched.

---

Decision

physics is promoted to expert_demo=true.

The signed claim entered into docs/reviewers.yaml is:

expert_demo_claims:
  - domain_id: physics
    evidence_lanes:
      - foundational_physics_ood
      - inference_closure
      - fabrication_control
    evidence_revision: "adr-0111:reviewed:2026-05-22"
    signed_by: shay-j
    claim_digest: "a104cad136f3219df05dc7ce6a78437c02f7b5827cd3cdce568db3acda6a43ed"

The evidence_revision follows the labeled form established by ADR-0110 (adr-<id>:reviewed:<date>); the load-bearing invariant remains replay byte-equality per ADR-0106 §1.5.

---

Invariants

adr_0111_physics_expert_demo_holds

ledger_report() must report physics with predicates.expert_demo == true and status == "expert-demo" so long as the signed claim in docs/reviewers.yaml resolves and the lane results on disk continue to produce the claimed digest.

adr_0111_replay_digest_byte_equality

Re-deriving the evidence-bundle digest from the lane result files at this commit must reproduce a104cad136f3219df05dc7ce6a78437c02f7b5827cd3cdce568db3acda6a43ed. Tested by tests/test_adr_0111_physics_expert_demo.py.

adr_0111_other_domains_unaffected

ADR-0111 promotes exactly one domain. mathematics_logic continues at expert-demo (per ADR-0110); systems_software, hebrew_greek_textual_reasoning, and philosophy_theology continue at reasoning-capable with expert_demo=false.

adr_0111_distinct_digest_from_adr_0110

The physics digest and the math digest must differ, demonstrating that two domains sharing two of three evidence lanes (inference_closure, fabrication_control) still produce distinct signed claims via the bundle's domain_id + evidence_revision fields. Tested by the same test module.

---

Acceptance evidence

Accepted when:

• docs/reviewers.yaml carries the signed expert_demo_claims entry for physics
• evals/foundational_physics_ood/holdouts/v1/cases_plaintext.jsonl exists (dev-mode fallback per ADR-0105)
• evals/foundational_physics_ood/results/v1_public_*.json and v1_holdout_*.json exist with accuracy=1.0 on both splits
• tests/test_adr_0111_physics_expert_demo.py pins the four invariants above
• ledger_report() confirms physics row at expert_demo=true / status="expert-demo"
• README "Accepted reasoning-capable domains" table updated to note physics at expert-demo

---

Consequences

• Second expert-demo promotion lands. The ADR-0106 + ADR-0109 contract has now succeeded against two distinct domains using three distinct lane shapes (accuracy, inference, refusal) plus one math-specific lane shape (accuracy on elementary_mathematics_ood vs foundational_physics_ood — both accuracy_shape but distinct lane ids).
• The "first promotion was bespoke to math" objection is retired. The bridges ADR-0110 landed were correctly scoped — the second promotion needed only a one-file fallback copy.
• Two remaining ratified domains (systems_software, hebrew_greek_textual_reasoning) are now next in line; both attach inference_closure + fabrication_control plus a third domain-specific lane, and both can follow this same path.

---

Out of scope

• This ADR does not amend ADR-0106 or ADR-0109.
• This ADR does not promote any other domain.
• Sealing the physics holdout under a real age recipient (ADR-0105's eventual design) remains future work. The dev-mode plaintext fallback is acceptable per ADR-0105's own §"Dev-mode fallback preserved" clause.
• Tightening evidence_revision to a raw git sha remains a candidate future amendment; the labeled form is still load-bearing for replay.