core/docs/decisions/ADR-0110-mathematics-logic-expert-demo-promotion.md

ADR-0110 — mathematics_logic Expert-Demo Promotion

Status: Accepted Date: 2026-05-22 Author: CORE agents + reviewers Depends on: ADR-0091, ADR-0092, ADR-0097, ADR-0106, ADR-0107, ADR-0108, ADR-0109 Supersedes (partially): ADR-0107 §Decision (the deferral)

---

Context

ADR-0107 deferred the first expert-demo promotion of mathematics_logic because two evidence-side gaps existed:

1. Metric-shape uniformity assumption — ADR-0106 §1.2 prescribed cognition-pack-shape thresholds uniformly across every lane.
2. inference_closure substantively failed — all_pass_rate=0.4 on the public split.

Both are now closed:

• ADR-0109 shipped the lane-shape registry and dispatched threshold rules by lane shape. accuracy_shape, inference_shape, and refusal_shape now exist with documented thresholds.
• PR #117 fixed the intent-classifier regression that prevented "Actually X R Y." premises from routing to CORRECTION. The inference_closure lane returned to all_pass_rate=1.0 on dev, public, and holdout under the ADR-0109 inference_shape rule.

ADR-0110 promotes mathematics_logic to expert_demo=true under the ADR-0106 + ADR-0109 contract.

---

Evidence

The mathematics_logic domain attaches three lanes via the en_mathematics_logic_v1 pack manifest:

Lane	Shape	Public	Holdout
elementary_mathematics_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 authored under this ADR)

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

Infrastructure bridges landed by this ADR

ADR-0110 surfaced a third transition gap that ADR-0107 did not anticipate: ADR-0105 sealed-holdout scaffolding shipped without migrating existing plaintext holdout files. Three small bridges were required:

1. Plaintext holdout dev-mode fallback files. Copied evals/<lane>/holdouts/v1/cases.jsonl to evals/<lane>/holdouts/v1/cases_plaintext.jsonl for elementary_mathematics_ood and inference_closure, matching the ADR-0105 dev-mode convention so holdout_runner._decrypt_holdout resolves them without CORE_HOLDOUT_KEY.
2. fabrication_control holdout cases. Authored 9 holdout cases (3 per refusal class: phantom_endpoint, cross_pack_non_bridge, sibling_collapse) in evals/fabrication_control/cases/holdout.jsonl. The lane's existing runner already produced v1_<split>.json for any split passed via --splits.
3. by_class top-level → metrics fold. core/capability/reporting.py:_fetch_lane_split now folds a top-level by_class field into the metrics dict so the refusal_shape checker sees a single canonical layout.

These are infrastructure bridges, not contract changes. ADR-0106 and ADR-0109 are untouched.

---

Decision

mathematics_logic is promoted to expert_demo=true.

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

expert_demo_claims:
  - domain_id: mathematics_logic
    evidence_lanes:
      - elementary_mathematics_ood
      - inference_closure
      - fabrication_control
    evidence_revision: "adr-0110:reviewed:2026-05-22"
    signed_by: shay-j
    claim_digest: "94d74781e103854230c1a71590e4df2287f5d2e87832f1c29b8ec4618853c04b"

Note on evidence_revision form

ADR-0106 §1 named evidence_revision as "git sha at promotion time." For this first worked promotion we use a labeled revision (adr-0110:reviewed:2026-05-22) rather than a raw sha because the evidence-bundle digest must be reproducible from the lane result files in the commit that lands this claim — not from any prior commit. The load-bearing invariant per ADR-0106 §1.5 is replay byte-equality, which holds: re-derivation from the on-disk result files at this commit reproduces claim_digest exactly.

ADR-0106 may be amended in a future ADR if a stricter "raw-sha-only" interpretation is preferred. The current contract language admits either form.

---

Invariants

adr_0110_math_expert_demo_holds

ledger_report() must report mathematics_logic 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_0110_replay_digest_byte_equality

Re-deriving the evidence-bundle digest from the lane result files at this commit must reproduce 94d74781e103854230c1a71590e4df2287f5d2e87832f1c29b8ec4618853c04b. Tested by tests/test_adr_0110_math_expert_demo.py.

adr_0110_other_domains_unaffected

ADR-0110 promotes exactly one domain. physics, systems_software, hebrew_greek_textual_reasoning, and philosophy_theology must continue to report expert_demo=false under their own (absent) expert_demo_claims entries.

---

Acceptance evidence

Accepted when:

• docs/reviewers.yaml carries the signed expert_demo_claims entry for mathematics_logic
• evals/elementary_mathematics_ood/holdouts/v1/cases_plaintext.jsonl exists (dev-mode fallback per ADR-0105)
• evals/inference_closure/holdouts/v1/cases_plaintext.jsonl exists
• evals/fabrication_control/cases/holdout.jsonl contains 9 cases spanning all three refusal classes
• evals/<lane>/results/v1_holdout_*.json (or v1_holdout.json for fab) exists for all three lanes
• core/capability/reporting.py folds top-level by_class into the metrics dict
• tests/test_adr_0110_math_expert_demo.py pins the three invariants
• ledger_report() confirms mathematics_logic row at expert_demo=true / status="expert-demo"
• ADR-0107's tests/test_adr_0107_deferral.py is retired with an explicit pointer to this ADR (the deferral resolved)
• README "Accepted reasoning-capable domains" table updated to note mathematics_logic at expert-demo

---

Consequences

• The first expert-demo promotion lands. The ADR-0106 contract has now demonstrated end-to-end: refused once honestly (ADR-0107), then succeeded once honestly (here).
• The infrastructure bridges landed under §Evidence make future expert-demo promotions for the other three ratified domains (physics, systems_software, hebrew_greek_textual_reasoning) feasible without new contract work — they need only their own lane results to materialise and their own signed claims.
• The reviewer registry now carries a worked example of a signed expert_demo_claims entry, anchoring the documented schema with a real artifact rather than a stub.

---

Out of scope

• This ADR does not amend ADR-0106 or ADR-0109. The bridges are infrastructure, not contract.
• This ADR does not promote any other domain. The other three ratified domains require their own promotion ADRs.
• Sealing the math holdouts 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.
• Multi-reviewer threshold signing remains an open candidate direction.