Finite lineage challenge — AI design review 001
Date: 2026-10-16. Reviewer: OpenAI GPT-6 Sol, a separate AI review agent. This is an AI assessment, not a human external expert review or Cycle 03 entry approval.
Reviewed material and limits
- Draft:
research/preparation/FINITE_LINEAGE_CHALLENGE_DRAFT.md, SHA-256bdf198bc95ec49dbcd720bfa156cf050c7d3c459484c0ded9cc616cfd3f91136. - Construction semantics excerpt, with only the question and construction/merge sections: SHA-256
44b22c907990094983f1a11dab7f087524cfc519379f204d91c18a682fc31138. - Payload-free RP004 structure export v2: SHA-256
b448013d9f849f51af1bf026c9bf7b666bd5110a32aa28315c84c1e2819ff1ea.
The reviewer did not open the original fixture, original RP004 protocol, RP006 code/evidence, or outcome files, and generated no cases or answers. A prior AI review attempt had read control-bearing prose in the operational protocol and was stopped; none of that attempt's findings were used.
Findings
Decision value. The challenge can change a corpus-specific representation decision if that decision is declared before execution. A same-projection/different-answer pair rejects the mask for that inquiry on any corpus containing the pair; no conflicts support only the declared 55 cases. FULL is justified when birth_kind and alive are needed. TOPOLOGY_STATE, a subset of FULL, has separate value only if a real downstream use excludes those two queries and could retain fewer fields. The draft must specify that use and the decision made under each possible result before outcomes.
Construction and state sensitivity. All eight event sequences are legal under the supplied semantics: live inputs, fresh outputs, arity, consumption, and copy/fork state preservation hold. Final live sets are F4-01 {a,c,d}, F4-02 {c,d}, F4-03 {a,d}, F4-04 {b,c,d}, F5-01 {d,e}, F5-02 {d,e}, F5-03 {a,e}, and F5-04 {c,e}. The illustrative vectors include repeated and distinct states but do not establish coverage. The forced-equality component counts are 3,2,3,3,3,3,4,4; partitions into at most three blocks produce 5,2,5,5,5,5,14,14, totaling 17 F4 and 38 F5 variants. Canonical A/B/C assignment must be fixed across cases: arbitrary per-case renaming can change cross-case S projection equality even when within-case same_state is unchanged.
Structural overlaps. D below means a parent-closed three-node parent-DAG match after ID relabeling. L also matches birth-operation labels and fork-sibling relation, but ignores independent event order. Blank cells have no match.
| Candidate | G0 | G1 | G2 | G3 | G4 | C4 |
|---|---|---|---|---|---|---|
| F4-01 | D {a,b,c} |
D {a,b,d} |
||||
| F4-02 | D {a,b,c}, {a,b,d} |
|||||
| F4-03 | D {a,b,c} |
|||||
| F4-04 | D {a,b,c}; D+L {a,b,d} |
|||||
| F5-01 | D {a,b,c}, {a,b,d} |
|||||
| F5-02 | D {a,c,d} |
D+L {a,c,d} |
D {a,b,c}, {a,b,d} |
|||
| F5-03 | D {a,b,c}; D+L {a,c,d} |
|||||
| F5-04 | D {b,c,d} |
D+L {a,b,c} |
The draft's “operation-closed” term is incorrect: parent closure does not include every output of a fork. F4-02 and F5-01's G0 triples and F5-02's C4 triples omit a fork sibling and are not replayable three-entity sublogs. A corrected report should distinguish parent-closed DAG matches, event-labelled matches, event-complete replayable sublogs, and strict ordered event subsequences/prefixes. A local merge neighborhood resembles G4 when ancestry before the merge inputs is ignored, but is not a parent-closed G4 triple.
The sanitized C4 events are root(a); root(b); continue(a→c). F5-04 begins root(a); root(b); continue(b→c). Swapping roots makes the structures event-labelled isomorphic, but maps the chronology to root(b); root(a); continue(a→c), so this is not a strict ordered prefix. It is a structural extension only if independent roots may commute. F4-04's {a,b,d} contains an ordered C4 subsequence with an interposed copy event; it is not a contiguous prefix. The reviewer could not check old-versus-new state-equality-pattern overlap because the v2 structure export contains no state values.
Protocol and arithmetic. Counts are correct: 17×60 + 38×90 = 4,440 FULL truth rows; 55×8 = 440 projections; 8×(C(17,2)+C(38,2)) = 6,712 pair/mask records; and 2×8×2 = 32 certificates. TOPOLOGY_STATE reuses its subset of FULL truth rows. Pair records need family-specific comparison status and findings because only projection equality is shared. A manifest cannot include its own hash; bind its hash through a separate receipt. The current toy probe does not bound challenge output volume, conflict-witness volume, disk behavior, or suspension/runaway-worker handling. Before a recorded run, define explicit output limits and partial-output handling; review a no-answer synthetic output-shape stress test, external suspension-aware supervisor, and process-group termination/reaping. The verifier should independently implement parsing, semantic reconstruction, enumeration, projection checks, and certificate validation; preserve disagreements.
Prior-art boundary. The draft's description of view determinacy and selected-information preservation as established work is defensible as written. The narrow possible contribution is an auditable finite benchmark for fixed lineage queries, aligned IDs, whole-field masks, and predeclared equality variants. It is not a new determinacy concept. The reviewer did not open cited papers or conduct another search; citation accuracy/completeness and exact benchmark overlap remain unresolved.
Disposition
Revise and return. Fix the overlap terminology/criteria, predeclare the two use profiles and family-retention rule, make cross-case canonical state labels explicit, define family-specific pair outcomes and manifest hash binding, and close the output/resource/supervisor limits. Then submit the corrected exact bytes to another independent review before any separate entry decision. This review does not authorize Cycle 03, a source freeze, package/Test registration, resource allocation, or execution.
Source record
Repository path: research/preparation/FINITE_LINEAGE_AI_REVIEW_001_2026-10-16.md
SHA-256: 3f56c55f69a8e8b6b78b355167b3597767466ecc1ecd452badbe7140466f4cf7
This reading view is generated from the source. The downloaded Markdown preserves the original bytes.