Scientific record
Status at specification: approved local diagnostic execution by the supplied user request; no cognitive-system trial authorized. Date: 2026-10-23.
Reconciliation before implementation
Current Fiction question, FH-01–FH-09, EXP-0001/2/3 and their gates are retained. DEC-0007 integrated only partial conversation recovery. The full original PT-001/Cycle 3 memorandum is unavailable here. This packet is a newly specified local realization of the supplied Cycle 3 review, not a recovered or verbatim revision of the missing protocol. CV-001 remains separate. The user's current review provisionally accepts fixed-goal versus problem-transformation as an experimental distinction, not separate mechanisms, and explicitly authorizes this offline execution. It supersedes the earlier recovery-first next-step requirement for this bounded work, without supplying the missing reports.
Prior canon is preserved in archive/research-snapshots/2026-10-23-before-pt-001/. Existing CWR files are hash-inventoried and are not edited/imported. We reuse its methodological pattern (separate replay evaluator and exhaustive reference), not its token semantics or code: allocation outcomes need a different contract.
Seven design statements (specified before first run)
- Hypothesis under diagnostic challenge: An apparent gain credited to reframing may be entirely reproduced by optimizing an explicitly specified value criterion or changing candidate search order/coverage. This is a claim about measurement identifiability, not cognition.
- Operational definition: Plans allocate indivisible resource units to three recipients. Validity is exact feasibility. Benefit is improvement in the declared synthetic stakeholder-value total; proxy gain, minimum recipient benefit, coverage and units allocated are separate measures. No text novelty or internal reframing detector is implemented. Candidate conflict is established only if maximum total value exceeds the best value among proxy-maximizing outcomes, avoiding arbitrary proxy tie-breaking.
- Alternatives: Better objective information, search ordering/coverage, differing feasible spaces/utility curvature, arbitrary stakeholder weights, and evaluator defects explain apparent gains.
- Controls: Four conditions share each world's facts, costs, capacities, horizon, legal actions and all value tables. P optimizes proxy (diagnostic reference). V optimizes value without reframing. R explicitly considers replacing the proxy formulation with stakeholder allocation and uses a balance-first candidate order. S expands/diversifies candidate search without explicit reframing, using high-utilization-first order. V/R/S use the same value score and candidate-evaluation ceilings. Also compare S at the next larger budget, explicitly unequal-resource. Rotate all six recipient enumeration orders. Full-coverage V/R/S establishes an objective ceiling. Procedural instructions differ and convey approach information; fixed task facts do not mean identical informational input.
- Measurement success: Independent allocation enumeration and sequential replay agree on *all* bounded plans, rejection boundaries, outcome vectors and metrics; intended conflict/no-conflict labels survive exact primary-value checking; invalid plans receive no outcome credit; wording/goal claims cannot change scores; raw strategy outputs, frontiers and sensitivity results are reproducible. Any requirement failing blocks that measurement claim. These exact checks are appropriate to finite apparatus, not statistical thresholds for humans/models.
- Failure / attribution criterion: A claimed reframing-specific benefit is unsupported if direct value optimization attains it, extra/order-changed search reproduces it, wording earns credit, invalid plans pass, or a result depends on undisclosed value choices. Even positive bounded R–V differences cannot identify reframing separately from R's programmed candidate ordering. No statistical cognitive claim can pass this rehearsal.
- Next decision: If apparatus fails, repair with versioned failure records. If it works but attribution fails, retain it as a measurement diagnostic and require a genuine controlled cognitive intervention with shared value information, matched resources, search controls and independent formulation evidence before attributing benefit to problem transformation.
Worlds and counts
worlds.json is the complete input. All start with allocations [0,0,0], six resource credits and recipients A/B/C with capacity four each. An operation delivers one unit to exactly one recipient, consumes its delivery cost, and increments that allocation. No borrowing, negative allocations, transfers, new recipients, rule changes, or textual redefinitions of value are allowed. Stopping at any point, including no action, is legal. Horizon is six deliveries. Resource costs are one in W1/W3 and [1,2,1] in W2/W4. Remaining credits are recorded, not valued in the primary measure.
W1/W2 are hypothesized proxy-conflict worlds; W3/W4 hypothesized no-conflict under primary total value. W1/W3 share a feasible geometry; W2/W4 share another. Opportunity labels are testable proxy/value conflict hypotheses, not proof of a cognitive reframing opportunity. The four worlds are selected synthetic examples, not random/independent samples or fully matched cognitive tasks. Different utility curvature, optimal-set density and candidate order can change difficulty even within geometry pairs.
Proxy is 3*A + 2*B + C. Total value is the sum of each recipient's table entry at its allocation. W1 favors two underserved groups; W2 has threshold benefits; W3/W4 align primary value with proxy. Tables are chosen normative toy preferences, not measured human welfare. Primary sum expresses aggregate benefit, alternative min-benefit-first expresses protection of the worst-served recipient, coverage-first expresses inclusion, and units-first expresses breadth of delivered service. These are defensible toy choices whose legitimacy in actual stakeholder settings remains unvalidated. No single clever plan is designated correct; every feasible plan is accepted and all optimal ties/frontiers are retained.
Strategies and resource accounting
The shared candidate catalog is generated by enumerating integer allocation vectors using only capacities, cost/horizon constraints, not utility or reference optimal solutions. V and P order it lexicographically after recipient permutation; R sorts by descending minimum allocation then descending coverage then the same lexical key; S sorts by descending total deliveries then lexical key. At ceilings 4, 8, 16 and complete catalog, each condition evaluates exactly the same number of distinct candidates. Selection uses only the condition's objective with lexical tie-breaking; no secondary value tie-break for P. Strategies output a canonical sequence implementing their selected vector. Independent replay scores it after selection.
This balance-first ordering is a constructed implementation of one possible reformulation, not a validated reframing manipulation. A search-only ablation using R's identical ordering is included: its outputs must be identical, exposing the non-identifiability directly. Candidate construction/constraint screening and sort-key computation scan the full small catalog for every condition. We record those counts and runtime but do not claim equal actual computation from equal evaluation counts. Equal budgets concern the number of objective-evaluated candidate plans only. The extra-search comparison explicitly spends more evaluations. No solver calls, value lookups during sorting, private optimum access, external information, learned model or random seed are used. Random seed: not applicable (deterministic exhaustive enumeration).
Independent verification and adversaries
reference.py enumerates all integer allocations algebraically and computes metrics separately. It does not import the replay evaluator. evaluator.py consumes sequential actions, enforces constraints at each prefix and returns exact final metrics. A full grammar audit covers all 1+3+...+3^6 = 1,093 plans per world, including illegal prefixes, plus explicit over-horizon and malformed cases. Order-equivalent plans are not independent outcomes. A combinatorial multinomial count independently checks sequence multiplicities for each allocation. Full feasible vectors, metrics, objective optima and multiobjective non-dominated outcome frontiers are emitted.
Counterexamples: direct value versus R ceiling; S/extra-search versus R; same plan with superficial narrative; easier claimed proximal goal with same plan; capacity/budget/horizon/unknown-action violations; alternative value criteria and ranking changes. Neither prose nor asserted objective changes the evaluator. This prevents undeserved novelty credit by withholding any novelty claim; it does not validate a novelty measure.
Reproduction · interpretation
Python 3.10+ standard library only. From this directory: python -B run.py --output runs/reproduction and python -B -m unittest discover -s tests -v. Existing output directories are refused to preserve runs. Input/code hashes, environment, UTC run timestamp, deterministic seed status, condition descriptions, strategy traces, candidate counts and actual times are retained. Use output content excluding timestamps/runtime to compare deterministic results. User-facing record date uses Africa/Lagos. All outputs are constructed-strategy/software observations or bounded exhaustive verification results. Human/model observations, participants, paid calls and spend are zero. No human cognitive mechanism or general innovation result follows.