What is demonstrated, what is internally reproducible, and what remains open.
The goal is to make the evidence boundary easier to inspect than the narrative.
Evidence ledger.
internally reproducible is not independently reproduced. Retained negative results are listed alongside positive ones.
| Question | Evidence | Reproducibility status | Boundary |
|---|---|---|---|
| R.01 Can information acquisition be premature? | Frozen 15-state diagnostic; full REC 15/15; principal timing/consequence ablations collapse pre-boundary performance. | Offline paper-scoped artifact with raw outputs, hashes, evaluator, reference implementation, and verification scripts. | Independent external reproduction and broad real-world prevalence are not yet established. |
| R.02 Can passive value-of-information choose the wrong RFQ policy? | Formal representability boundary; 10,080 heterogeneous tests; fresh 5,000-world IID holdout; mechanism controls and shifts. | Exact enumeration; separated discovery/holdout seeds; frozen holdout ranges. | Stylized mechanism test, not an empirical estimate of real RFQ leakage frequency. |
| R.03 When does private-credit model error become decision-critical? | Frozen 20-world holdout; line-cut control; 14/20 failures under omitted working-capital mechanism; stressed optimum feasible 20/20. | Pre-specified holdout plus separately frozen, source-anchored 40-world follow-up. | Original/follow-up are not hash-continuous; synthetic environment; no deployed-system prevalence claim. |
| Negative result Is REC a new generic planning primitive? | Negative result: exact depth-3 expectimax solves 27/27 small worlds. | Retained explicitly in the reproduction package. | No generic-planner novelty claim. |
| AEI Does the broader AEI thesis generalize? | Multiple internal hidden-rule, rare-event, constraint, document, and adaptive-environment evaluations. | Large private forensic archive with frozen specs, hashes, raw outputs, and reconstruction material. | External replication, externally specified worlds, and real-world validation remain decisive next tests. |
How claims are handled.
Freeze before confirmatory evaluation
Where a result is treated as prospective evidence, the environment, seed rule, analysis, or evaluation criteria are fixed before the corresponding holdout is run.
Keep negative evidence
Generic-planner successes, oracle contamination, non-identifiability, failed controls, and null results are retained rather than silently dropped.
Separate discovery from confirmation
Development worlds and post-hoc follow-ups are not relabeled as prospective confirmation.
State external-validity limits
Synthetic mechanism tests are not presented as prevalence estimates for deployed systems.
Public, paper-scoped, and private layers.
Public
Research summaries, selected evidence, neutral manuscript metadata, and safe research notes.
Paper-scoped
REC paper-scoped offline artifact; finance reproduction packages can be published as separate safe repositories.
Private
Full AEI architecture, reconstruction-grade implementation details, complete internal evaluation suite, and successor design.