01
Use strong baselines to shape architecture
7.3 records live-model literature development and a frozen eight-state diagnostic. It asks whether observations change subsequent actions and whether the conditional ledger and control structure outperform the same information or ordinary active design. Strong baselines match the result, motivating less unnecessary complexity and more model computation on unknown rivals and action gaps.
02
The actual literature-loop output
Authorized literature development uses six model calls, three local acquisitions and 137,456 reported tokens. Source IDs, action fields, context and acquisition-receipt gaps require development repairs; failures and costs remain recorded. The final mechanism concerns stagnation after refutation and remains PROPOSED_UNTESTED. Its causal experiments on generation, retrieval and selection have not run.
03
Eight states across two computational families
The diagnostic uses four polynomial-response states and four BFS/DFS trajectory states. Each has two response worlds requiring different follow-up probes. Public predictions, semantics, budgets, code and prompts freeze before the formal batch. It uses a computational host rather than the paper corpus. Each condition makes one follow-up decision per world, without running full multi-branch frontier scheduling.
04
Five conditions and matching strong baselines
A uses a conditional ledger and focus fields, C a same-time flat presentation of the same information, and E ordinary active design. All pass 8/8 pairs and 16/16 worlds. Fixed actions B and disconnected feedback D pass 0/8 pairs and 8/16 worlds. E needs no model calls when public predictions are provided. Feedback-based selection is supported; separate ledger or harness performance gains are unestablished.
05
Source retrieval and bounded-context repairs
Retrieval replaces short-word-dominated substring matching with whole-word matching, IDF, term-frequency saturation and length normalization, downweighting references. Scientific queries stay separate from operational prompts and retain exact query windows. Context protects pending questions and actions, compresses repeated observations and budgets actual serialization. Later proposals can cite reading receipts, which remain evidence-acquisition records.
06
Conditional updates, stopping and action registration
Attested observations exclude incompatible rivals while unknown or missing predictions stay unresolved. The frontier preserves lineage and deferred obligations; actions over old rival sets are rechecked. A historical question can close after one rival remains. Executed new actions enter the known catalogue, preventing repeated alias rejection. Direct compilation of existing actions remains a next-version plan.
07
Frozen results and failed-development costs
The formal comparison uses forty model calls, 790,465 measured tokens and 160 host executions; shared preparation and preflight add 208 executions. An earlier naming cue stops a development batch, whose seven launched calls retain failed-development costs rather than joining formal successes. Results belong to frozen source; later contract repairs use regression checks and are not presented as rerun model results.
08
Inputs, readable reports and audit
Literature inputs include a source manifest and bounded text; diagnostic inputs include frozen states, response worlds, conditions and budgets. Outputs retain sources, rivals, actions, receipts, frontier, costs and stop decisions in JSON, Markdown, HTML and signed projections. Source hashes lock identities and locations, while support relations still need semantic review. Local HMAC checks continuity rather than substituting for independent scientific judgement.
09
The next selective-repair study
The proposed next design selects and compiles existing discriminating actions directly, calling the model for mismatches, missing rivals or action aliases. It would separate stable QuestionSpec from changing QuestionState and diagnose stagnation through generation, retrieval and selection interventions. New actions need language boundaries, predictions, attestation and observed counterexamples. These prospective contracts are not promoted as implemented 7.3 mechanisms.
10
Where current evidence ends
The release reports 2,138 passes, 41 skips and no failures, with static checks passing. The eight states reuse a developed generator and seeds rather than forming unseen real-domain families; full human time and model dollar costs are unmeasured. Evidence covers proposals, acquisition, bounded computational feedback and engineering contracts. Open-domain actions, long-horizon graph gains, transfer, originality and RSI still need independent experiments.
Reading example · Explaining the workflow
The diagnostic decision logic
Two responses under a frozen state and shared public action predictions.
- 01
The first observation leaves different compatible rival sets.
- 02
A and C use the same response information to choose valid probes.
- 03
E directly maximizes rival-pair separation from public predictions.
- 04
Fixed or disconnected B/D choices cannot resolve both worlds together.
Per-state results and costs for five conditions; feedback is supported, separate harness gains are not established.
An explanation of the recorded computational diagnostic, adding no new successes or real-domain generalization.
Changes in this version
Evaluation evidence
- Earlier approach
7.2 records development repairs on one state pair.
- This version’s approach
7.3 freezes eight states and five conditions, retaining matching strong-baseline results.
Separate useful feedback from independent control-structure gains.
Evidence acquisition
- Earlier approach
Reading receipts and query contexts have interface gaps.
- This version’s approach
Exact source windows, typed acquisition receipts and serialized budgets enter the loop.
Make acquired text citeable by subsequent proposals.
Architecture direction
- Earlier approach
Admitted known actions still use the model proposal contract.
- This version’s approach
Plan direct compilation of existing actions and reserve model use for rival and action gaps.
Ordinary active design already matches the diagnostic result.