01
From correct inference to admitted execution
7.2 connects a live model and checks whether observations lead to different next measurements. Correct inference can still fail execution: explanatory fields inside action definitions or incorrect descriptions of excluded rivals trigger rejection. The version therefore studies how intentions, exact fields and executable semantics connect, separating research decisions into inspectable stages.
02
Separate proposal provider and controller
CodexCLIProvider uses existing local authentication for structured outputs, preserving requests, configuration, proposals and reported token usage. Proposal mode disables tools, plugins, connectors, hooks and delegation. The model suggests JSON while Aletheia validates and executes. Successful calls use the client default; the resolved model identity was not returned and remains unreported.
03
Two observation worlds under one public state
The paired pilot uses one frozen four-rival state. Both worlds share the public prediction table, action language and remaining budget. An amber response leaves r7/r2 and requires slope; violet leaves r9/r4 and requires phase. The model sees only its observed response, while hidden rival indices and evaluator-computed survivors stay on the evaluation side. This makes feedback dependence testable.
04
Evaluate each transition separately
The host performs a common first measurement and records its receipt. The provider proposes compatible rivals, a revision contract and a follow-up action. The controller separately checks rival sets, action choice, evidence links, question transition, semantics and costs. Follow-up execution waits for the complete proposal contract. Changing an ID or wording is not counted as success; probes and responses enter the record.
05
What the initial failure explains
Under the initial interface, both next actions are correct, but transition contracts pass 0/2 and host execution contracts 1/2. The model adds explanatory fields to an operation and lists already refuted rivals as removed live rivals. The failures call for precise action and transition semantics rather than weaker scientific gates or allowing a correct answer to erase an invalid execution path.
06
Same-case repair through clearer fields
The second input clarifies the exact allowed action fields and the meaning of revision fields. Both actions, transitions and host contracts pass 2/2. Two model calls accompany four host measurements: initial measurements and conditional follow-ups. Independent reassessment preserves the passes. Costs from both same-case runs remain recorded; these are development repairs, not independent held-out performance trials.
07
Invocation, timeouts and usage contracts
Prompts are written before process launch to prevent blocked input from bypassing time limits, and every exit path cleans up the process group. Event bytes, calls and wall time are bounded. Tokens are reserved, then settled against provider usage; overruns remain recorded and stop the run. Failed proposals retain their charged usage; dollar prices are unconfigured. Tests cover tool activity, lingering processes, invalid references and tampered reports.
08
Signed pilots and readable reports
Inputs comprise frozen cases, shared predictions, a world-specific receipt, budgets and model configuration. Outputs distinguish correct choices, passed contracts, attestation and actual execution, retaining proposals, identities and usage. Readable pilot reports match signed records and frozen inputs field by field. Event continuity checks engineering provenance; scientific validity remains a separate judgement.
09
Pilot complete; broader comparison still prospective
The historical release reports 2,110 passes, 41 skips and no failures, with Ruff, Mypy and version checks passing. A broader contract freezes eight states and five conditions: adaptive control, fixed actions, same-information reading, disconnected feedback and ordinary active design. At the 7.2 historical point that comparison is still prospective; the later 7.3 results belong on their own page.
10
The proposed executable-idea compiler
The next route would let a model express explanations, intentions and discriminating predictions, then compile them into host operations with types, units, costs and source bindings. Equal-cost controls include direct model reading, ordinary active design and graph editing, with compiler, update and frontier ablations. This remains a research proposal, without demonstrated open-domain action invention, transfer or separate harness gains.
Reading example · Explaining the workflow
Recorded paired development: amber and violet
Two observations under one four-rival state, with matched public predictions and budgets.
- 01
Amber leaves r7/r2; the model requests slope.
- 02
Violet leaves r9/r4; the model requests phase.
- 03
Initial correct choices still fail field or revision contracts.
- 04
After same-case field clarification, both follow-up measurements pass and execute.
Two admitted follow-up measurements and separate contract records, preserving development failures.
Reported same-case development, not an unseen-task success rate.
Changes in this version
Provider
- Earlier approach
The 7.1 report has no live-model calls.
- This version’s approach
7.2 adds Codex proposal mode and usage settlement.
Check whether a live model changes actions after observation.
Execution gate
- Earlier approach
A correct choice can still contain invalid transition fields.
- This version’s approach
Host execution waits for the complete proposal contract.
Check correctness and admitted execution separately.
Follow-up evaluation
- Earlier approach
One state pair can diagnose interface problems.
- This version’s approach
Freeze a follow-up eight-state, five-condition contract.
Further distinguish feedback effects from control-structure gains.