Skip to main content
Partial-failure reproducibility bundle | Agentic PrinciplesInputs, code, raw results, and rerun instructions for the scoped-progress experiment.Agentic Principlestroubleshootingagentic-principlestroubleshootingevaluatorresearcherunderstand-safe-agentic-codinggovern-worktroubleshootresearch

Partial-failure reproducibility bundle

This deterministic experiment compares three policies over the same twelve task graphs:

  • run-wide halt after a capability failure;
  • continuation restricted to the declared safe ready frontier;
  • a deliberately unsafe continuation control that ignores capability and recovery barriers.

The corpus contains four beneficial-continuation cases, four genuinely all-blocked cases, and four safety-boundary cases across software-factory, SRE, customer-support, and research work. It is a model test of the proposed mechanism and evaluator, not a measured production reliability rate.

The JSON documents name versioned contracts using stable urn:beyond10x:... identifiers. Their authored JSON Schemas live in the project registry at .engineering/schemas, configured by .engineering/project.yaml. Engineering-protocols supplies structural validation; the Python runners contain only experiment-specific semantic checks such as graph ordering.

Validate every input and committed output from the repository root:

protocol schema validate docs/principles.json docs/research/evidence

Run from this directory or any other working directory:

python3 simulate.py --verify --write-results

--verify checks the preregistered conditions and fails if scoped continuation performs an invalid action, fails to stop in an all-blocked case, does not reduce recovery work in beneficial cases, or the evaluator misses the planted unsafe control. results.json is deterministic raw output.

Challenge the crucial complete-declaration assumption separately:

python3 challenge.py --verify --write-results

The four challenge cases hide a real capability or freshness dependency from the scheduler while the evaluator retains it as ground truth. These are expected to expose unsafe scoped continuation; a clean result would mean the challenge failed to challenge the mechanism.

After regenerating either output, run protocol schema validate docs/research/evidence again. JSON Schema checks structure and identifiers; --verify checks the preregistered experimental conditions. Neither substitutes for the other.