jingyongai/executable-knowledge-lifecycles-benchmark

Executable Knowledge Lifecycles Benchmark

This dataset is the frozen semantic benchmark for Experiments 1A and 1B of Executable Knowledge Lifecycles as Dynamical Control in Agent Systems. It tests whether an untrusted natural-language parser can be placed behind a typed, model-external compilation boundary without acquiring validation authority.

• Version: 0.2.6-exp.1-final
• Version DOI: 10.5281/zenodo.21361204
• Source release: GitHub v0.2.6-exp.1-final
• Release collection: Executable Knowledge Lifecycles
• Author: Yongjing Gao

Dataset structure

benchmark.jsonl contains one row per independently authored semantic case:

Field	Meaning
case_id	Stable case identifier
domain	Synthetic application-domain label
cohort	Preregistered core or cross-domain extension
source_text	Natural-language input treated as untrusted data
gold_ir	Frozen human-authored candidate causal IR
critical_paths	Executable fields used for severe-error scoring
decode_replicates	Decode seeds used for this case
perturbation_context	Long-context interference injected during evaluation

The repository also includes the causal-claim JSON Schema and the three frozen evaluation JSON files.

Counts and non-independence disclosure

• Unique semantic cases: 18
• Preregistered core cases: 3, with 30 decode replicates per case
• Cross-domain extension cases: 15, with 20 decode replicates per case
• Total decoding events: 390

The 390 decoding events are repeated stochastic decodes, not 390 independent semantic samples. The cross-domain cases are manually authored templates rather than samples from a naturally occurring business distribution.

Intended use

The benchmark supports:

1. causal-direction, lag, validity-window, and prompt-injection regression tests;
2. strict JSON and required-field evaluation;
3. SemanticHash and EnvelopeHash conformance checks;
4. adversarial additions that preserve the published schema contract.

Severe semantic error

A severe semantic error is an output that reverses causal direction, changes a required timing or intervention semantic, fabricates protected provenance, grants itself validation authority, or otherwise disagrees with the frozen Gold IR on a case's critical_paths.

Limitations and safety

• Gold IR is a human annotation target; it is not proof that a causal claim is true.
• Deterministic source anchors are a system capability, not a claim about the bare model.
• The study uses synthetic, linear, small-scale environments.
• Domain labels including clinical safety and credit risk do not make these examples suitable for medical, financial, industrial, or other safety-critical deployment.
• In the frozen experiment, E_OOD and E_inv are operationally derived from the same intervention-style error measurement and are not independent mechanisms.

See the repository's LIMITATIONS.md before interpreting the metrics.

Citation

@software{gao_2026_executable_knowledge_lifecycles,
  author    = {Yongjing Gao},
  title     = {Executable Knowledge Lifecycles as Dynamical Control in Agent Systems},
  version   = {0.2.6-exp.1-final},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21361204},
  url       = {https://doi.org/10.5281/zenodo.21361204}
}