bdistill-validate
When to use
- After bdistill-extract, before exporting -- filter out unreliable entries so your exported KB only contains claims the model reproduces consistently.
- When thresholds in your KB seem suspiciously precise -- a rule that says "trigger at 47.3%" deserves a consistency check. Real knowledge reproduces; confabulations drift.
- When building rules for a deterministic system -- if the rules will drive automation or monitoring, every numeric threshold must survive rephrasing. One unstable number can cascade into bad decisions.
- Before bdistill-operationalize -- only promote entries to production monitoring if they are stable. Unstable entries need re-extraction or external sourcing.
Input contract
required:
domain: string # Domain name matching your KB (e.g. "aml-compliance")
optional:
source_type: string # "knowledge" or "rules" (default: "rules")
num_rephrases: int # Number of rephrased questions per entry (default: 5)
max_entries: int # Maximum entries to probe (default: 20)
Output contract
More from francyjglisboa/bdistill-skills
bdistill-extract
Extract structured, adversarially validated domain knowledge or IF-THEN decision rules from AI training knowledge. Builds a compounding knowledge base — one file per domain, deduplicated across sessions. Triggers on "extract knowledge", "build KB", "distill", "extract rules", "decision thresholds", "what do you know about". Outputs JSONL entries to {domain}.jsonl.
1bdistill-export
Export a bdistill knowledge base into any format — system prompt for Claude Projects/Cursor/Copilot/ChatGPT, Python harness module with build_prompt(), JSON for agent consumption, Excel with quality color-coding, audit checklist CSV, or fine-tuning JSONL. Triggers on "export", "system prompt", "harness", "training data", "Excel export", "export for Claude Project". Outputs file on disk.
1bdistill-xray
Probe any AI model's behavioral patterns across 6 dimensions — tool use, refusal boundaries, formatting defaults, reasoning style, persona stability, and grounding/hallucination resistance. The model probes itself, no API key needed. Generates a visual report. Triggers on "x-ray", "probe behavior", "behavioral analysis", "model evaluation", "how does this model behave". Outputs behavioral profile with scores.
1bdistill-predict
Assemble structured predictions with decomposed evidence, adversarial self-challenge, and calibrated probability. Supports binary YES/NO mode (prediction markets, any yes/no question) and directional mode. Optionally recalls from your KB and searches the web for current data. Triggers on "predict", "forecast", "what happens if", "probability of", "will X happen". Outputs prediction card with evidence chain.
1bdistill-operationalize
Connect exported rules to live data for automated monitoring. Loads a bdistill rules export, fetches current data from free APIs or local feeds, contrasts each rule's conditions against reality, and reports which rules triggered with current values and impact estimates. Works with any domain — weather, market, compliance, clinical. Triggers on "operationalize", "monitor", "check against live data", "contrast rules", "what's triggered". Outputs decision report.
1bdistill-abstract
Extract structural rules from one domain, abstract to bare skeletons at three granularity levels, then re-instantiate in other domains to discover non-obvious cross-domain correspondences. Filters with mandatory web-grounded novelty check AND adversarial validity challenge AND reverse round-trip validation. Triggers on "abstract rules", "cross-domain", "structural analogy", "what pattern in X applies to Y", "transfer rules between domains". Outputs validated cross-domain rule correspondences with structured testable predictions.
1