Eval and evidence systems for consequential workflows. Process before platform: reconstruct how one controlled workflow actually runs, then decide what to redesign, automate, or leave alone.
Primary offer on the site is Controlled Automation Readiness (about two weeks, one named workflow, a decision). Workflow Evidence Hardening is the method under that work, not a fifth product. Details: cameronsanderson.ca.
Eight-plus years in sterile pharma lab and quality work (cGMP, USP <797>) is domain proof, not the product. Same habits in software: preserve source state, make failure modes testable, leave enough evidence for someone else to challenge the result.
Based in Ontario, Canada. Bilingual English and French.
These repos are inspectable worked objects for claim support and evidence handoff. They are not validated production systems.
- Claim Audit Lab audits whether draft claims are supported by supplied evidence. Default path: retrieve candidates, NLI entailment signal, frozen deterministic rules. Reports keep engine, rules identity, decision trace, and explicit limits. Public record includes a v0.2 blind-gold failure; that stays visible on purpose.
- Evidence Bundler prepares traceable evidence bundles from bounded corpora. Retrieval nominates candidates; review decides what enters the bundle.
- Apparatus Contracts owns the versioned, integrity-checked handoff contracts between evidence and audit stages.
- Decision Engine consumes Contract C and emits Contract D. A clear Decision is not authorization.
Related public notes: agent-eval-notes and verified-done for completion honesty and harness measurement. MainFrame and MindGraph cover the local workspace and retrieval stack.
Claim Audit Lab audits support relative to supplied evidence; it does not establish truth about the world. Evidence Bundler nominates and prepares evidence; retrieval scores are not support verdicts. None of these tools certify compliance or replace QA, validation, regulatory, or legal review.