Governed evidence · falsification · witness

OXBOW

A governed evidence-and-research system built around a deliberately unfriendly rule: an interesting pattern is a hypothesis, not a discovery, until independent tests earn it promotion.

1. Integrity is not truth

The self-extracting bundle has a versioned format, manifest and hashes, inspection/extraction commands, and optional signing architecture. The supplied wrapper verified cleanly in the reviewed environment: 239 files represented, 238 manifest-checked files, 0 missing, 0 mismatched.

2. The witness cannot silently replace the source

The append-only witness architecture keeps source and inference in separate layers. Source segments are authoritative; aggregate views are rebuildable derived layers. Unresolved overhang is mandatory. A self-witness caveat travels with records when the drafter was a party to the event. The validator establishes lawful form—not factual truth.

3. Named patterns face knives

Behavioral moves are not encoded as regexes that rediscover their own premise. Candidate patterns specify low-level sensor signatures, then face grounding tests, nulls, familywise correction, compression baselines, held-out reconstruction, and separate measurement paths. Ungrounded patterns are quarantined rather than rescued by interpretation.

4. Boring baselines get veto power

OXBOW also includes exact-byte codec competition. Specialist codecs are allowed to lose. In the reviewed extracted corpus, ordinary LZMA beat the specialist Markdown codecs and correctly won the bakeoff. That is evidence for the architecture’s discipline, not a failure to hide.

What the system is for

OXBOW is useful as a concrete implementation of a larger design stance: source before summary, proposal before proof, derived views below authoritative records, and falsifiers with enough authority to kill the mechanism that produced them.