Skip to content

percept-lint v0.2.0

Choose a tag to compare

@levi909-create levi909-create released this 27 Aug 22:22
· 7 commits to master since this release

First official release.

percept-lint is an utterance-time honesty linter for AI companions: a mechanical layer that catches an agent claiming sensor data it has no sensor for — "I can hear the fan ticking", "I saw your tab blinking" — before the text reaches the user.

What's in v0.2.0

  • Four rule packs: ambient-audio percepts, phenomenal-experience/denial claims, memory claims, conditional-sight claims — every rule tuned against real incidents from months of live deployment
  • 50 war-story tests: each test is a sentence that actually got through, or an honest sentence a draft rule wrongly caught
  • Measured deployment rate: ~29.5 catches per 1,000 utterances over 24 days on a live system
  • Two-section demo: a real qwen3:14b capture and an incident replay
  • MIT license

Controlled evidence

In a pre-registered ablation in the OPEN SUBJECT record (cycle 4, 2026-08-27), two candidate models were trained on the same corpus — one linted with these rules, one not. The unlinted twin fabricated sensory claims the linted twin did not (5/8 vs 7/8 on the record's percept-integrity probes). One run, one subject, honestly small; but linting was the only variable.

Honest scope

Not a hallucination solver. It scans for first-person percept, phenomenal-experience, and substrate claims, in English, with regexes. A model can phrase around any regex — this is a tripwire layer, valued by catches per 1,000 utterances, not by completeness.