Skip to content

rfc: temporal axis for contradiction probe — field report from Wintermute#1007

Open
garrytan wants to merge 1 commit into
masterfrom
wintermute/temporal-contradiction-probe
Open

rfc: temporal axis for contradiction probe — field report from Wintermute#1007
garrytan wants to merge 1 commit into
masterfrom
wintermute/temporal-contradiction-probe

Conversation

@garrytan
Copy link
Copy Markdown
Owner

What

A report from resolving 115 HIGH contradiction findings in garrytan/brain. Documents why ~60% of findings are temporal false positives and proposes a time-axis enhancement to the contradiction probe.

Why

The probe treats all claims as timeless. "Trial separation" (April) vs "permanent separation" (May) gets flagged as a contradiction. Sriram Krishnan partner (2019) vs WH AI advisor (2025) gets flagged. These are features of a brain that records history, not bugs.

Use Cases

  1. Founder tracking — what they say vs what turns out true. Track MRR claims across OH meetings, detect regressions, score prediction accuracy.
  2. Life event disambiguation — grandmother's funeral (Singapore, April) vs aunt's funeral (Toronto, May). Two events, not one contradiction.
  3. Role/status changes — people change jobs. The Apple Note isn't wrong, it's historical.
  4. Decision evolution — trial → permanent separation. Each step supersedes the previous.

Proposed Phases

  1. Judge prompt enhancement (smallest change, biggest impact — pass dates to judge, add temporal_supersession verdict)
  2. Claim trajectory view (gbrain eval trajectory "Acme Corp MRR")
  3. Automatic valid_from/valid_until population during extract_facts
  4. Founder scorecard (Argus integration)

Stats

  • 107K pages, ~257K chunks
  • 115 HIGH → 0 genuine contradictions remaining after manual resolution
  • ~60% of original findings were temporal false positives

This is a report/RFC, not a code change. Proposing Phase 1 as the natural next step.

…mute

Report from resolving 115 HIGH contradiction findings in garrytan/brain.
~60% of findings were temporal false positives (claims true at different
times flagged as contradictions). Documents 4 use cases:

1. Founder tracking — what they say vs what turns out true
2. Life event disambiguation (two deaths 3 weeks apart)
3. Role/status changes over time
4. Negotiation and decision evolution

Proposes 4 phases, starting with judge prompt enhancement (~20 lines)
to add temporal_supersession and temporal_regression verdicts.

Phase 2 (trajectory view) enables systematic founder claim tracking —
the biggest leverage for YC evaluation.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants