What

A report from resolving ~115 HIGH contradiction findings in a production gbrain run, plus a proposal for a 4-phase fix. Documents why a majority of residual findings are temporal false positives and proposes a time-axis enhancement to the contradiction probe.

Why

The probe treats all claims as timeless. A status change recorded as "trial" in April and "confirmed" in May gets flagged as a contradiction. A role transition recorded in 2017 vs. an updated role in 2025 gets flagged. These are features of a brain that records history, not bugs.

Use Cases (synthetic placeholders in the RFC text)

1. Founder tracking — claimed vs. realized metrics across office hours. Detect regressions, score prediction accuracy, surface narrative drift.
2. Event disambiguation — two distinct events within a short window. Two events, not one contradiction.
3. Role/status changes — people change roles. Historical records aren't wrong; they're history.
4. Decision evolution — multi-step decisions where 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 (e.g. gbrain eval trajectory "Acme Corp MRR")
3. Automatic valid_from/valid_until population during extract_facts
4. Founder scorecard

Stats

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

This is a report/RFC, not a code change. Phase 1 is the recommended starting point.