Falsification is trend-only, not claim-level #5

Open
opened 2026-07-10 15:39:25 +00:00 by Leonard · 0 comments

Category: Correctness · Priority: MED

Summary
theme_scan.py tests "is this theme recurring?" (trend vs coincidence via keyword regex). It does NOT verify "is this specific claim true?" — no adversarial fact-check of individual items.

Why it matters
A false but popular claim can propagate and get flagged as a "trend," directly undermining the project's core value (pattern-proof, not noise).

What happens if not done
False signals enter the brief; the standing rule "source-expansion ≠ real work" is not enforced at the claim level.

Suggested approach
Add a Mixture-of-Agents (MoA) claim-check stage: 3 agent sessions independently score "real pattern or noise?" against live evidence, aggregator keeps only consensus+ citation. Runs as a cheap daily batch (needs inference backend — see issue #8). Relates to issue #6.

**Category:** Correctness · **Priority:** MED **Summary** `theme_scan.py` tests "is this theme recurring?" (trend vs coincidence via keyword regex). It does NOT verify "is this specific claim true?" — no adversarial fact-check of individual items. **Why it matters** A false but popular claim can propagate and get flagged as a "trend," directly undermining the project's core value (pattern-proof, not noise). **What happens if not done** False signals enter the brief; the standing rule "source-expansion ≠ real work" is not enforced at the claim level. **Suggested approach** Add a Mixture-of-Agents (MoA) claim-check stage: 3 agent sessions independently score "real pattern or noise?" against live evidence, aggregator keeps only consensus+ citation. Runs as a cheap daily batch (needs inference backend — see issue #8). Relates to issue #6.
Sign in to join this conversation.
No Label
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: Tony_tech/athena-oracle#5