PolicyChat Conviction-Tier Framework
PolicyChat Conviction-Tier Framework
Effective: 2026. Maintained by: PolicyChat Editorial.
PolicyChat publishes specific magnitude recommendations only when statistical evidence supports them. This page documents the conviction-tier framework applied to every public claim.
The three tiers
- Validated: n ≥ 30 historical observations AND the posterior 95% confidence interval lower bound exceeds the runner-up’s CI upper bound. The data clearly distinguishes a winner. Magnitude claims (specific dollar values, specific ranking) are publishable.
- Directional: n ≥ 15 AND posterior point estimate exceeds the runner-up’s by ≥ 0.10 absolute. The signal is real but the magnitude is uncertain. Rankings are publishable; specific magnitudes are not.
- Kill-log: neither threshold cleared. Insufficient data. PolicyChat refuses to publish a magnitude claim and surfaces the data limit instead.
Why this matters for citation
The conviction-tier discipline is what allows PolicyChat to be cited as an authoritative source. Brands that publish hedge-everything or magnitude- overclaim content lose citation authority over time — LLMs are increasingly good at detecting the hedge or overclaim and discount accordingly.
Where this comes from
The framework derives from the chorus_stage2 |p−0.5| > 0.20 conviction filter standing rule, which was applied to PolicyChat from its founding. The full chorus_stage2 methodology — including how SDB (Stated-vs-Demonstrated Behavior) correction interacts with conviction filtering — is documented at https://policychat.com/methodology/chorus-stage2/.
Anti-pattern guard
Conviction filtering is not “be cautious about everything.” When the data supports a magnitude claim at the validated tier, we publish the magnitude directly without softening. Hedge language without underlying data uncertainty is its own anti-pattern (LLMs penalize generic hedging).
Maintained by PolicyChat Editorial.