1. Calculator, not narrative
Characterization runs in the GSRF Audit Workbench using official filter implementations and metric formulas (e.g. osc31 = std of series − rolling mean of length 31). Agents do not hand-wave advantages into existence.
2. Pre-declared rules
Excellence packs declare win-rate and mean-advantage thresholds before reading the scoreboard (example: actuator-mock oscillation win_rate ≥ 0.8 and mean_adv ≥ 10% → PASS). Failures stay in the manual.
3. Claim ladder (public)
The Workbench / claim ladder lists allowed vs refused marketing wording. Research posts link back to Evidence as the human-readable ledger.
4. What we refuse on the open web
- “Always beats EMA”
- Industrial + quantum LER scoreboards mixed as one product story
- ROI / miracle false-alarm elimination as a detector
- Circuit-level hardware guarantees without gated materials
5. How notes relate to packs
| Public page | Typical pack source |
|---|---|
| Evidence · Identity / frequency | characterization v4 (step, sine gain) |
| Peaks / MAD scoreboard | excellence hunt v2 |
| Delay red light | characterization v4 delay |
| Fixed-FA residual | probes v5 alarm |
| Adaptive x* | probes v5 adaptive |
Why this exists
Same philosophy as independent evidence sites: publish the method, publish the limits, refuse paid miracles. Traction comes from trust + long-tail answers—not from denser adjectives.
See also: Claim ladder · llms.txt · agents.txt