How each status is set
The AI reads and quotes. Fixed rules set the status. Every AI run is traced in PRISM.
Real runs, traced in PRISM
How a status is set
- 1AI, through NicrronReads each messy pass note and turns it into one to three standard reasons.
- 2akta.pro and Hacker NewsReturn news, company profile updates and launch posts dated after the pass.
- 3AI, through NicrronLabels each item as supports, contradicts or irrelevant, and must copy a quote word for word.
- 4Fixed rules, no AISet each status and the conviction score. Same input, same answer, every time.
- 5PRISMTraces every AI run and checks it against our rules. Breaks are caught before a partner sees them.
Rules checked in PRISM
Claude Opus 5.5, through Nicrron
- Every "supports" label must include a word-for-word quote found in the source.
- Evidence must be dated after the pass date.
| Version | “Supports” labels | Broke the quote rule | Broke the date rule | What changed |
|---|---|---|---|---|
| labeler-v1 | 17 | 8 | 7 | Stitched quotes together with "..." and cited undated company-profile text as proof. |
| labeler-v2 | 11 | 0 | 0 | Quote must be one passage copied character for character; evidence dated before the pass, or undated, is rejected. |
Same 47 real Akta articles for both versions.
How conviction is scored
- De-risking: how many pass risks are cleared50%
- Thesis fit30%
- Traction: positive signals since the pass20%
These weights are a starting point, not a finding. The breakdown is always shown next to the conviction score, and the status comes before the rank.