EVIDENCE
Evidence, expiry, and conditions
Evidence is not permanent memory. The map tracks its operational freshness.
TTL evidence decay
| Status | Operational meaning |
|---|---|
| Current | Can satisfy an evidence predicate. |
| Stale | Still visible, but needs review. |
| Expired | Cannot authorize the move. |
Separating responsibilities
The integration obtains and identifies evidence. The map owns freshness rules and transition conditions. The agent may interpret them, but it does not modify them within the same move.
Tested live, not just counted
We apply the same discipline here — evidence has a real state, not a guessed one — to our own product claims. The numbers on the “Why” page are not just tokens counted on paper: we ran live model calls in two conditions, on the same 5 decision points — the compiled slice from the map and the exact quoted raw FPF fragment for that decision. We compared what the model answered, not just how long the prompt was.
And we say clearly where the comparison stops: nobody would ever use the full raw spec directly — 2,247,567 tokens, beyond any practical context window. We tried it so this would not be a guess, and the attempt failed with “context length exceeded”. So you are not comparing the compiled map to something impossible to use. You compare it to the best raw variant that can actually run: the exact quoted fragment, trimmed to the sections relevant to that decision, not the whole document. Even against that carefully reduced variant, the compiled map won on size, speed, and match rate against the expected result.
| Date | Decision points | Compared conditions | Result |
|---|---|---|---|
| 2026-07-10 | 5 | compiled slice vs. exact quoted raw fragment, live | 80% vs. 40% match with expected result; 259× live cost gap per decision |
What stands out most is not just cost — it is accuracy: 80% versus 40% means the compiled version hit the expected result twice as often. We do not claim to know exactly why, end to end. The document offers a partial clue: the reduced raw fragment still carries a stricter FPF ontology and demands explicit authority structures where the compiled map is willing to continue — that is where the miss on role_conflict comes from. But that explains part of it, not all of it. We say it plainly: this is a measured observation, not a fully proven cause.
It is one round, on 5 points, with one current model — a measurement, not a universal proof. More rows will be added to the table as they appear, not promised all at once now.