Confidence and verification
How the blind second-pass audit works, and what an UNVERIFIED slip means.
solvemax reports how sure it is at three separate levels, deliberately kept apart so one can't paper over another.
Win rate & robustness
The win rate ("wins X% of 5,000 simulated futures") comes from the Monte Carlo stress test: the locked weights and scores are jittered (weights ±30%, scores ±1) 5,000 times, and the leader's win rate across all 5,000 is reported directly. A leader that only edges out narrowly is flagged CLOSE CALL; a comfortable, repeated winner is flagged ROBUST. This is the honesty check against a fragile-looking win being sold as a sure thing.
The scored-matrix fold also reports flip-points (the smallest weight increase on each criterion that would dethrone the leader) and a minimax-regret cross-check (a second scoring method — "assume the worst case, minimize your maximum regret" — which sometimes disagrees with the win-rate leader; when it does, the slip says so explicitly).
Confidence %
The confidence number printed on the slip is the engine's own calibrated estimate, capped when
the underlying result is fragile. It reads alongside — never instead of — the robustness flag.
Independent verification (the audit)
After the deep run produces a winner, a second, independent model pass — by default, a different model in the routing pool than the one that ran the deep analysis — receives only the bare facts (the decision, the options, your clarifying answers). It sees no scores and no conclusion, and derives its own winner from scratch. The slip reports one of four outcomes:
| Verdict | Stamp on the slip | Meaning |
|---|---|---|
CONFIRMED | CONFIRMED | The independent pass landed on the same winner. |
CONFIRMED_WITH_AMENDMENTS | CONFIRMED* | Same winner, with a noted caveat. |
DIVERGED | CONTESTED | The independent pass picked a different winner — reported, not hidden. |
UNVERIFIED | UNAUDITED | No independent audit completed — see below. |
What UNVERIFIED means, honestly
Every run has a fixed time budget. If the verify step can't finish inside it (a slow or congested
model provider, most often), solvemax's default behavior is to ship the slip anyway, flagged
UNVERIFIED, rather than fail the whole run and refund your credit for a deep analysis that
already completed successfully. The winner shown in that case is the deep run's own pick, carried
over unaudited — not a second pass's confirmation. The slip is explicit about this both in the
visual stamp (marked UNAUDITED, not CONFIRMED) and in the verification section's text.
This is a deliberate trade-off, not a bug: a completed-but-unaudited verdict is more useful to
you than no verdict at all, as long as it's labeled honestly — which it always is. UNVERIFIED
never silently renders as CONFIRMED.
If verify fails outright rather than merely running out of budget, or any other stage of the pipeline fails, the run fails and the credit is refunded automatically — see Stopping a run.