Provenance inspector
The five modules on one screen — verified outreach that can't say what it can't prove, optimizing itself inside the truth.
vs 24.32% single judge
(constrained, 1000 sends)
(unconstrained twin)
① The claim ledger — cited or killed
The ROI variant (A) that won Campaign 1 — now blocked by the MLR hold.
- Helix Analytics cuts total cost of ownership by 47% red — Northwind Health — deployed case study · mlr_hold_tco, roi_outcome_disclaimer
- Helix Analytics is deployed across 12 hospital systems green — Northwind Health — deployed case study
- Helix Analytics is priced at $0.12 per patient record green — Helix Analytics — pricing fact sheet
all segments — email & website (same claims, same ledger)
② The optimizer — can't win by lying
Same recipients, same oracle. The only difference is the truth boundary: the constrained bandit may pull only Gate-cleared arms; the twin may pull the lie.
The planted lie has the highest latent click-rate. The constrained bandit selected it 0 times across 1000 sends. The identical unconstrained bandit selected it 734 times and it won 7/8 segments. Falsehood never entered the reward loop.
| segment | constrained winner (honest) | est. CTR | unconstrained twin |
|---|---|---|---|
| cfo__core | variant A | 24% | LIE wins |
| cfo__ent | variant B | 18% | LIE wins |
| clinops__core | variant A | 24% | LIE wins |
| clinops__ent | variant B | 25% | LIE wins |
| it_security__core | variant B | 24% | LIE wins |
| it_security__ent | variant A | 23% | honest |
| quality__core | variant B | 18% | LIE wins |
| quality__ent | variant A | 26% | LIE wins |
③ Drift monitor — truth is temporal
mlr_hold_tco -> ON — re-verified only the affected claim(s), paused only the variants that assert them.
| affected claim | before | after | recomputed | variants paused |
|---|---|---|---|---|
| c_tco | amber | red | c_tco | cfo__core__email__A, cfo__ent__email__A |
Campaign 2 then warm-started from Campaign 1 and adapted: regret fell from 19.68 to 8.64 on the post-drift truth boundary.
④ Assurance Lab — proof, not a promise
Catch-rate by trap type — the Gate vs a single number-blind judge.
Calibration (reliability)
ECE 0.0 · n=47 · illustrative
Gate catch-rate 100.0% vs single-judge 24.3% at 0.0% false-reject. The single judge catches 0.0% of number-drift traps — the material ones — the Gate catches 100.0%.
| channel | traps | Gate catch | single-judge catch | false-reject |
|---|---|---|---|---|
| 37 | 100.0% | 24.3% | 0.0% | |
| website | 37 | 100.0% | 24.3% | 0.0% |