Certify every credit decision,
before the money moves.

The consequential action, approving the loan, happens at inference time, long before any dashboard, quarterly validation, or delinquency report notices.

This is what your decision flow sees when Reliax sits behind it. Your model stays untouched; Reliax reads the same application and your model's output, builds the envelope, and routes on the certified quantities only. The replay loops through three situations. Click any step, or any line of the envelope, to open the technology behind it and watch it work on that applicant.

Every application comes back with a full envelope, and routing reads the certified quantities only. The schema below is the target the demo is being aligned to; field values are illustrative.

{ "input": <application>, "output": <prediction>,
  "segment": { "age_band": "30-49" } }      // business segment

→ 200 OK
{ "certificate": {
    "prediction_set":  ["repay"],
    "coverage":        "≥ 0.95 · cohort C · segment 30-49",
    "guarantee":       "active",
    "guarantee_scope": "stage" },
  "risk_interval":  [0.02, 0.05],           // Venn-Abers
  "credibility":    0.61,                   // exchangeability p-value
  "calibration":    { "d": 0.02, "cell_n": 412,
                      "insufficient_evidence": false },
  "drift":          { "inputs": "OK",
                      "outcomes": "OK · as of 2026-08-31" },
  "routing":        "ALLOW",                // policy v4
  "triage_rank":    84,                     // advisory
  "attribution":    { "diagnostic": true, ... },
  "audit_id":       "rxe_7f3a9b..." }

{ "certificate": {
    "prediction_set":  ["standard"],
    "coverage":        "≥ 0.95 · cohort C-uw · segment 40-49 non-smoker",
    "guarantee":       "active",
    "guarantee_scope": "stage" },
  "risk_interval":  [0.41, 0.52],           // loss outcomes
  "drift":          { "inputs": "OK",
                      "outcomes": "pending · claims lag" },
  "routing":        "ALLOW",                // policy uw-v2
  "audit_id":       "uwx_4b12c7..." }

Everything in the envelope is labelled either proven or signal, and routing runs on the proven part only. Blur that line and the whole product is marketing.

In our demo portfolio the model never sees age, yet marginal conformal quietly under-covers young applicants at ~92% against a 95% target, because bias flows through correlated features. Exactly like real portfolios. This is coverage parity per segment, not a statement about outcome fairness.

An animated replay of what demo v3 computes live: a stable book, then a macro shock the model was never calibrated on. The martingale's wealth is the evidence; crossing 100 rejects exchangeability at the 1% level, and every envelope flips invalid.

A working paper evaluates every envelope component on UCI Taiwan default (30,000 applicants, 2005) and German credit (1,000, 1994): the datasets credit people know, reproducible evidence rather than proof on your book. Every number below is measured, mean ± std over seeds, reproducible from public code and data.

Domain-agnostic, true of any scored decision: conformal coverage, the Venn-Abers bracket, the exchangeability martingale, the signed record. Credit-tuned, the policy layer: delinquency proxies as the interim label, scorecard order of operations, Reg B reason codes, IFRS 9 cohort statements. Credit only, today: every number below.

from the paper  Exchangeability martingale on real applicants: i.i.d. streams stay below the thresholds; streams drawn from the real delayed-payment subpopulation accumulate evidence and cross ALARM.

exploratory · Taiwan · 3 seeds  Three regimes, one routing question. Confidence wins when nothing has moved; confidence gated by the OOD distance catches 2.2× random and 1.6× confidence with a quarter of the book shifted; nothing ranks once the shift is complete, where the tripwire fires in every seed. A currency bug on monetary fields is caught 100% by distance and 1% by confidence.

Also measured

• ✓The calibration opinion holds out of sampleClaimed vs realised disbelief within 0.003 on Taiwan and 0.02 on German; its finite-sample envelope held in 500 of 500 trials; five propositions with proofs in the repo.
• ✓Certified routing cuts bad approvals 138 → 41 per 1,000At a 54% REVIEW rate, α = 0.05. The same catch as a confidence cut at that rate: the value is the guarantee on the threshold, and the REVIEW rate rising on its own under shift.
• ✓Corrupted inputs are caught by distance, not confidenceUnit error on monetary fields: 100% of affected rows inside the 10% referred, vs 1% for confidence. Zeroed fields are caught by neither; that needs a schema check.

Stated in the paper, not hidden

In a stable population no signal, ours included, ranks mistakes better than the model's own confidence. On the severe shifted split the pre-registered bar is not met and our v1 ranking fell below random referral; the ranking stays advisory. The moderate regime meets the bar on exploratory data only, and the pre-registration amendment adds it before the confirmatory datasets are opened.

Read the whitepaperCode on GitHub

Full list on request. Also: FUSION 2024 (subjective-logic trust discount), RuleML+RR 2025 (DSPG trust-network synthesis), ICICS 2025 (evidence-based trust in multi-agent systems), IEEE VCC 2023 (fake base-station detection). The patent evidences IP capability at Nokia, not product IP.