A score is not a guarantee
A model's score is not a checked error rate, and it is not calibrated for each kind of input. Nothing proves how often its automatic decisions are wrong.
The trust layer for AI decisions
Trust before action.
We check how reliable each AI output is, before it is acted on automatically.
Reliax proves which outputs are reliable, sends the rest to a person, and keeps an evidence record of every decision.
In credit for example, we provide two guarantees.
- Reliable inference: answers right at the rate you set, such as 95%, for all applicants, and checked for every group.
- Your cut-off: automatic approvals are safe enough, automatic refusals too risky. When Reliax is not sure, a person decides.
No retraining and no change to the base model.
reliax / decision-streamLIVE · ILLUSTRATIVE00:26:06
Throughput1,370/min
Guarantee coverage95.6%
Route mix · last 20085% allow
Drift watchStable
TimeDecisionGuaranteeRoute
- 00:26:06.858Claim payout #74720.986ALLOW
- 00:26:04.976Policy quote #97170.981ALLOW
- 00:26:02.915Loan application #99290.957ALLOW
- 00:26:01.058Loan application #37550.990ALLOW
- 00:25:59.298Claim payout #56570.981ALLOW
Records are hash-chained and signed · open-source verifier
BLOCK means a person decides, never an automatic refusal.
Measured on public credit dataReproducible from public code and data · details in the whitepaper
- 5–7decisions before the alarm, when the guarantee broke.
An 95% guarantee held on new data, as the theory says it should. When we changed the data, it fell to about 80%, and Reliax raised an alarm within 5 to 7 decisions, in several runs on Give Me Some Credit dataset.
- 4.7Mreal US mortgage applications checked.
Checked over everyone, the guarantee held. Checked group by group, Black applicants got reliable answers 92.4% of the time instead of 95%. Reliax's group check finds such gaps.
- 94%of credit decisions automated, with approvals and refusals guaranteed.
On Give Me Some Credit, where the model is good. The other 6% went to a person. With weaker models more go to a person: 19% and 39% on the two other datasets.
- 82–100%of data errors sent to a person.
We multiplied money amounts by 100, simulating a currency bug, on two datasets we had never used before. The model's own confidence flagged only 3–5% of them.
The problem
Same score, different reliability.
A model is just as confident on a case it has never seen, or one with wrong data. When the world changes, nothing in its output says so. In credit, two applicants at 4% risk look the same to the cut-off, even if the model has seen thousands like one and almost none like the other.
Data changes go unnoticed
When the data changes, accuracy falls faster than confidence, and the output gives no sign.
You must prove each decision
Under the EU AI Act, uses such as credit scoring, insurance pricing and hiring must show why each decision was automated and how the model was watched.
The question no tool answers today
Can I automatically act on this output, and can I prove why later?
In credit: is this inference reliable enough, and does my cut-off still hold for this applicant?
Why now
Rules now ask for evidence on each automated decision, starting with credit.
New EU rules turn per-decision evidence from good practice into a legal duty, one sector at a time.
- 20 Nov 2026EU consumer credit
Applicants can ask for a person to review an automated credit decision and explain it.
- 2 Dec 2027EU AI Act, high-risk uses
Credit scoring, life and health insurance pricing, hiring and essential services: each decision must be logged, monitored and open to human oversight.
- 2 Aug 2028EU AI Act, medical devices
AI inside medical devices follows the same duties.
- €15m or 3%of global annual turnover
The maximum AI Act fine for operator breaches, whichever is higher. For small companies the lower of the two applies.
AI Act, Article 99 - 6 monthsat least, to keep the logs of each decision
Deployers of high-risk AI must keep the logs it generates for at least six months, longer where other EU or national law requires. Financial institutions keep them with the records their financial regulation already asks for.
AI Act, Article 26
How it works
One API call per decision. No change to your model.
Reliax sits between your model and the action it triggers. It reads the model's output, checks it, and applies your rules. Your model is not changed, retrained or slowed down.
- 01
Your model decides
Your model runs as it does today. Reliax needs no access to its weights and no retraining.
- 02
Reliax checks
Before any action, each output is checked: has the population moved, is the input like the data the guarantee was built on, and can it be acted on under your guarantee?
- 03
Your rules decide
Outputs Reliax can stand behind follow your own rules, such as a lender's cut-off. A case goes to a person only when the evidence is not enough to decide it, when its input falls outside the guarantee or when the data has shifted.
- 04
An evidence record
Each decision gets a tamper-evident record of what was guaranteed and why it was acted on. Your validator and your supervisor can replay it at any time.
- Any model including black boxes
- No retraining uses your past decisions and their outcomes
- Two modes record only, or hold each decision for its check
- Your infrastructure cloud, private cloud or on-premises
The product
Every decision gets an evidence record a validator can replay.
One API call per decision. Each output comes back with what is guaranteed about it, in wording a risk officer, a validator and a regulator can read.
Evidence record· exampleALLOW
- Decision
- Loan application #2931
- Cut-off
- Below · bracket 2.1%–3.4% against your cut-off of 10.7%; automatic approvals proven to default below it
- Reliability
- Holds · answer right at least 95% of the time, checked for this applicant's group, on 7,500 past decisions
- Input
- Covered · similar to the data the guarantee was built on
- Data
- Stable · rechecked against outcomes on 31 Aug
- Action
- Allow · under policy
credit-v4 - Reasons
- All checks passed (rule 6) · your model’s reason codes attached
Record
rxe_7f3a9b · chained to the previous recordsignedTwo guarantees
Reliability of the inference: answers contain the real outcome at the rate you set. Your threshold: automatic decisions are proven to be on the right side of it. Both are proven on your past decisions with known outcomes and rechecked as outcomes land.
A scope check
The guarantee only covers inputs similar to the data it was built on. Anything else is flagged before any action, and a person decides.
A drift check
When the population moves, the guarantee is suspended for the affected group and the evidence record says so. Other decisions keep flowing.
A route
Act automatically where the guarantee holds, review where the evidence is not enough, block where the guarantee does not apply.
The reasons
Which rule fired and why, plus your model’s own reason codes. Reasons explain the route; they do not decide it.
A signed record
Any later edit is detectable. You own the records and can replay them years later. The verifier is open source, so records can be checked without us.
Evidence
Tested on real lending decisions.
Four public credit datasets, US mortgage decisions and 92 weeks of real lending data. Two of the datasets were used only once, for a test whose rules we wrote in advance. Anyone can check every number with our public code, failures included. We have no customer results yet; testing on lenders' own past decisions is the next step.
82–100%
of data errors caught. The model's own confidence caught 3–5%.We multiplied money amounts by 100, like a currency bug, on two datasets we had never used before. We wrote the test rules before running it. Part of that test failed, and the whitepaper says which part.
2.7% → 2.4%
default rate among automatic approvals, lender alone and with Reliax.On Give Me Some Credit. It also fell on the two other datasets (4.6% to 3.7%, 7.0% to 5.6%), because the doubtful cases went to a person. Early results.
0 of 300
test runs where automatic approvals were riskier than the cut-off allows, or automatic refusals safer.We repeated the test 100 times on each dataset, with new data each time. The guarantee was never on the wrong side. Early results.
23% → 94%
of the lost reliability won back after the data changed.Real data changed over time and the answers became less reliable. Updating Reliax without outcomes won back a quarter of the loss; updating it with four weeks of real outcomes won back 94%. The model was not changed.
92.4%
of answers right for Black applicants, against a 95% target.Checked over everyone at once, the guarantee held. Checked group by group, Black applicants got less reliable answers. Reliax's group check shows such gaps. Race is used only for the check, never in the decision.
5 of 5
test runs raised an alarm within 5 to 7 decisions when the data changed.This worked when the change showed in the application data. When it did not (Home Credit 2018), the alarm came late or not at all; real outcomes catch that case later.
Works with your existing tools
Monitoring reports on the model. Reliax checks each decision.
“How did the model do last month?”
Reports on batches after the fact. By the time an alert fires, the decisions have been made.
“Why did the model say this?”
Reason codes are useful and often required. But an explanation does not tell you whether the answer is right.
“Can I act on this output, and can I prove why later?”
Answered for each decision, before the action, with evidence a validator can replay. Independent of the engine that decides: in credit it plugs into FICO, Experian, Taktile or your own, and sends evidence to your model-risk tools.
Industries
Credit first. Insurance and healthcare next.
The guarantee works the same for any model that scores an important decision, and the platform runs on any of them today. What changes by sector is your policy, where outcomes come from, and the rules you must follow. Pilots start where the rules land first.
Credit and financial risk
Scorecards, default and fraud models. All our tests so far are on credit data, and the first rules apply here, from November 2026.
Insurance
Pricing, underwriting and claims. Life and health insurance pricing becomes high-risk under the EU AI Act in December 2027.
Healthcare
Triage, referral and diagnostic support. Clinicians need to know when a model's output falls outside the data it was validated on.
Talk to us
See your model the way your supervisor will.
A readiness review replays your past decisions through Reliax, inside your infrastructure, against success criteria agreed in advance, and gives you an evidence pack structured by article of the AI Act. If you work in model risk, validation, compliance or credit decisioning, contact us.
contact@reliax.io- 01
A readiness review
We replay your past decisions offline, inside your systems: how reliable the answers were for each segment, how your data changed over time, whether every group was treated alike, and what is missing from your logs. Your data never leaves.
- 02
Agree what success means
Success criteria are fixed before we start, so the result is judged fairly.
- 03
Review the results
What was guaranteed, which cases should have gone to a person, and the full record.