Decision ReceiptModel DecisionDR-K7M9-2P4Q
Verified- Decision
- Approved by Risk Committee
- Policy
- Third-party risk — Tier 1, v7
- Evidence
- 3 attestations, 2 documents
- Integrity
sha256:0xdead…beef
Use casesAI decisions
AI decisions you cannot reproduce.
Model outputs are consumed as decisions every minute. When asked to explain or replay one later, most organisations cannot produce a verifiable Decision Receipt.
MeshQu changes that.
Signals
Input data
Feature snapshot
Model version
Policy wrap
Reviewer actor
Decision
Model decision at execution
Approve · Decline · Override
Proof
Decision Receipt
Verified
01The moment
A model produces a result.
An input is submitted. Features are generated. A model returns an output.
A decision is made: Approve, Decline, Score, or Classify.
Model output
Application · APP-77241
- Model
- underwriting · v3.14
- Output
- 0.78 · decline band
- Confidence
- Low margin Override eligible
ApproveDeclineOverride
02The failure
You are asked to explain it later.
Why did the model return this result? Which version was used? What thresholds applied? What inputs influenced the outcome?
You try to reconstruct it from model outputs, logs, feature pipelines, and configuration. It takes time. And even then — you can't reproduce the decision.
Reconstructing the inference…
- Pull model registry20 min
- Recover feature snapshot60+ min
- Find threshold config30 min
- Search override notes40 min
- Replay — best effortoften impossible
Cannot reproduce.
03The reason
It was never captured at execution time.
Model systems produce outputs. They do not capture the decision as a verifiable, replayable event. No Decision Receipt is produced.
The model runs. The result exists. But the exact decision — as it happened — is lost.
Signals received
Logs written
The gap
No replayable evidence captured
04The shift
Model decisions should be reproducible.
Not approximate. Not inferred. Not explained after the fact.
A decision should be captured in a way that can be replayed and verified exactly as it occurred.
Signals received
Logs written
Decision + Proof
At execution
05At execution
Captured as it happens.
Most AI systems log outputs after the model runs.
MeshQu captures the decision at execution time.
The inputs, policy, and outcome are bound in the moment. After the fact is interpretation. At execution is proof.
INFERE
MeshQu Decision Layer
Decision Receipt
Verified
06The receipt
Every decision produces a receipt.
A Decision Receipt contains the input data, the feature snapshot, the model version and configuration, the policy and threshold logic, the outcome, and the actor.
The model version and the policy version are pinned in the receipt — exactly the version that ran, and exactly the policy that was ratified at the moment of the call. Versions are not metadata. They are proof.
Signed. Verifiable. Replayable. The same decision can be run again — and produce the same result.
07A new layer
It works with what you already use.
Model pipelines. Feature stores. Inference services. Decisioning systems. MeshQu doesn't replace them.
Your systems run the model. MeshQu proves the outcome.
Pipelines
Features
Inference
Decisioning
MeshQu
— receipt
decisiondecline
modeluw-v3.14
actorsystem
sealverified ◎
08Consistency
Deterministic by design.
Same input. Same policy. Same version. Same result.
Every decision can be replayed and verified independently.
Same input
Identical features
Same policy
Identical thresholds
Same version
Identical weights
Same result. Every replay.
09The result
You can answer precisely.
The receipt resolves instantly — what inputs were used, which version ran, what logic applied, what outcome was produced. Reproducible. Verifiable.
Asked
Why did the model return this result?
RCP-01H7M12B7G6AI4Q VerifiedResolved in instantly
10Close
You ran the model.
The question is — can you run it again and get the same result?
Decision AssuranceCapture every decision. Prove every time.