The Decision Provenance Engine (DPE)¶
Prove what the model did, not what it is¶
The DPE is CRP's runtime answer to the AI black-box problem. It does not try to understand the model's weights. It breaks every response into claims, checks each claim against verifiable sources, and binds the verdict to a tamper-evident audit trail. The result: you can ship AI outputs with measurable evidence of grounding, safety, and quality.
The DPE runs 13 stages on every response before it is delivered. It is fully specified in SPEC-005.
This page explains each stage in plain English for non-technical readers.
| # | Stage | What it checks | Latency budget |
|---|---|---|---|
| 1 | Claim Detection | Decomposes the response into atomic, verifiable claims. | < 5 ms |
| 2 | Attribution Analysis | Maps each claim back to facts in the envelope. | < 3 ms |
| 3 | Fabrication Detection | Flags claims with no envelope support and no plausible inferential path. | < 4 ms |
| 4 | Distortion Detection | Detects altered numbers, dates, names, quantities. | < 3 ms |
| 5 | Entailment Scoring | Scores claim-to-source semantic entailment. | < 8 ms |
| 6 | Contradiction Detection | Cross-checks against the current envelope and prior windows. | < 4 ms |
| 7 | Repetition Detection | n-gram and semantic overlap against prior windows. | < 2 ms |
| 8 | Completeness Verification | Tracks how many required sections / outputs were produced. | < 2 ms |
| 9 | Flow Analysis | Opening coherence, topic continuity, register, transitional markers. | < 4 ms |
| 10 | Hallucination Risk Scoring | Aggregates 1-6 into a calibrated risk score. | < 2 ms |
| 11 | Quality Tiering | S / A / B / C / D classification. | < 1 ms |
| 12 | Safety Policy Evaluation | Applies the active CRP-Safety-Policy directives. | < 2 ms |
| 13 | Provenance Binding | HMAC-chains the verdict to the audit trail. | < 5 ms |
Total budget: < 50 ms for an average 1,500-token response.
Why 13 Stages Is the Right Number¶
Each stage answers a question that a regulator, auditor, or safety reviewer genuinely asks:
- "Did the model invent something?" → Stage 3
- "Did the model change a number or date?" → Stage 4
- "Does the answer logically follow from the source?" → Stage 5
- "Did it contradict an earlier turn?" → Stage 6
- "Did it just rephrase what it already said?" → Stage 7
- "Did it finish the job?" → Stage 8
- "Does the prose actually read as continuous?" → Stage 9
- "How risky is this response overall?" → Stage 10
- "Can I prove all of this happened?" → Stage 13
If any one stage is removed, a class of safety question goes unanswered.