EVIDENCE

Don't trust us. Check the evidence.

Verification artifacts, test telemetry, and audit trail demos from actual test runs. Every number is reproducible. Every repository is public.

VERIFICATION CERTIFICATE

What a proof artifact contains

Every verified decision produces a certificate: a machine-readable record that ties the model, inputs, outputs, and integrity chain together. This is what one looks like.

verification_certificate.log Verified
timestamp 2026-04-13T09:41:22Z
model llama-3.1-8b-instruct
model_hash sha256:9f3a7b2c...d4e1c7e1
input_hash sha256:a1b2c3d4...8f9e0a1b
output_hash sha256:e5f6a7b8...2c3d4e5f
pipeline RAG+CSWR > CQL policy > verified
tests 3,000+ passed across 6 suites
hallucination 0% (controlled)
latency ~10s on RTX 5070 (local)
chain PASS | integrity intact
prev_hash sha256:7c8d9e0f...1a2b3c4d
15/20 blocks verified
FLIGHT RECORDER

Audit trail: contradiction detection in action

A real scenario from the 14-phase Contradiction Detection pipeline. The system receives a query, retrieves conflicting sources, detects the contradiction, and produces a resolution with confidence scores.

abes_flight_recorder.log Recorded
Phase 01 PASS Query ingestion: "What is the recommended dosage for treatment X in patients over 65?"
Phase 02 PASS Context retrieval: 4 documents retrieved from knowledge base. Relevance scores: 0.94, 0.91, 0.87, 0.72.
Phase 03 PASS Belief extraction: 3 belief statements extracted from retrieved context.
Phase 04 WARN Entropy scoring: belief entropy = 0.73. Threshold exceeded (0.60). Sources disagree on dosage range.
Phase 05 FLAG Contradiction detected: Source A recommends 10mg, Source B recommends 5mg for patients over 65. Publication dates differ by 3 years.
Phase 06 PASS Source recency weighting applied. Source B (2025) prioritized over Source A (2022).
Phase 07 PASS Confidence calibration: adjusted confidence = 0.78. Contradiction noted in output.
Phase 08 PASS Response generation with explicit uncertainty disclosure. Both sources cited with dates.
Phase 09 PASS Output hash generated: sha256:4f7a...9c2d. Chained to previous decision record.
Phase 10 PASS Provenance record written. Chain integrity verified. Certificate issued.
CONFIDENCE DECAY

Verification expires. That is the point.

A verification certificate is not forever. Confidence decays deterministically over time. Models update, data shifts, and the world changes. This tool shows how confidence decreases using exponential decay with a 72-hour half-life.

confidence_decay_calculator INTERACTIVE
Hours since verification 0h
95.0%
Confidence
VALID
Status
72h
Half-life
73h
Reverify in
TEST SUITE SUMMARY

Aggregate metrics across all projects

Four verification systems. 3,962+ total tests. Every result from actual test runs on local hardware, not projections or estimates.

IRONROOT
Provenance Layer
Tests 140
Coverage 51%+
Capability layers 5
AASMS
Adversarial Verification
Evolution cycles 6
Proposals verified 24
Errors 0
Pass requirement 100%
ABES
Contradiction Detection
Tests 822
Prompts evaluated 1,000
Pipeline phases 14
Hallucination rate 0%
RLFUSION
Decision Auditing
Test suites 6
Test iterations 3,000+
Pass rate 100%
Latency ~10s
Aggregate
3,962+ Total tests
0% Hallucination rate
100% Required pass rate
RTX 5070 Inference hardware
VERIFY FOR YOURSELF

Every repository is public

Source code, test suites, and results. Clone the repo, run the tests, check the hashes. That is what verifiable means.

Provenance Layer IRONROOT
Adversarial Verification AASMS
Contradiction Detection ABES
Decision Auditing RLFUSION

If you need AI you can prove works, let's talk.

Contact Us