Verification Protocol  ·  #63/996,819
7 active claims| 71 tests| PASS| patent pending
PROOF,
NOT
TRUST.
What this is

Any computational result — ML accuracy, physics simulation, pipeline output — packaged into a tamper-evident evidence bundle. Any third party verifies it offline with one command, without your code, data, or environment. PASS or FAIL. No grey zone. No trust required.

SCROLL
7Active Claims
71Tests Passing
2Verification Layers
5Domains
0Trust Required
PASSsteward_audit
MITLicense
7Active Claims
71Tests Passing
2Verification Layers
5Domains
0Trust Required
PASSsteward_audit
MITLicense
The Problem

AI makes claims.
Nobody can verify.

Every ML team claims state-of-the-art results. Every simulation produces numbers. Every pipeline outputs certificates. There is no standard way to check any of it independently.

Without MetaGenesis

Trust the number.
Or re-run everything.

A reviewer cannot check a reported ML accuracy without the original model, training data, and compute environment. The only options are blind trust or complete reproduction.

With MetaGenesis

One command.
PASS or FAIL.

Any computational result is packaged into a tamper-evident evidence bundle. A third party runs one command and gets an unambiguous verdict offline, without access to your environment. No grey zone.

How It Works

Four steps.
Immutable proof.

Every run produces an evidence bundle with two independent verification layers. Any deviation surfaces as FAIL with a specific reason.

01
runner.run_job()
Executes computation → produces run_artifact.json + ledger_snapshot.jsonl
→ artifact
02
evidence_index.json
Maps run artifacts to registered claims with provenance chain
→ index
03
mg.py pack build
Bundles artifacts + SHA-256 manifest + root_hash into submission pack
→ bundle
04
mg.py verify
Integrity check (SHA-256) + semantic check (job_snapshot, canary flag, kind)
PASS / FAIL
Live Demo

Five minutes
from zero to proof.

bash — metagenesis-core-public
$
cloning...
installing requirements...
→ running verification demo
✓ PASSbundle integrity verified
✓ PASSsemantic invariants verified
$ python -m pytest tests/steward tests/materials tests/ml -q
71 passedin 2.70s
Steward Audit
statusSTEWARD AUDIT: PASS
required filesall present
phase 42locked ✓
claim coveragebidirectional ✓
Canonical State
MTR-1,2,3materials
SYSID-01system id
DATA-PIPE-01pipelines
DRIFT-01drift
ML_BENCH-01NEW — ML accuracy
Why SHA-256 Is Not Enough

The bypass attack.
Caught. Proven.

Every system using only file hashes is vulnerable. An adversary removes content, recomputes all hashes, and passes integrity checks silently.

The attack
1
Remove job_snapshot from run_artifact.json — stripping the core evidence
2
Recompute all SHA-256 hashes to match modified files — restoring apparent integrity
3
Submit bundle with no real evidence that passes all standard integrity checks
✗ Standard check: PASS (attack succeeds silently)
MetaGenesis defense
1
Integrity layer — SHA-256 + root_hash detects any file modification after manifest generation.
2
Semantic layer runs independently — checks job_snapshot present, payload.kind matches, canary_mode correct.
3
Even with all hashes recomputed, semantic check fails: job_snapshot missing.
✓ Semantic check: FAIL — job_snapshot missing (caught)
Adversarial test: tests/steward/test_cert02_pack_includes_evidence_and_semantic_verify.py
7
Active Claims
Across 5 domains: materials, ML/AI, system ID, data pipelines, drift monitoring
71
Tests Passing
Including adversarial tamper detection, determinism checks, and boundary conditions
2
Verification Layers
SHA-256 integrity + semantic invariants. Each layer catches attacks the other misses.
0 req.
Trust Required
No GPU, no internet, no code access. Verify on any machine with Python.
Verified Claims

Seven claims.
All bidirectionally enforced.

Every claim has an implementation, runner dispatch, threshold, and tests. Enforced on every PR — not by human review.

MTR-1
Materials Science
Young’s Modulus Calibration
relative_error ≤ 0.01
MTR-2
Materials Science
Thermal Paste Conductivity Calibration
relative_error ≤ 0.02
MTR-3
Materials Science
Multilayer Thermal Contact Calibration
rel_err_k ≤ 0.03 · rel_err_r ≤ 0.05
SYSID-01
System Identification
ARX Model Calibration
rel_err_a ≤ 0.03 · rel_err_b ≤ 0.03
DATA-PIPE-01
Data Pipelines
Data Pipeline Quality Certificate
schema pass · range pass
DRIFT-01
Drift Monitoring
Calibration Anchor & Drift Monitor
drift_threshold 5.0%
ML_BENCH-01NEW
ML / AI Benchmarking
ML Model Accuracy Certificate
|actualclaimed| ≤ 0.02
Governance

Code enforces.
Not people.

Bidirectional claim coverage checked on every pull request. A claim without implementation blocks merge. An implementation without claim blocks merge.

$ python scripts/steward_audit.pySTEWARD AUDIT: PASS
canonical_state: ['MTR-1','MTR-2','MTR-3','SYSID-01','DATA-PIPE-01','DRIFT-01','ML_BENCH-01']
claim_index: ['MTR-1','MTR-2','MTR-3','DATA-PIPE-01','SYSID-01','DRIFT-01','ML_BENCH-01']
coverage check: all job_kinds dispatched — runner kinds in claim index
canonical sync: PASS — bidirectional coverage verified
Open Protocol

Not a tool.
A standard.

MetaGenesis Verification Protocol (MVP) v0.1 — open spec for packaging computational claims into independently verifiable evidence bundles.

What MVP defines

A minimal, concrete spec for what “independently verifiable” means for any computational result.

Bundle: pack_manifest + evidence_index + per-claim artifacts
Integrity: SHA-256 hashes + root_hash over all files
Semantic: job_snapshot present, kind matches, canary flag correct
Governance: runner kinds == claim_index kinds == canonical_state
Output: PASS or FAIL with specific reason — no ambiguity

What it is not

Not a simulation platform
Not an AI system
Not “tamper-proof” — tamper-evident under trusted verifier assumptions
Does not guarantee algorithm correctness — only evidence integrity

Planned domains

PHARMA-01 — ADMET certificates (FDA 21 CFR Part 11)
CARBON-01 — carbon sequestration model outputs
FINRISK-01 — VaR model validation (Basel III/IV)
Use Cases

Five verticals.
One protocol.

01
ML / AI

Benchmark Certification

Any ML accuracy claim packaged into a tamper-evident bundle. Reviewers verify offline — no model access, no environment, no GPU.

02
Pharma / Biotech

Regulatory Submission

ADMET predictions, PK/PD simulation outputs — packaged with full provenance. FDA 21 CFR Part 11 compatible audit trail.

03
Carbon Markets

ESG Model Auditing

Carbon sequestration and deforestation models become independently auditable. Corporate buyers verify without proprietary model access.

04
Financial Services

Risk Model Validation

VaR, credit scoring, stress test outputs packaged for Basel III/IV model risk management.

05
Materials / Engineering

Calibration Handoff

Young’s modulus, thermal conductivity — delivered with machine-verifiable proof. Drift detection against verified baselines included.

One command.
Proof.

Clone. Run the demo. See PASS.
No account. No API key. No GPU. No network.

View on GitHub → Request Free Pilot
MIT · Patent Pending #63/996,819 · Inventor: Yehor Bazhynov