Verifiable evidence for computational claims — proof, not trust.
Implements the MetaGenesis Verification Protocol (MVP) v1.0.

Any computational result — ML accuracy, calibration output, FEM simulation,
ADMET prediction, VaR estimate, IoT sensor stream — can be packaged into a
self-contained tamper-evident evidence bundle that anyone verifies offline with
one command, getting PASS or FAIL with a specific reason. No zero-knowledge
proving overhead, no distributed-ledger infrastructure — just SHA-256 on any
machine. Works offline. No network. No model access. No trust required.

The mechanism is a notary for computations: it certifies what was computed,
when, and how — and, for in-scope claims, agreement with a known physical anchor
(SCOPE_001). That mechanism is secondary to the outcome above.

NOT a simulation platform. NOT an AI system. NOT a modeling tool.
A governance + evidence + verification protocol. An open standard.

What this PROVES: tamper-evidence; independent offline verifiability;
timestamped provenance (NIST Beacon temporal layer); physical-anchor agreement
for the in-scope claims (SCOPE_001).
What this does NOT prove: that a model's reported number is "correct"
(FAULT_006); detection of data contamination; it does not create or improve any
model or simulation; it does not replace physical validation or measurement.

## What this repo does

1. Wraps any computational result in a tamper-evident evidence bundle
2. Verifies bundles offline with 5 independent layers:
   - Layer 1: SHA-256 integrity (file modification)
   - Layer 2: Semantic (evidence stripped, hashes recomputed — still caught)
   - Layer 3: Step Chain (execution inputs changed — trace_root_hash breaks)
   - Layer 4: Bundle Signing (HMAC-SHA256 + Ed25519 — unauthorized creator)
   - Layer 5: Temporal Commitment (NIST Beacon — proves WHEN signed)
3. Enforces governance: every claim has implementation, every implementation has claim
4. Detects drift against verified calibration anchors
5. Cross-Claim Chain: MTR-1 → DT-FEM-01 → DRIFT-01 cryptographically linked end-to-end

## 20 domain verification templates

MTR-1        mtr1_youngs_modulus_calibration          (materials — physical anchor E=70GPa)
MTR-2        mtr2_thermal_paste_conductivity_calibration (materials — physical anchor)
MTR-3        mtr3_thermal_multilayer_contact_calibration (materials — physical anchor)
SYSID-01     sysid1_arx_calibration                   (system identification)
DATA-PIPE-01 datapipe1_quality_certificate             (data pipelines)
DRIFT-01     drift_calibration_monitor                 (drift detection — MTR-1 anchor)
ML_BENCH-01  mlbench1_accuracy_certificate             (ML classification + Step Chain)
DT-FEM-01    dtfem1_displacement_verification          (digital twin / FEM — MTR-1 anchor)
ML_BENCH-02  mlbench2_regression_certificate           (ML regression: RMSE, MAE, R²)
ML_BENCH-03  mlbench3_timeseries_certificate           (ML time-series: MAPE)
PHARMA-01    pharma1_admet_certificate                 (pharma ADMET — FDA 21 CFR Part 11)
FINRISK-01   finrisk1_var_certificate                  (finance VaR — Basel III/IV)
DT-SENSOR-01 dtsensor1_iot_certificate                 (IoT sensor data integrity)
DT-CALIB-LOOP-01 dtcalib1_convergence_certificate      (calibration convergence)
AGENT-DRIFT-01   agent_drift_monitor                    (agent quality drift — recursive self-verification)
MTR-4        mtr4_titanium_modulus_calibration        (materials — physical anchor E=114GPa Ti-6Al-4V)
MTR-5        mtr5_steel_modulus_calibration           (materials — physical anchor E=193GPa SS316L)
MTR-6        mtr6_copper_conductivity_calibration     (materials — physical anchor k=401 W/(m·K) Cu)
PHYS-01      phys01_boltzmann                         (fundamental physics — SI 2019 exact kB=1.380649e-23 J/K)
PHYS-02      phys02_avogadro                          (fundamental physics — SI 2019 exact NA=6.022e23 mol-1)

## Physical Anchor Principle (SCOPE_001)

Two distinct concepts — never conflate:
- Tamper-evident provenance: "was the bundle modified?" → ALL 20 claims
- Physical anchor: "does the number agree with physical reality?" → MTR-1/2/3/4/5/6, PHYS-01/02, DT-FEM-01, DRIFT-01, DT-CALIB-LOOP-01 ONLY

E = 70 GPa for aluminium is a physical constant measured in thousands of labs worldwide.
NOT a chosen threshold. MTR-1 verifies rel_err ≤ 1% vs this constant.
DT-FEM-01 verifies FEM output vs MTR-1 anchor.
DRIFT-01 monitors deviation ≤ 5%.
DT-CALIB-LOOP-01 verifies convergence toward that anchor.
Scope boundary documented in reports/known_faults.yaml :: SCOPE_001.

## Cross-Claim Cryptographic Chain

```
E = 70 GPa (physical reality)
    ↓
MTR-1  trace_root_hash: "abc..."  (calibration against physical constant)
    ↓  anchor_hash="abc..." baked into DT-FEM-01 Step 1
DT-FEM-01  trace_root_hash: "def..."  (FEM output vs physical reference)
    ↓  anchor_hash="def..." baked into DRIFT-01 Step 1
DRIFT-01   trace_root_hash: "ghi..."  (drift vs anchor)
```

DRIFT-01's trace_root_hash commits to entire chain from physical reality to simulation state.
Proven: tests/steward/test_cross_claim_chain.py

## Governance invariant

runner dispatch kinds == scientific_claim_index job_kinds == canonical_state claims
Enforced by: python scripts/steward_audit.py → must output STEWARD AUDIT: PASS

## Hard rules (never violate)

- Never use "tamper-proof" — use "tamper-evident"
- Never invent implementations — if file not in repo, say so
- Never add a claim without: implementation + test + claim_index entry + canonical_state update
- steward_audit.py must PASS after every change
- known_faults.yaml is authoritative — do not contradict it

## 10 innovations — 4 filed in PPA #63/996,819 + 4 post-filing implemented + 2 architecturally designed

PPA filed 2026-03-05 (§0020-0033 — four independent claims):
1. Bidirectional Claim Coverage (PPA §0020-0023) → steward_audit.py :: _claim_coverage_bidirectional()
2. Tamper-Evident Bundle with Semantic Verification (PPA §0024-0027) → mg.py :: _verify_semantic()  [test_cert02]
3. Policy-Gate Immutable Evidence Anchors (PPA §0028-0030) → mg_policy_gate_policy.json + .github/workflows/mg_policy_gate.yml
4. Dual-Mode Canary Pipeline (PPA §0031-0033) → runner.py :: run_job(canary_mode=True/False)

Post-filing implemented (to include in non-provisional by 2027-03-05 — not in filed PPA text):
5. Step Chain + Cross-Claim Cryptographic Chain → all 20 claims + anchor_hash  [test_cert03 + test_cross_claim_chain]
6. Bundle Signing (HMAC-SHA256 + Ed25519) → mg_sign.py + mg_ed25519.py  [test_cert07 + test_cert09]
7. Temporal Commitment (NIST Randomness Beacon) → mg_temporal.py  [test_temporal + test_cert10]
8. 5-Layer Independence Proof → test_cert11 (coordinated multi-vector) + test_cert12 (encoding attacks)  [proves each layer catches attacks others miss]

Architecturally designed (committed in repo before non-provisional filing — implementation in progress):
9. Self-Verifying Recursive Learning Layer (AGENT-LEARNING-01, v3.2.0 target) → docs/INNOVATION_09_SELF_VERIFYING_LEARNING.md  [the protocol's own learning record verified by the same 5-layer chain it applies to external claims]
10. Federated Verification Registry (FVR, v4.0.0 vision) → docs/INNOVATION_10_FEDERATED_REGISTRY.md  [multiple independent instances aggregate signed observations without trust in any central party; no consensus, no token]

Non-provisional filing will consolidate all 10 under a single unified claim family.

## How to verify everything works

python scripts/steward_audit.py                              # PASS
python -m pytest tests/ -q                                   # 2428 passed (2422 + 1 skip on win32)
python demos/open_data_demo_01/run_demo.py                   # PASS PASS

## Current state (2026-04-12)

- Claims: 20
- Tests: 2428 passing (2428 on Linux/CI; 2422 on win32 — 1 platform skip)
- Verification layers: 5 (integrity + semantic + step chain + bundle signing + temporal commitment)
- Domains: 8 (materials, sysid, data, ml, digital_twin, pharma_biotech, financial_risk, fundamental_physics)
- Agent checks: 22 (agent_evolution.py)
- Real ratio: 51.2% (21 real / 20 synthetic)
- Cross-Claim Chain: MTR-1 → DT-FEM-01 → DRIFT-01 cryptographically linked via anchor_hash
- steward_audit: PASS
- known_faults: 9 entries (ENV_001 + SCOPE_001 + ENV_002 + FAULT_007/008/009)
- DOI: 10.5281/zenodo.19521091 (Zenodo)
- SDK: sdk/metagenesis.py — MetaGenesisClient (verify, pack, sign, chain, receipt)
- GitHub Action: .github/actions/verify-bundle/ — one-line CI integration
- Site: metagenesis-core.dev — deployed on Vercel, auto-deploys from main
- PPA filed: 2026-03-05 — non-provisional deadline 2027-03-05
- Payment: $299/bundle — yehor@metagenesis-core.dev
- GitHub Release: v1.0.0-rc1
- Agent memory: session-level record rebuilt 2026-04-28 after a working-copy data-loss event (full audit: reports/AGENT_LEARN_FULL_AUDIT_2026_04_28.md; deeper validation: reports/AGENT_LEARN_DEEPER_VALIDATION_2026_04_28.md; event registered in EVOLUTION_LOG.md). Decision history of the 30-day pre-loss window preserved in commits, tags, weekly reports, and 47 quick-task slugs.
- JOSS paper: paper.md ready (resubmit Sep 2026)

## Key files

- README.md — problem, solution, quickstart
- docs/PROTOCOL.md — v1.0 protocol specification (5 layers + Cross-Claim Chain)
- docs/ARCHITECTURE.md — v1.0 architecture diagrams
- docs/VISION.md — why every computational result needs a tamper-evident artifact
- docs/PHILOSOPHICAL_FOUNDATION.md — epistemological foundation (proof, not trust)
- docs/EVOLUTIONARY_ARCHITECTURE.md — how the protocol verifies itself
- docs/REGULATORY_GAPS.md — FDA, EU AI Act, Basel III coverage assessment
- docs/ROADMAP.md — 4-level evolution: Protocol → Standard → Infrastructure → Universal
- reports/scientific_claim_index.md — all 20 claims with thresholds
- reports/canonical_state.md — authoritative claims list
- reports/known_faults.yaml — known limitations (SCOPE_001 + ENV_001)
- backend/progress/runner.py — job dispatch (20 claims)
- scripts/steward_audit.py — governance enforcement (SEALED)
- scripts/steward_dossier.py — generates per-claim dossier .md files from claim index
- scripts/steward_submission_pack.py — builds submission pack with evidence bundles
- scripts/mg.py — pack build + verify (3 layers) + verify-chain CLI
  verify-chain: python scripts/mg.py verify-chain bundle1/ bundle2/ bundle3/
- scripts/deep_verify.py — 13-test proof-not-trust verification script
- demos/open_data_demo_01/run_demo.py — working end-to-end demo
- demos/client_scenarios/ — 4 client scenarios: ML/AI, Pharma, Finance, Digital Twin

## Agent evolution system

Embedded agent governance with 22 evolution checks (`scripts/agent_evolution.py`)
and 8 daily CI agents enforcing documentation consistency, test coverage,
manifest accuracy, forbidden terms, and branch synchronization automatically.
Includes cumulative learning (`scripts/agent_learn.py`) that records recurring
patterns across sessions. AGENT-DRIFT-01 monitors agent quality drift through
the same verification mechanism the agents extend.

## Git workflow

Git repo: C:\Users\999ye\Downloads\metagenesis-core-public
Push via PR only — main branch is protected.
Policy gate: canonical_state.md LOCKED (normal state).

git checkout -b feat/your-branch-name
git add <files>
git commit -m "type: description"
git push origin feat/your-branch-name
# then open PR on GitHub → CI PASS → merge