metagenesis-core-public: Instructions file for Claude Code

MetaGenesis Core — Context for AI Agents (GSD)

START HERE: python scripts/session_close.py --read → shows full current state in 5 lines

Loaded automatically by all GSD agents via CLAUDE.md. Last updated: 2026-04-12 | v1.0.0-rc1 LIVE | 20 templates | 2428 tests

---

MISSION — WHY THIS EXISTS

One sentence: MetaGenesis Core is a notary for computations.

Like a notary certifies a document — we certify that a computer produced exactly this result, at exactly this time, in exactly this way. Without access to the computer. Without trusting anyone. In 60 seconds.

python scripts/mg.py verify --pack bundle.zip  →  PASS / FAIL

The problem: Every day billions of computations produce numbers that everyone must simply trust. ML model says 94% accuracy — trust it. FEM simulation says the part holds 10 tons — trust it. Drug calculation passed — trust it. Risk model says all clear — trust it. There is no standard of proof. MetaGenesis Core is that standard.

Six domains. Six real pains. One protocol.

Domain	Pain	What we give
ML / AI	Benchmark gaming. "94% accuracy" unverifiable	Cryptographic proof + timestamp. Impossible to backdate.
Pharma / Biotech	FDA 2025 requires verifiable AI artifacts for IND filing	Bundle = audit artifact. $299 vs $47M raise.
Finance	Basel III: independent VaR validation required	Regulator verifies offline. No model access needed.
Digital Twin	Calibration chain from physics to simulation — who proves accuracy?	Cryptographic chain from physical constant to result.
Science	70% of results cannot be reproduced	Reproducibility as math, not words. Hash equality = proof.
Climate / ESG	Carbon credits backed by unverifiable models	Proof the model produced exactly this result on this data.

The physical anchor — what makes us unique: We anchor computation to declared physical constants.

kB = 1.380649e-23 J/K  (SI 2019, exact, zero uncertainty) → PHYS-01
NA = 6.02214076e23 mol-1 (SI 2019, exact, zero uncertainty) → PHYS-02
E  = 70 GPa aluminum   (NIST measured, ~1% uncertainty)   → MTR-1
    ↓ anchor_hash baked in
DT-FEM-01: FEM output vs physical reference → rel_err ≤ 2% → PASS
    ↓ anchor_hash baked in
DRIFT-01: deviation from anchor ≤ 5% → PASS

Change any link in the chain — the cryptographic hash breaks everywhere downstream.