Skip to content

Navigation Menu

Sign in
Sign up

Repository files navigation

MathLedger

Governance substrate for verifiable learning: deterministic replay, dual attestation, fail-closed claims.

Live Demo arXiv License: Apache-2.0


Pipeline

 MathLedger Demo Pipeline
 ========================
 ┌─────────┐ ┌─────────┐ ┌────────────┐ ┌───────────┐
 │ Input │ ──▶ │ Run │ ──▶ │ Artifacts │ ──▶ │ verify.py │
 │ (seed) │ │ (demo) │ │ (output/) │ │ │
 └─────────┘ └─────────┘ └────────────┘ └───────────┘
 │
 ▼
 ┌────────────┐
 │ PASS / FAIL│
 └────────────┘
 Determinism guarantee: same seed → byte-identical artifacts → same hash

Quick Start

# Prerequisites: Python 3.11+, uv package manager
uv run python scripts/run_dropin_demo.py --seed 42 --output demo_output/
cd demo_output && python verify.py

Expected output:

[PASS] Composite root verified: H_t == SHA256(R_t || U_t)

No database or external services required. The demo runs fully offline.


Project Map

You want to... Start here
Run the demo (2 min) scripts/run_dropin_demo.py
Understand what you're looking at docs/HOW_THE_DEMO_EXPLAINS_ITSELF.md
Audit artifacts independently docs/pilot/AUDIT_WALKTHROUGH.md
See the sealed FOL specification docs/FOL_FIN_EQ_PHASE3_CLOSURE.md
Browse all documentation docs/README.md
Try the live hosted demo mathledger.ai/demo
Read the paper docs/PAPERS/mathledger_arxiv_preprint.pdf

What Gets Generated

After running the demo, demo_output/ contains:

demo_output/
├── verify.py # Self-contained verifier (run this)
├── evidence_pack.json # Complete audit artifact
├── manifest.json # Execution metadata
├── u_t.txt # UI Merkle root
├── r_t.txt # Reasoning Merkle root
├── h_t.txt # Composite root: SHA256(R_t || U_t)
└── replay_instructions.md # Third-party reproduction steps

The verify.py script recomputes all hashes locally with no external calls. If replay produces different hashes, the pack is invalid.


What This Is

  • Governance substrate: separation of exploration from authority at the data model level
  • Deterministic attestation: same inputs → same hashes, every time
  • Fail-closed claims: ABSTAINED is a first-class outcome, not a missing value
  • Auditable by design: evidence packs enable independent replay verification

What This Is NOT

  • A capability demonstration (the demo does very little on purpose)
  • A production system (governance substrate only)
  • A safety claim (we claim legibility, not alignment)
  • A convergence guarantee (learning dynamics are Phase II)

Example Output

Running the boundary demo at mathledger.ai/demo produces:

1. ADV (Advisory) "2 + 2 = 4" → ABSTAINED Excluded from authority stream
2. PA (Attested) "2 + 2 = 4" → ABSTAINED Authority-bearing but no validator
3. MV (Validated) "2 + 2 = 4" → VERIFIED Arithmetic validator confirmed
4. MV (False) "3 * 3 = 8" → REFUTED Arithmetic validator disproved
Conclusion: Same claim text, different trust class → different outcome.
 Same trust class, different truth → VERIFIED vs REFUTED.

This demonstrates authority routing, not proof generation.


FOL_FIN_EQ_v1 (First-Order Logic Verification)

Verify first-order logic formulas with equality over finite domains:

python -m scripts.run_fol_fin_eq_demo --domain z2 --output demo_z2
python demo_z2/verify.py

Expected output:

PASS: All certificates verified

Golden manifest SHA256: 096ee79e4e20c94fffbc2ec9964dde98f8058cba47a887031085e0800d6d2113

See docs/FOL_FIN_EQ_PHASE3_CLOSURE.md for the sealed specification.


Independent Audit (SHADOW-OBSERVE)

For external reviewers conducting artifact verification:

Document Purpose
docs/pilot/AUDIT_WALKTHROUGH.md Complete audit procedure
docs/pilot/PILOT_NON_CLAIMS.md Binding non-claims
docs/pilot/PILOT_EVALUATION_CHECKLIST.md 22-item PASS/FAIL checklist
docs/FOR_AUDITORS.md Cold-start auditor guide

This audit verifies artifact integrity and determinism only; it makes no correctness, safety, or compliance claims.


Paper / Citation

MathLedger: A Verifiable Learning Substrate with Ledger-Attested Feedback

@misc{mathledger2025,
 title={MathLedger: A Verifiable Learning Substrate with Ledger-Attested Feedback},
 author={MathLedger Contributors},
 year={2025},
 eprint={2601.00816},
 archivePrefix={arXiv},
 primaryClass={cs.AI},
 url={https://arxiv.org/abs/2601.00816}
}

Documentation Index

Full documentation index: docs/README.md

Category Key Documents
Demo HOW_THE_DEMO_EXPLAINS_ITSELF.md, V0_LOCK.md
Audit FOR_AUDITORS.md, pilot/AUDIT_WALKTHROUGH.md
Specs FOL_FIN_EQ_PHASE3_CLOSURE.md, invariants_status.md
Ops DEPLOY_BY_TAG_DOCTRINE.md, VERSION_NUMBER_DOCTRINE.md

License

Apache-2.0

About

Governance substrate for verifiable learning: deterministic replay, dual attestation, fail-closed claims.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

AltStyle によって変換されたページ (->オリジナル) /