`pacta` is a local CLI prototype for interpreting formal-verification evidence in cryptographic tooling. It is not a trading bot, does not move funds, and does not make financial decisions.
The immediate corpus is the `saymrwulf/*-verified` family of repositories. The shipped Lean files are treated as the verification artifact. This project intentionally does not run Charon, Aeneas, extraction, or Rust-to-Lean regeneration.
## Purpose
An autonomous economic agent needs to answer a narrow question before trusting infrastructure:
> Does this theorem cover the exact code path that will protect my funds?
`pacta` helps answer that by replaying pure Lean checks where possible, auditing axioms and proof hygiene, generating machine-readable claim cards, and assigning residual-risk classifications with explicit exclusions.
The prototype is written for Python 3.11+ and macOS on Apple Silicon. It does not assume GNU coreutils, Linux `free`, Linux `taskset`, GNU `timeout`, Docker, Nix, or x86_64.
Lean tooling is detected with `shutil.which("lean")` and `shutil.which("lake")`. If neither is available, `pacta` reports clear diagnostics and still supports offline claim-card generation and static hygiene scans.
-`06a_provider_build_the_log.ipynb` / `06b_agent_verify_inclusion.ipynb`: the MIRRORED PAIR - one provider (builds, Lean-verifies, signs with the merkleized library, self-checks its own inclusion), many agents (verify inclusion from scratch in ~25 lines, no Lean); the domain separation is the design and the lecture structure mirrors it.
-`10_verified_custody_wallet.ipynb`: warden - the quorum custody boundary and signing firewall, ratchet-rule (toy 3-of-3, then the real four proven forks), plus the counterparty recomputing a custody card's inclusion proof.
The course states and keeps a "ratchet rule": every load-bearing idea runs twice - napkin scale, then real scale - and every pair is executable in the notebook.
The notebooks are committed without execution output. They can be opened in Jupyter, VS Code, or any notebook reader. They import `pacta` directly from this repository and avoid external notebook-only dependencies.
`pacta agent` turns evaluation into an operational consequence.
-`build-library` requires `R3` by default. It builds a small Rust proof-gated component capsule under `artifacts/`. The capsule embeds the claim card and exposes whether downstream automation may use the component for lower-layer cryptographic code only.
-`build-wallet-demo` requires `R4`. An `R3` Ed25519 arithmetic claim will refuse this action and write a machine-readable denial artifact instead of building a wallet.
This is intentional. Arithmetic proof evidence can authorize a constrained lower-layer library decision, but it must not contaminate wallet, transaction, custody, or trading-agent risk scoring.
In live mode, `--clone --run-axioms` downloads the configured repository, replays the local Lean checks, runs the axiom audit, writes `claims.yaml` and `report.md`, and only builds the capsule if the resulting score satisfies the policy threshold. Failed replay is a hard consequence: no artifact is built.
Some verified repositories rely on a pinned Aeneas Lean project, usually exposed by an environment script such as `~/aeneas-toolchain/env.sh`. `pacta` can use that environment without running extraction:
```bash
pacta doctor --config examples/repos.yaml --repo-name dalek-ed25519-verified
If those are missing, the result is `R0` for local replay because this machine lacks verifier capability. That is different from saying the theorem is false. It means the agent cannot trust the repository from local machine-checked evidence yet.
## Third-Party Attestation
For agents that should not build the full Lean/Aeneas environment locally, `pacta` also supports an attestation lane. A specialized proof-checking service can replay the proofs in its own controlled environment and publish a certificate describing:
- repository URL and commit,
- theorem/certificate names,
- observed axioms,
- Lean/toolchain environment,
- service identity and signature metadata.
The agent can consume that certificate only when the provider is explicitly trusted:
This changes the trusted base. The agent is no longer trusting local Lean replay; it is trusting the proof-checking service, its environment, signing key custody, and log retention. Without an explicitly trusted provider, attestation evidence scores `R0`.
The included `examples/dalek-ed25519.attestation.yaml` is an unsigned schema/demo fixture and requires `--allow-unsigned-attestation`. Real provider certificates should be signed and consumed with `--attestation-public-key`.
Standalone signatures prove who signed an attestation, but they do not make the provider accountable for equivocation or silent replacement. The nested provider can also append attestations to a local RFC 9162-style Merkle transparency log and issue inclusion receipts.
The log uses:
-`RFC9162_SHA256` Merkle leaf/node hashing with `0x00` leaf and `0x01` node domain separation.
- Signed Tree Heads over canonical JSON tree-head payloads.
- OpenSSL Ed25519 signatures today.
- An explicit `ML-DSA-65` / FIPS 204 signature slot that is `unavailable` unless the host has a real backend. If an agent policy requires both signatures, verification fails closed.
On a host without ML-DSA support, that command should fail. That is intentional. The system records the missing capability as a deployment blocker instead of treating the Ed25519 signature as quantum-robust.
This repository includes a nested provider prototype under `provider/`. It searches read-only under your home/GitClone tree for reusable Lean/Aeneas infrastructure, runs the proof replay, signs the result with OpenSSL Ed25519, and emits an attestation.
This is the intended trust transformation: local agents can avoid constructing the full verifier environment, but they must explicitly trust the provider identity and verification key.
Unknown log: pin on first use (recorded as trust-on-first-use). Same tree size: the root must match the pin - a mismatch is named EQUIVOCATION and rejected forever. Larger tree: a consistency proof from the pinned size is required (receipts embed a from-previous anchor whose root is itself checked against the pin). Smaller tree: rollback, rejected. `--max-sth-age-seconds` additionally rejects stale or future-dated tree heads.
## Dogfood Verified Cryptography
`pacta` can verify its own evidence signatures through the PROVEN code path - a small Rust binary built from the pinned proven source workspace (the exact commit the dalek certificates pin, serial backend pinned as the verified extraction pins it):
pacta receipt-verify ... --require-verified-verifier # fail closed unless the proven path ran
```
The backend that actually verified each signature (`verified-dalek-serial` or the `openssl` fallback) is recorded in receipts' signature statuses and attestation evidence; the fallback is never silent. A provenance sidecar records the source commit, backend cfg, and the honest coverage note (the certificates cover the extraction image of this verify path; SHA-512 and the wire glue remain the documented trusted base). ML-DSA is deliberately NOT dogfooded: no proven implementation exists, so the slot stays fail-closed - the hybrid-PQC posture is one proven-classical signature plus one required-but-honest post-quantum slot.
Every ed25519 repository in the corpus carries Lean-checked certificates for field arithmetic over `F_p` (`p = 2^255 - 19`), the complete twisted Edwards laws, full scalar arithmetic mod l, encoding/decoding canonicality with constructive decompression, and a FOUR-TIER signature apex culminating in the full lift: the extracted verifier accepts iff the signature's R decompresses to a valid on-curve point equal to `[k](-A) + [s]B`. Each apex tier's axiom cone is pinned to EXACTLY the fork's documented SHA-512/wire-format boundary by that repository's own check script (`pacta` mirrors those boundary sets per fork and re-checks observed cones against them - the authority is always the repo's button).
-`R2`: Formal model exists, but it is incomplete, weakly tied to production code, or major proof gaps remain.
-`R3`: A specific lower-layer implementation artifact is Lean-checked for a specific backend and theorem boundary.
-`R4`: End-to-end primitive proof covers public API, parsing/encoding, scalar arithmetic, hashing interface, signature equation, rejection rules, and implementation boundary.
-`R5`: `R4` plus reproducible production builds, compiler/build assurance, side-channel analysis, hardware/KMS/MPC integration, and operational controls.
Expected classifications: the arithmetic pair alone reaches `R3`. The full configured set - arithmetic, scalars, encoding/decoding, and the four apex tiers with boundary-exact cones - reaches `R4`, always with explicit residual blockers (the SHA-512 oracle, hypothesis-parametric parses, translation faithfulness, missing side-channel/build assurance). `R5` remains future work. The `build-wallet-demo` gate therefore genuinely opens on real evidence now - and still emits only a policy scaffold, never a production wallet.
