Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 14 additions & 7 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,15 +2,15 @@

`mcproof` is the verification harness, Plane 1 rulebook, dataset suite, weight compiler, and serving benchmark for the Siliconcordat Formal Verification MDE. It checks Lean 4.35.0 proofs across six kernels, grades exact-math checkers against planted bugs, compiles `weft` adapters, and audits live `warp-modeld` mounts. Proof authors, verifiers, and researchers use it to test formal reasoning without weakening deterministic gates.

Blueprint and gates live in [siliconcordat/mde issue 4](https://github.com/siliconcordat/mde/issues/4) under the [Master Plan](https://github.com/siliconcordat/.github/blob/main/MASTER-PLAN.md). Companion serving changes live in [siliconcordat/exo](https://github.com/siliconcordat/exo/tree/feat/mde4-phase4-phase6-stack) and [siliconcordat/warp PR 5](https://github.com/siliconcordat/warp/pull/5).
Blueprint and gates live in [siliconcordat/mde issue 4](https://github.com/siliconcordat/mde/issues/4) under the [Master Plan](https://github.com/siliconcordat/.github/blob/main/MASTER-PLAN.md). Companion serving changes live in [siliconcordat/exo](https://github.com/siliconcordat/exo/tree/feat/mde4-phase4-phase6-stack) and [siliconcordat/warp](https://github.com/siliconcordat/warp/tree/feat/mde4-phase4-live-gate).

## Unified Repository Layout

| Path | Contents |
| --- | --- |
| `src/mcproof/`, `bin/mcproof`, & `bin/bootstrap-deps` | Python 3.11+ Tier 2 oracle runner (`oracles.py`, `lean_harness.py`, `mutation.py`, `lineage.py`, vendored `prose_gate.py`, `cli.py`) and pinned external checkout bootstrap |
| `docs/CONTINUITY.md` | Capability steward record (`mde:formal-verification`), three-tier contracts, scratch reproduction runbook, and empirical Phase 1 to 6 lessons |
| `lean/` | Stock-`Init` Lean `v4.35.0` reference library, challenge/solution modules, 190-case differential test suite, and 6-kernel `lake comparator --paranoid` config |
| `src/mcproof/`, `bin/mcproof`, & `bin/bootstrap-deps` | Python 3.11+ Tier 2 oracle runner, pluggable `DomainOracle` protocol, label-free `repair_loop.py`, `F-01..F-08` certificate auditor, vendored `prose_gate.py`, and CLI |
| `docs/CONTINUITY.md` & `docs/CROSS_POLLINATION.md` | Capability steward record (`mde:formal-verification`), bidirectional `mcproof`/`soko`/collaborator transfer ledger, and Phase 1 to 6 lessons |
| `lean/` | Stock-`Init` Lean `v4.35.0` reference library (`29` theorems across `Mcproof.lean`, `TransitionBounds.lean`, and `AdapterBounds.lean`), `370`-case differential suite, and `6`-kernel `lake comparator --paranoid` config |
| `spec/` & `schemas/` | Audited `32,216`-token (`Qwen3.5-4B`, `30,873` `cl100k_base`) Plane 1 specification (`plane1_formal_spec.md`), SHA-256 `lineage.json`, and canonical `zone-manifest.v0.1.json` |
| `datasets/` | Frozen public `compile.jsonl` (`96`), `validation.jsonl` (`48`), `hybrid_k0.jsonl` (`2,048` noise + `24` `y=0` anchors), `144` gold targets, `scripts/grade_item.py`, and `SHA256SUMS` (`312953bc...`) |
| `commitments/` | Public-safe cryptographic commitment (`PUBLIC_HOLDOUT_COMMITMENT.json`, zero prompt text) binding the `242`-item sealed holdout and canary vault |
Expand Down Expand Up @@ -42,7 +42,14 @@ On public `validation.jsonl` (`48` items), `P2-B` regressed below bare `B0` and
| Rollback fidelity | `0.0` next-token logprob diff after `O(1)` kill switch | `0.0` exact top-20 logprob diff (`5.528 ms` multi-slot flush) | **PASS** |
| Phase 6 `T-C1` solo preservation | `>= 90%` live behavioral preservation in paired/triple stacks | `0.0%` min live `step0` retention (`weft-eido` `0.0611` live drift vs `0.0019` solo shift, `99.85%` CPU subspace diagnostic) | **FAIL (Live `step0` Drift)** |
| Phase 6 `T-C3`, `T-C4`, `H1`-`H7` | Base retention, order invariance, and 7 live hazards | `85.0%` vs `84.0%` `W0` (`T-C3`), `0.0` perm/rollback diff (`T-C4`, `4.602 ms`), `7/7` hazards | **PASS** |
| Phase 6 `T-C5` serving budget | `<= 15%` overhead (`>= -5%` floor), `>= 8k` KV, `<= 7.2 GiB` NVML VRAM | `-33.1%` same-server latency (`benchmark_anomaly = true`), `8192` KV/slot, analytical `4.98 GiB` VRAM (no 8 GiB NVML GPU online) | **FAIL (Timing Anomaly & Analytical VRAM)** |
| Phase 6 `T-C5` serving budget | `<= 15%` overhead (`>= -5%` floor), `>= 8k` KV, `<= 7.2 GiB` NVML VRAM | Live `UX582HS` RTX 3080 (`-c 65536 -np 8`): `Q8_0` `6.8525 GiB`, `1`-slot `+0.38%`, `8`-slot `-0.66%` (`Q4_K_M` `5.2314 GiB`, `+0.47%` / `+0.48%`) | **PASS (v0.2 Live NVML)** |

### v0.2 Closed-Loop Plane 1 Retrieval and Tier 2 Repair ([Proven-Commons/mcproof issue 5](https://github.com/Proven-Commons/mcproof/issues/5))

Under `100%` label-free `default_inference_oracle` stopping (`<= 1,500`-token Plane 1 retrieval + AST/token state-hash + `seen_fail_sigs` stagnation escape, receipts in `compile/artifacts/evaluation/repair_loop/`):

- **Public `validation.jsonl` (`48` items)**: `B0` climbed from `4/48` (`8.33%` open-loop) to `5/48` (`10.42%`, `k=1`) to `6/48` (`12.50%`, `k=2`) to **`8/48` (`16.67%`, `k=4`, `2.0x` open-loop)** (`A: 1/12, B: 5/12, C: 0/12, D: 2/12`), while single-layer `blk.31` `P2-B` scored `0/48` (`67` stagnation events).
- **Sealed `private_holdout.jsonl` (`96` items)**: broke the `0/96` single-pass ceiling across both arms (`B0` `0/96 -> 1/96` [`HOLD-C-04` at `k=2..4`], `P2-B` `0/96 -> 1/96` [`HOLD-D-10` at `k=3..4`]).

## Prerequisites and Fresh-Clone Reproduction

Expand Down Expand Up @@ -87,11 +94,11 @@ Run the full Tier 2 oracle gates and Family `A` to `D` graders. On Linux, `lake

- Passing Tier 2 oracles proves only the machine-checked properties enforced by each checker. Whether a theorem statement matches human intent remains a Tier 1 human specification judgment.
- `datasets/scripts/grade_item.py` grades Family C items by deterministic structural claim-tier and `meta["bar_status"]` replay rather than re-running Lean compilation on prose-only reports.
- Live GPU runs executed on a local 64 GiB Apple M1 Max (`llama-server 0.5.0`, `$0` cloud). Analytical 8 GiB RTX 3080 VRAM projections (`4.9753 GiB` Linux headless / `6.1753 GiB` Windows WDDM on `Q8_0`) remain unmeasured on live 8 GiB NVML hardware (`vram_measurement_kind = analytical_projection_not_nvml`).
- Live `8` GiB RTX 3080 `T-C5` telemetry (`stack/results/final/ux582hs_rtx3080_tc5_live_receipt.json`) measured Windows 11 NVIDIA Optimus (`driver 566.07`, `0 MiB` idle WDDM VRAM), while bare-metal Linux kernel execution on `UX582HS` remains tracked separately.

## What you must trust

- Lean `v4.35.0` and its five bundled external kernels (`leanchecker-paranoid`, `lean4lean`, `nanoda`, `con-leche`, `con-ron`), plus the Lean compiler and C runtime for native binaries (`Classical.choice`, `Quot.sound`, and `propext` measured via `collectAxioms` on `Mcproof.Proofs`).
- Lean `v4.35.0` and its five bundled external kernels (`leanchecker-paranoid`, `lean4lean`, `nanoda`, `con-leche`, `con-ron`), plus the Lean compiler and C runtime for native binaries. Across the `29` reference theorems, `collectAxioms` and `#print axioms` on `Mcproof.Proofs` use only `{propext, Classical.choice, Quot.sound}`. Axiom-free: `bezoutCheck_sound`, `bezoutCheck_complete`, `ratLeCross_sound`, `scalar_projection_retention_illusion`. Also `Classical.choice`: `clampNat_in_range`, `cleanSegments_safe`, `cleanSegments_singleton`, `prefundedVal_bounded_drop_witness`. The other `21` use `propext` and/or `Quot.sound` only (no custom axioms).
- Python 3 exact `int` and `fractions.Fraction` arithmetic (`-I -O -B -W error`), `numpy` CPU `float64` linear algebra, and the Rust `tree-sitter` CST parser in `siliconcordat/mde`.
- Host OS sandboxing (`bwrap` network-unshared read-only root on Linux, or subprocess `-I` environment isolation without OS syscall confinement on macOS and Windows) and SHA-256 digest implementations.

Expand Down
Loading
Loading