Repository navigation
ProofServe 0.1.0 — final materials, not published
Review: 自审. Frozen native checker e000ed6,
Lean 4.34.1, ABI6 / independent-closures-v8. Measured archive runtime:
818f742. Later documentation commits are not
substituted as its build provenance. All finite campaigns ended; not every
planned point or requested module completed. No public release action occurred.
Product and CLI migration
check PROJECT checks requested compiled-root dependency closures with the
original kernel, parallel workers, local checkpoints and receipt v1. Source
must already be built; default roots do not cover every source file.
audit PROJECT reports static coverage, explicit computation syntax, optional
supplied costs and compiled evidence; unavailable information stays explicit.
serve keeps the Kimina-compatible behavior. The former candidate JSONL check
command is now batch; no service feature was added.
Bundle: original matching SDK/checker/audit, Linux x86_64, Python 3.11+.
cargo install provides the Rust launcher/service, not the complete SDK.
Different declared Lean versions use their matching official toolchain when
available or fail explicitly. No silent migration or download occurs.
Final native matrix
Median seconds, three completed repeats per point. Same native input ledger;
within-tool speedups, not speedups over official leanchecker.
Nat.Basic is a 1,330-module closure, not all Mathlib.
| Closure | 1-worker wall | 8-worker wall | 1→8 wall speedup | CPU, 1 / 8 |
|---|---|---|---|---|
| Nat.Basic (1,330) | 110.809 | 65.312 | 1.697× | 110.246 / 113.455 |
| Grid (3,793) | 753.992 | 352.349 | 2.140× | 751.162 / 835.761 |
| Barrier (4,449) | 735.591 | 425.906 | 1.727× | 732.639 / 766.344 |
Nat.Basic/Grid also completed 2/4-worker points; Barrier 2/4 were not planned.
Of 71 planned cells, 61 attempts returned:10 unattempted and two incomplete.
No partial sample fills a median. Final matrix, commit 818f742.
Strict Nat.Basic at 8 workers costs 276.053 wall /283.782 CPU s:
4.227× wall /2.501× CPU versus default 8. Strict Grid timed out124 after
1,800.010 wall s, with 1,803.564 cgroup CPU s. No strict Grid repeat completed;
the other two were not run. No completed-run Grid strict multiplier exists.
Stock default trusts imports; --fresh replays a final constant environment,
not the complete module-local proof ledger. Discovery also differs.
Their observations do not establish an equal-obligation official speedup.
Final installed-package external evaluation
All 3 projects and 9 native/official/audit arms completed in 52.42 minutes,
with fresh checkpoints, one worker and n=1 per arm. No interrupted preview
cost was reused.
| Project | Native closure / official roots | Native CPU / wall s | Official CPU / wall s |
|---|---|---|---|
| Newspeak | 754 /105 | 109.055 /110.639 | 38.637 /38.736 |
| Solar | 2,368 /12 | 167.209 /169.321 | 5.512 /5.539 |
| samuelburbulla/proofs | 10,753 /9 | 2,481.041 /2,579.748 | 37.377 /40.993 |
All 126 common local roots agree. No same-obligation external speedup or
full-closure official equivalence is established: native checks the closure;
official default trusts imported proofs.
Final data/pins, commit 0516279.
AI upstream says “AI generated proofs of open math problems”: a repository-wide
claim, with per-proof/model/human-editing fractions undisclosed, not an independent
authorship audit. Audit coverage: Newspeak 105/105, Solar 12/32 (20 outside the
static default graph), AI 9/9. Lexical proxies are not executions/cost estimates.
Source binding/minimal used-axiom closure are not established.
Timings used P0–7/shared lease; preparation E8–9, no Lean -M.
zettai-dev/agentsight-dev kept running; not an exclusively idle host.
The 12 GiB external policy remained; AI recorded 888 memory.max events, no OOM/kill.
Supplementary corpus check: incomplete certified prefix
Certified 968/162,170 requested built source modules in 15 complete
batches, with 12,756 closure modules including dependencies.
161,202 requests are uncertified. Another 271 complete-segment accepts and 217
in-flight accepts are provisional; one header has no verdict. No kernel rejection
was reported. Operational deadline exit124 keeps collection completion false.
Native cgroup cost: 16,194.749 CPU /16,530.349 wall s, including unfinished
work; neither completed-corpus cost nor extrapolation. Requests excluded 926
user-deferred unbuilt sources and one lakefile configuration unit. The prefix
is dependency-order-dependent, not representative.
Exact coverage,
commit 5f49168. A successful segment is not a completed batch receipt.
Trust and known differences
Trust includes original kernel, adapter/toolchain, reported axioms and protected
local checkpoints. Default shared-checked can accept a crafted proof depending
on another module's theorem body that official original-context replay rejects;
the proposition follows from checked dependencies and reported axioms.
Strict matches the supplied controls, not universal hostile-artifact parity.
Local certificates/serialized environments are editable trusted state, not
portable cryptographic proofs. Noncryptographic identities detect ordinary
staleness, not forged metadata/collisions. Inputs must stay stable.
Incomplete/operational results do not certify the request; Git/source digests
do not bind source to artifacts. No sandbox for arbitrary hostile binaries.
Open-parameter reduceBool/reduceNat mentions work; closed/higher-order native
evaluation may be refused. --official is not a guaranteed IR workaround.
Incomplete required module parts are refused. Axiom fields are declaration
inventories, including sorryAx/compiler/runtime names, not minimal used-axiom
closures or sorry-free guarantees. Boundary.
Exact qualified archive
- File: proofserve-0.1.0-linux-x86_64.tar.gz
- Size: 954,865,527 bytes
- SHA-256: 5c49603e3f8e34b72b89e7959b00a74fbcc910c534da90dcb342dae01381da61
- Embedded Lean 4.34.1 / checker e000ed6 / runtime818f7422; dirty=false.
- Native executable equals the qualified frozen binary; 169 qualified source
files match Git's e000ed6 snapshot. - Debian13/glibc2.41-qualified only. Older-glibc compatibility is unverified.
Ubuntu22.04 workflow output is a distinct artifact with separate provenance.
Archive path/provenance. SHA-256 is requested publication
provenance, not a new install/acceptance gate. Embedded documentation stays
at 818f742; these notes are its updated sidecar, not a mutated measured tarball.
Functional qualification and pending public decisions
Frozen archive/container preview passed relocation, check/audit, strict/default
forwarding, version refusal and all 3 existing kernel-rejection classes.
Cold image and standalone network-disabled tests included two Mathlib imports
and accepted/rejected/sorry responses. Evidence.
Image caches were reclaimed; retained evidence is not an available image.
Procfs ESRCH race fixed at 7a931ad: 102 local Rust tests, fmt/clippy, four crate
dry-runs; CI separately ran 6 native tests. Ordinary push CI skips its manual-only
container job; a green push alone does not qualify an image. Historical
passes/failures retain their pins in qualification JSONs.
Readiness: conditional on the user's public release decisions, including ABI
scope and coupled tag/crates effects. Release checklist.
Research data/history stay in Git. Nothing has been tagged or published.