Skip to content
Draft
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
2 changes: 1 addition & 1 deletion .github/lexlean-revision
Original file line number Diff line number Diff line change
@@ -1 +1 @@
15c667bd36604c94c28e39502edec3fcad38e0ce
eadeff7cfe5c8e127b26543da04bfc2352d89047
21 changes: 17 additions & 4 deletions AGENTS.md
Original file line number Diff line number Diff line change
Expand Up @@ -156,7 +156,14 @@ Its only events are host replies and cancellation. `ResearchRecord` captures the
exact inputs and report reads; `ResearchEvidence` retains both the library commit
and canonical proof record under verified Git references before workspace removal.
Only then can the session return a checked answer. A stored record itself confers
no proof authority. Failed archival or cleanup retains the workspace. Stopped or
no proof authority.
`ResearchAutonomy` composes owned inspection, bounded portfolio dispatch, recovery
and batch claim publication. `lexlean inspect` is supplied by the v0.5 development
fork in upstream PR #482 (stacked on #439). Inspection data replaces caller goal
arenas but never acts as a proof receipt. Publication re-verifies retained source,
registers its generated host theorem through the repository API, and uses the
unchanged commit/push hooks and an exact source-revision lease. A bounded run
finishing does not mean every task was proved or published. Failed archival or cleanup retains the workspace. Stopped or
signalled tools also retain it: direct-child cleanup does not establish that all
descendants have terminated. Exact name matching alone does not establish the
identity of an arbitrary imported environment.
Expand All @@ -175,9 +182,10 @@ uncertain execution retains the library and stops the cohort. Its
`r30/research-cohort-result/1` report preserves individual outcomes, and exit 0
requires all checked results and no cancellation. Cohort execution is bounded
reuse, not durable recovery or automatic strategy selection.
Source inspection, strategy scheduling, durable failure diagnostics, recovery,
and claim publication still need completion; these modules do not replace the
legacy engine yet. Add every new application or library module and its modeled
The autonomous command connects inspection, bounded scheduling, recovery and
publication for supplied tasks. It does not generate a complete research strategy,
resume uncertain stored processes, or replace the legacy scheduled engine yet.
Add every new application or library module and its modeled
features to `ResearchCatalog`, and add behavioral evidence for the capability;
a module-count theorem alone does not establish feature completeness.

Expand Down Expand Up @@ -417,6 +425,11 @@ State the missing executable connection and its acceptance condition explicitly.
- Before another costly run after a failure, identify the failing obligation,
distinguish proof failure from resource failure, and make one evidence-based
correction. Preserve every existing obligation when partitioning proofs.
- Semantic `axioms` annotations are exact sets, not upper bounds. Audit new and
changed declarations against Lean's actual dependencies before the full gate;
do not copy an annotation from a neighboring definition. Check the whole
changed declaration set together so unused or missing annotations do not
produce a succession of late policy failures.
- When several behavioral proofs evaluate the same expensive serialized fixture,
establish its decoding once and reuse that checked equality. Keep the actual
wire-decoding obligation and each behavioral statement; do not substitute an
Expand Down
Binary file modified Rule30-Core.pdf
Binary file not shown.
Binary file modified Rule30-Observations.pdf
Binary file not shown.
Binary file modified Rule30-Paper.pdf
Binary file not shown.
10 changes: 5 additions & 5 deletions pdf-manifest.json
Original file line number Diff line number Diff line change
@@ -1,18 +1,18 @@
{
"documents": {
"Core": {
"pdf_sha256": "0412d2a3bd4a1f856877b4576fa9d3c74cf20110254b3e8c08e88f9c513ef5c6",
"pdf_sha256": "4096cc4d659f31be9cd86bdb962639e8ae7f67b2fb1d10b3456a84ff41f36c06",
"tex_sha256": "a7cd1807077baa1d1ca34015faf500a24477991859427f4db646e5f230fe54e2"
},
"Observations": {
"pdf_sha256": "4c3af528e93971b79c6ff12a31ac92ce502e40fe6c35656036156256a4739af6",
"pdf_sha256": "98c0f20646a1c0c2153355c5fda10dcaf4c26fe98fec37d86d9416f7e14d17e2",
"tex_sha256": "056820f37ac40133e341f46ad6c81125e35981150a1cc86f2063faad3f26fbd4"
},
"Paper": {
"pdf_sha256": "28247b25fc3192f97ca2454315ae49530d4fb91cf7d363cb4cfee14c8e6291a6",
"pdf_sha256": "a60f48e7c14ec6ab0e5aa17ae9cc4e848f5c5c61a612339acf1091666f355247",
"tex_sha256": "000f988561d9621f9ba2f9a99aba971670864367ebd336e6fbad0e1f71da01cb"
}
},
"generated_at_utc": "2026-11-11T21:48:32.936749+00:00",
"source_digest": "9290b3cca596a4281b852998983ea680eb6706ed332321ae9034b400112ea3e4"
"generated_at_utc": "2026-11-12T20:06:19.560839+00:00",
"source_digest": "a0895002261a43f4ad4af9368d22ad039c69ae02739db0d827297b94a3348578"
}
79 changes: 71 additions & 8 deletions rule30-api/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -25,8 +25,8 @@ gate verifies these modules with the rest of the application.
`production/research-request.json` exports `Research.routeDead`, `Research.step` and
`ResearchCatalog.initial` through the pinned compiler. These pure library roots
do not replace the running Python engine. Native Host drivers now execute
individual attempts and shared-library cohorts through `ResearchCommand` and
`ResearchCohortCommand`. Owned candidate verification binds kernel replay and
individual attempts, shared-library portfolios, and journaled autonomy through
`ResearchCommand` and `ResearchCohortCommand`. Owned candidate verification binds kernel replay and
axiom audit to the exact goal; the controller's receipt matching checks
applicability, not mathematical truth.
Legacy audit classifications and universally closed hypothesis schemas are
Expand All @@ -39,9 +39,61 @@ successful, normally exited owned exchange acknowledges the transition.
Duplicate successors, stale predecessors, cancellation and uncertain process
outcomes cannot acknowledge a commit. `ResearchJournalTraces` checks these
boundaries. A journal acknowledgement is neither a perpetual lease nor proof
authority: enclosing recovery still needs fresh workspaces and quarantine limits,
and claim publication still needs source integration, verification and register
updates. Inspection and autonomous portfolio execution remain unfinished.
authority. `ResearchRecovery` allocates fresh generations with bounded quarantine;
`ResearchPublication` owns source integration, fresh verification, register updates
and revision-conditional publication. `ResearchAutonomy` connects their lifecycles.

`ResearchPortfolio` connects the existing scheduler to owned sessions. The native
`ResearchCohortCommand` also accepts `r30/research-portfolio/1`, with a `base`,
`tasks` and `budget` (at most 64). The base uses
`r30/research-portfolio-base/1` and the attempt configuration fields except
`arena` and `proof`. Tasks use the existing ready-task schema and have distinct
IDs. Owned local kernel inspection replaces caller-supplied goal arenas before
dispatch. Unsupported inspected types block only the affected tasks. After each
outcome the scheduler selects the next task; local constructive
synthesis supplies its candidate. One owned library serves the run, with tracked
sources rechecked between attempts. Unsupported synthesis blocks that executor;
it does not refute or exhaust the goal. Budget completion and cancellation clear
unstarted selections. Uncertain execution retains ownership and stops dispatch.
The result schema is `r30/research-portfolio-result/1`; `completed` means the
bounded run ended, not that every task was proved or published. Each result
retains its individual proof status and evidence identity; `tasks` reports the
remaining task statuses. This command runs supplied ready tasks. Automatic
generation of a complete research strategy is a separate capability.

The same native command accepts `r30/research-autonomy/1`, containing `portfolio`,
`record_path` and `max_quarantined` (at most 8). Recovery reads the durable journal
through owned Git processes, reserves the next generation with compare-and-swap,
and assigns fresh workspace and log paths. Earlier active generations enter
quarantine; recovery never signals a stored PID or reuses an uncertain workspace.
The generation is sealed only after the portfolio and publication have released
their resources.

Checked portfolio results enter one publication batch. Each archive must match
its canonical evidence identity, complete request and exact library revision.
The controller renames the retained LexLean source to an evidence-addressed
module, inserts its source path into the sorted, duplicate-free paper entrypoints,
then owns a new compiler snapshot and full verification. It registers
the attested generated host theorem through the existing `bin/r30 module add`
and `theorem add` operations. A closed proof may update its still-open matching
frontier claim; a premise proof publishes a supporting lemma and leaves the
parent claim open. These existing repository API operations remain dependencies
of the LexLean-authored controller; no new Python runtime is introduced.

All proofs in a batch share one publication workspace, one unchanged pre-commit
gate and one push against the exact source predecessor. The commit is retained
locally under an evidence-and-commit-addressed reference before pushing. Repeating
retention preserves the same commit without overwriting a different attempt.
A failed or uncertain effect retains the workspace;
only a successful remote acknowledgment permits normal publication cleanup.
`r30/research-autonomy-result/1` distinguishes attempt `results`, remaining
`tasks`, `published_evidence`, `published_commit`, and the `generation_record`.
An empty publication list is not a solved claim. The current bounded constructive
search does not establish that Rule 30 Problem 1 can be solved automatically.
Publication refuses statements the existing theorem API cannot express and retains
their evidence and workspace. Recovery starts a fresh generation and quarantines
uncertain previous work; it does not assume stored process identifiers are live.
No remote model is used by these stages.

`ResearchArchitecture.lex.tex` exposes the autonomous pipeline directly to
PrismPM's architecture, security and quality projection: nine components,
Expand All @@ -67,7 +119,18 @@ Each expensive native-binding closed-goal case has its own proof unit.
Verification lifecycle, policy, process and metadata traces use batches of
at most eight theorems, preserving their original obligations and axiom policy.
CI retains the pinned compiler's native timing and memory telemetry on success
and failure, without changing the gate, resource limits or process width.
and failure, without changing proof acceptance or process width. The API allows
at most 32 MiB of child output: the 30-root autonomous extraction measured
21,298,345 bytes, exceeding the prior 16 MiB allowance. The complete 30-root
program, crate and certificate bundle measures 104,325,479 bytes, so its aggregate
source budget is 128 MiB. All proof, replay and axiom checks remain enforced.
The pinned compiler partitions large preservation certificates at complete
declaration-command boundaries. Every part is elaborated and replayed under the
unchanged child timeout, and the original certificate's axiom audit covers the
imported chain. Parts live outside original certificate module prefixes so
`leanchecker` does not select them all again during the final replay. This
addresses the combined controller's measured timeout; it does not provide
cross-revision proof caching or reduce the proof obligations.

<h1 align="center">Centerline</h1>
<p align="center"><em>Every arrow a theorem.</em></p>
Expand Down Expand Up @@ -703,8 +766,8 @@ where this repository uses it or why it does not.
```
bin/r30 api sync # regenerate every module, register, schema, feature, CONFORMANCE.md and the logo; lock
bin/r30 api check # refuse drift (run by the tests and the gate)
bin/r30 api production # export the production roots to certified Rust and replay the bundle (about a minute)
bin/r30 api production # export the production roots to certified Rust and replay the bundle
bin/r30 api production --targets # the compiler's target descriptors
bin/r30 verify api # lexlean verify of the API modules (about five minutes; Convert.lex.tex decides 28 conversions, Rule30Core.lex.tex the first rows)
bin/r30 verify api # lexlean verify of the API modules; cost depends on the selected proof closure
bin/r30 view render [--out DIR] # the current view (the site, statements.json, explainer.json, rule30.svg, the PDFs)
```
4 changes: 2 additions & 2 deletions rule30-api/lexlean.lock
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
spec = "lexlean/lock/2"
language = "1.2"
compiler_semantics = "601e9f4a1824bb6aacf7c3c2243d4b33aaa315abdebd5ceaf920d178e196019e"
project_config_sha256 = "6ceb296fae3095cf9f42d6c53a9eeeee55cf9df8e6e1eba87fd6248553fd000d"
compiler_semantics = "2e4c0ddd6084eba8fa99f04ab0235bfc5775dbc0455216b6ec51abe8ce8d9750"
project_config_sha256 = "3aab69e83aaa43aee298ada2603049770fd5ba35d441dde63a941209cecdd225"
lean_toolchain = "leanprover/lean4:v4.32.1"

[[workspace_file]]
Expand Down
Loading
Loading