@crewhaus/provability 0.0.0-stage → 0.8.0

package/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

package/NOTICE

@@ -0,0 +1,9 @@
+CrewHaus Factory
+Copyright 2026 Max Meier
+
+This product includes software developed by Max Meier
+(operating as StudioMax — https://studiomax.io) and contributors.
+
+The CrewHaus name and logo are trademarks of Max Meier. See
+https://github.com/crewhaus/factory/blob/main/TRADEMARK.md for the
+use-of-marks policy.

package/README.md

@@ -1,3 +1,301 @@
-# Temporary Holding Version
+# @crewhaus/provability
 
-This version is a temporary placeholder for this package. An operational version to replace this has been submitted for review and is awaiting a staged release.
+The provability ladder: `cited`, `grounded`, `checked`, `proven`. A spec sets a floor, and every
+claim in an answer must bind to evidence at or above it. This package is the deterministic part
+of that: the evidence index built from the run's own event log, the four verifiers behind one
+function (`verifyDocument`), the proof bundle they produce, the ways to render it, and where
+bundles and proof artifacts are stored.
+
+It makes no model call and imports no model adapter (`crewhaus doctor --philosophy-alignment`
+checks that). The one model-backed step in the feature, the opt-in paraphrase waiver, is a
+callback the caller injects, and so are the other two things that reach outside the process: the
+HEAD request behind `sources.resolve: head` and the replay of a pure tool. The in-loop gate, the
+`ProvabilityPreflight` tool, `crewhaus provability check` and the research report gate all call
+the same `verifyDocument`, so a draft the model preflights is graded exactly as the gate will
+grade it. Every text operation (claims, markers, definitions, the evidence-reference grammar,
+locating a claim in a body, numbers, quotations, coverage) is
+[`@crewhaus/claim-ledger`](../claim-ledger)'s; nothing here re-implements one.
+
+```ts
+import {
+  buildEvidenceIndex,
+  createBundleStore,
+  createFileArtifactStore,
+  evidenceFromMessages,
+  evidenceFromSessionLog,
+  proofsRoot,
+  renderFooter,
+  verifyDocument,
+} from "@crewhaus/provability";
+
+const root = proofsRoot(process.cwd(), "support-bot"); // <cwd>/.crewhaus/proofs/support-bot
+const evidence = buildEvidenceIndex(
+  [
+    ...(await evidenceFromSessionLog({ sessionId })), // the log, sub-agent children included
+    ...evidenceFromMessages(turnMessages), // the turn still in flight
+  ],
+  { sessionId },
+);
+const bundle = await verifyDocument(answerText, {
+  config: ir.provability, // the lowered, fully resolved IrProvability
+  evidence,
+  artifacts: createFileArtifactStore({ rootDir: root }),
+  replay, // optional: re-run a tool by name with its logged input
+  isPureTool, // optional: (toolName) => whether replay may run it
+  resolveHead, // optional: a guarded HEAD, for `sources.resolve: head`
+  judgeWaiver, // optional: the paraphrase waiver
+  retryIndex,
+});
+const path = await createBundleStore({ rootDir: root }).put(bundle, { sessionId, turnNumber });
+console.log(renderFooter(bundle, { path }));
+// ── provability · grounded ✓ · 6/6 claims · 3 sources · 0 assumptions · .crewhaus/proofs/…
+```
+
+## API
+
+**Verifying.** `verifyDocument(text, opts: VerifyOptions): Promise<ProofBundle>`. Besides
+`config` and `evidence`, the options are `workspaceRoot` and `baseDir` (where path sources
+resolve), `stateRoots` (the run's own state directories, never a source), `artifacts` (a
+`ProofArtifactStore`; absent, every `proof:` binding is `proof_missing`), `recheckProof`,
+`trustStoredProofs`, `replay` and `isPureTool`, `resolveHead` and `headDeadlineMs`,
+`judgeWaiver`, `turnNumbers` (`scope: turn`), `retryIndex` and `now`. Beside it:
+`unverifiedBundle(config, error, retryIndex?, now?)` for a caller that caught a throw,
+`bundleHash(bundle)`, `claimReasons(claim)`, `isToolAllowed(toolName, allow)`,
+`statementMismatch(proved, sentence, bar)`, `independenceKey(origin)`, and the bounds
+`MAX_HEAD_REQUESTS` (50), `HEAD_CONCURRENCY` (6), `HEAD_DEADLINE_MS` (15 000) and
+`MAX_GROUNDING_TOKENS`. The errors this package raises are `ProvabilityError`s (a
+`CrewhausError`): `code: "config"` for a malformed argument (a config without a valid `level`, an
+invalid session id, spec name, turn number or artifact), `"runtime"` for a file the store
+refused or could not read.
+
+**Evidence.** `evidenceFromSessionLog({ sessionId, rootDir?, includeChildren?, maxDepth?,
+workspaceRoot?, toolResultsRoot?, inputSchemaOf?, onUnreadable? })` and
+`evidenceFromMessages(messages, opts?)` produce `EvidenceRecord`s; `recordFromCall(call, opts?)`
+builds one; `buildEvidenceIndex(records, { sessionId? }): EvidenceIndex` merges them;
+`restrictToTurns(index, turnNumbers)` scopes one. `evidenceFromSessionLog` throws a
+`ProvabilityError` for a session id that is not `sess_<16 hex>`; a log it cannot read is
+reported, not thrown (below). The readers behind them are exported too: `groundingBody`,
+`servedFrom`, `servesPage`, `spilledOutput`, `sha256Of`.
+
+**Rendering.** `renderFooter(bundle, { path? })` (the one-line footer), `renderBundleJson`,
+`renderMarkdownSection` (the research report's `## Provability`), `renderNudge` (the retry
+message) and `summarizeBundle(bundle): ProvabilityGradedSummary` (the `provability_graded` and
+`provability_evaluated` payload).
+
+**Storage.** `proofsRoot(cwd, specName)`, `createBundleStore({ rootDir }): BundleStore`
+(`put`, `get`, `list`, `prune`), `createFileArtifactStore({ rootDir })`, and one definition of a
+proof id and an artifact's shape: `PROOF_ID_REGEX`, `newProofId`, `isProofId`,
+`isProofArtifact`.
+
+**Levels, origins, purity.** `levelRank` (none 0, cited 1, grounded 2, checked 3, proven 4),
+`rungOfRank`, `meetsFloor`, `maxRung`, `isProvabilityLevel`, `PROVABILITY_LEVELS`;
+`normalizeUrl`, `normalizePath`, `originKey`, `hostOf`, `hostMatches`, `registrableDomain`,
+`isPrivateHost`, `hasUserinfo`, `doiUrl`; and `isPureTool(tool)`, which tools replay may re-run
+(read-only, concurrency-safe, not destructive, no sandbox, internal scope, no io capability).
+
+**Types** (`types.ts`): `ProofBundle`, `ProofClaim`, `ProofBinding`, `UnmetReason`,
+`EvidenceRecord`, `EvidenceIndex`, `ProofArtifact`, `ProofArtifactStore`, `VerifyOptions`,
+`ProvabilityGradedSummary`, and `ProvabilityConfig` (the lowered `IrProvability`). The bundle's
+wire schema is the zod `ProofBundle` in `@crewhaus/gateway-protocol`, held to this type by its
+tests. `private-address.ts` is the repo's synchronised private-address classifier, a copy that
+`apps/cli/src/tool-registry.test.ts` sweeps and executes: never edit it here alone.
+
+## Who calls it
+
+- `@crewhaus/evaluation-service`'s gate, which builds the index from the session log and the
+  transcript, wires the stores, replay through the catalog, the guarded HEAD and the waiver judge,
+  and hands runtime-core the closures.
+- `@crewhaus/tool-proof`'s `ProvabilityPreflight`, which skips the HEAD, the replay and the judge.
+- `crewhaus provability check`, offline over any document (with `trustStoredProofs` when it has
+  no session log to read).
+- `@crewhaus/report-writer`'s research report gate, over the citation tracker's cache.
+
+## The rungs, as checked
+
+- **cited.** Every marker in the claim has a definition, and the definition names something
+  locatable: an `http(s)` URL or a `doi:` whose host passes `sources.allow` / `sources.deny`
+  (deny wins; a DOI is checked as `doi.org`) and is not a private address, or a workspace path
+  that stays inside the workspace with every symlink followed, and exists. A URL that puts a user
+  name before its host (`https://acme.example@evil.example/`) is `host_denied`. Free text counts
+  only under `sources.bibliographic: true`, and can never be grounded. With `sources.resolve:
+  head` the URL must answer one HEAD with 2xx or 3xx: at most `MAX_HEAD_REQUESTS` per
+  verification, `HEAD_CONCURRENCY` at once, all within `headDeadlineMs`; a URL past the cap or the
+  deadline is `unresolved`. A dangling `[2]` beside a good `[1]` leaves the claim unmet
+  (`no_definition`). A defined second marker that fails does not unmeet a claim another binding
+  meets; the bundle still shows it.
+- **grounded.** The cited origin was fetched or read in this run (not an error result, and not a
+  fetch that answered outside 2xx), and claim-ledger's `locateClaim` finds the claim in that body
+  at `grounded.coverage` within `grounded.window_tokens`. A quotation must appear verbatim; a
+  figure or a negation is never excused by a partial bar, at `coverage: 1` too. The binding
+  freezes an excerpt (240 characters at most, the claim at its tightest) and the body's digest,
+  so the bundle stays checkable after the transcript is evicted. A path under the harness's own
+  state (`.crewhaus/`, `.crewhaus-shared/`, and the `stateRoots` the caller names) is never a
+  source, and neither is a copy of one (`CopyPath`, `MovePath`, `ConcatFiles`, the archive tools,
+  followed back up to four steps): a session log holds every answer verbatim (`not_fetched`). A
+  body a redirect brought from a host `sources.deny` or the private-address rule refuses does not
+  ground the URL that led there. One verification reads at most `MAX_GROUNDING_TOKENS` of fetched
+  text; a source past the budget reads `below_floor`, with a detail naming the cap.
+- **checked.** `tool:<toolUseId>` resolves to a recorded, non-error call of a tool on
+  `execution.allow`: the graded session's own call first, then a sub-agent's; an id only several
+  sub-agents share is `tool_missing`. Every number the claim states is in the result as an exact
+  decimal, and every quotation word for word; a claim with neither must share
+  `grounded.coverage` of its words with the call. The result is read with the claim's words the
+  input handed over blanked out where the tool repeats them, so a tool that only echoes the draft
+  (NormalizeText of the sentence, a lint run over the answer) checks nothing, while a value
+  selected from a document (JsonQuery over a record) does. A `FactCrossCheck` result counts only
+  where it found this sentence `supported`. A figure's polarity is read word by word
+  (claim-ledger's `governingNegation`) in the statement it sits in: a figure the claim denies
+  must be denied in the result too, and "91 is not prime" or "91 (not prime)" does not check
+  "91 is prime". At a `checked` floor, with `execution.replay` on and a pure tool, the gate
+  re-runs it once per call (at most `max_replays` per verification) and compares digests: a
+  mismatch is `replay_mismatch`; a replay that could not run leaves the binding `checked` with
+  `replayed: false`.
+- **proven.** `proof:prf_<16 hex>` names a stored artifact a checker on `formal.checkers`
+  accepted, and the run's evidence holds the `ProofCheck` call that wrote it: accepted, with the
+  assumptions and axioms the file lists, given exactly the checker, claim, statement and proof
+  the file holds. The file sits in the workspace, where the model's file tools could forge one;
+  the run's record of the check is what counts. Where `recheckProof` is wired (the gate and the
+  preflight wire tool-proof's, the builtin checker), the checker accepts the stored proof again.
+  The sentence must state the artifact's recorded `claim` (`statementMismatch`): its words cover
+  the claim's at `grounded.coverage`, both state the same figures, and each word of the proof's
+  claim is denied in both or in neither, so a proof that no two meetings overlap does not prove
+  "two meetings overlap". The checker, the formal statement, the hypotheses the proof rests on
+  (builtin `given` steps, SMT asserts and definitions) and the axioms go into the bundle beside
+  the prose. Hypotheses are assumptions: counted in the footer, refused under `assumptions:
+  forbid` and capped by `max_assumptions`.
+
+## How each outcome is represented
+
+- A claim with no marker is `unmet` with zero bindings; `claimReasons(claim)` reports it as
+  `no_citation`.
+- An allowed assumption binding is `ok: false` with no `reason`: not evidence, not a failure. A
+  claim whose only bindings are allowed assumptions has status `assumption`, is left out of the
+  score on both sides, and is always counted in the bundle and the footer. A refused assumption
+  (`assumptions: forbid`, or past `max_assumptions` in document order) is `assumption_forbidden`
+  and fails the verdict whatever the score, with `achieved: "none"`; the footer shows every
+  declared one (`3 assumptions (2 refused)`).
+- Under `grounded.independent: n` at a `grounded` floor, a claim met only through sources needs n
+  distinct origins (`independenceKey`: registrable domain, DOI registrant prefix, or file), each
+  read off the host that served the body. Short of that, each of its source bindings becomes
+  `not_independent`. A claim met by an execution or formal binding is met regardless.
+- A source the run wrote itself (a `Write` or `Edit` of the file, a PUT to the URL, or a copy of
+  one) still grounds, because the file does say it now, but the binding carries
+  `writtenBy: { toolUseId, toolName }`, the bundle counts it in `selfWritten`, and the footer
+  reads `1 source (1 written by this run)`. A `FactCrossCheck` that found the claim only in such
+  files reaches `grounded` at most. An MCP result that only hands back text an earlier call of
+  the run sent (a stored row read back) is `not_fetched`.
+
+## Things worth knowing
+
+**The outside calls happen only where the floor needs them.** Rungs above the floor are still
+computed, because `achieved` is the highest floor the answer would have passed, but without the
+outside calls: the HEAD runs only at `cited` and `grounded`, the waiver judge only at `grounded`
+and only for a miss claim-ledger calls waivable (`isWaivableMiss`: words partly found, none of
+the missing ones a figure or a negation), and only while the claim still needs it. Replay runs
+only at `checked`.
+
+**Evidence comes from the log, merged with the transcript.** `evidenceFromSessionLog` walks child
+sessions through their `sub_agent_start` brackets and places a child's calls where the bracket
+sits. An id logged twice pairs its last `tool_use` with its last `tool_result`, the rule
+continuity-store applies. A result the runtime spilled to a file is read whole from this call's
+own file under the spill root, never through a link. A log that cannot be read to its end keeps
+the records before that point and reports it through `onUnreadable`: fewer records fails closed.
+A call is its session and its id together, because an OpenAI-compatible server numbers ids per
+conversation and a parent and its child can share one.
+
+**Origins are read off the call, and only where the result is that origin's content.** Among
+builtins: `WebFetch`'s and `Fetch`'s `url` (a `Fetch` only when its method is GET, HEAD or
+unset) and `Read`'s `path`, plus the final URL a WebFetch header names after a redirect. An MCP
+tool's `url`/`uri`/`href`/`path`/`file`/`file_path`/`filePath` names an origin only when the
+tool's declared schema lists that field (`inputSchemaOf`, default `defaultCatalog`'s), and, where no
+schema is known, only when the input carries nothing else a call could have been pointed at.
+The runtime's own tools whose results are model output (`Task`, `Consult`, `Escalate`, `Skill`,
+the memory, plan and goal tools) name no origin. Hits are indexed under their own origins:
+`Retrieve`'s `[n] id=… doc=…` blocks in exactly the structure Retrieve prints, and JSON
+`url`/`uri`/`href`/`link`/`source` hits from an MCP tool whose input names no URL or path. When a
+body is grounded against, WebFetch's header and the `[user prompt: …]` line it prepends, and
+Fetch's status line and headers, are removed.
+
+**The text a definition line shows is read.** `[8]: https://acme.example/x Payments clear
+within 5 days` renders as that sentence, so what it shows past its destination
+(`definitionGloss`) is verified as claims citing that definition (`definition: "[8]"`). A link
+title (`"…"`, `'…'`, `(…)`) renders as nothing and is not read, which is how to keep a page
+title out.
+
+**A vacuous pass is backstopped.** An answer with no claims passes every floor, which is right
+for a greeting or a clarifying question and wrong for an answer whose claim the extractor did
+not see. So when nothing counted is extracted (`vacuous.ts`), what a reader sees, code,
+definition lines, URLs and list-item numbers aside, is cut far more bluntly: at every line break
+(U+2028, `<br>` and the rest), every sentence terminator, every semicolon, em dash, en dash (not
+one between figures) and table pipe, never at a comma. The answer is vacuous only when no segment
+carries a marker and no segment outside a question states a figure; otherwise each such segment
+is a claim, verified like any other.
+
+**The nudge crosses no trust boundary.** `renderNudge` is built from the model's own sentences,
+the identifiers it cited and fixed reasons, ASCII apart from the model's own words. It never
+includes a fetched body, an excerpt or judge text; the one checker output it repeats is a
+rejected proof's reason and counterexample.
+
+**`verifyDocument` does not throw for anything a document can contain.** An unreadable artifact,
+a failing judge or replay and a refused HEAD each become a binding's reason; it throws a
+`ProvabilityError` only for a malformed config. A caller that catches anything else serves
+`unverifiedBundle`, so the footer reads `? unverified`, never `✓`. `bundleHash` is a sha256 over
+the canonical JSON of every other field, keys sorted at every depth.
+
+**Storage.** Bundles go to `.crewhaus/proofs/<spec>/<sessionId>/turn-<n>.json`, and a bundle put
+without a session id under `no-session/` (`NO_SESSION`). A later pass of the same turn (a retry)
+goes beside it as `turn-<n>.<k>.json`, never over it, so every `bundleHash` an audit record
+carries names a bundle still on disk. `get` reads a turn's last pass and `list` each turn's last
+pass unless asked for all. Artifacts go to
+`.crewhaus/proofs/<spec>/artifacts/<prf_id>.json` and are never replaced. Files are mode 0600,
+written atomically through `@crewhaus/tool-safety/fs`, never through a link; spec names, session
+ids, turn numbers and proof ids are validated before any path is built from them. Bundles have
+no TTL; `prune` removes old ones.
+
+## Limits
+
+Deliberate, or known and not closed in this release:
+
+- A fabricated but well-formed URL passes `cited`: that rung checks the reference, not the page.
+  `isPrivateHost` is offline, so a public-looking name that resolves privately is caught by the
+  fetch that reads it, not here.
+- `grounded` is faithfulness, not truth: a source that is wrong grounds a wrong claim.
+- `registrableDomain` is a pragmatic eTLD+1 table, not the Public Suffix List: every
+  `*.github.io` site counts as one origin under `grounded.independent`.
+- At `grounded`, a negation before the span the claim is found in counts only when it governs
+  the claim's first word, so "It is not the case that the meetings overlap on Tuesday afternoon"
+  grounds "The meetings overlap on Tuesday afternoon".
+- At `checked`, a result's negation counts only where it governs the figure or one of the
+  claim's own words (`governingNegation`: before it, at most two words between): "1,284.50 is
+  not correct" and "it is not the case that the total is 1,284.50" both check "The total is
+  1,284.50". A question answered by a bare "No." is not read either. An identifier the claim
+  restates unquoted (a UUID, a hash) is compared by value only where the number layer reads a
+  figure in it, and otherwise falls back to word coverage: quote identifiers. A claim a tool
+  reassembles from several input strings, or a draft padded past twice the claim's length, is
+  not caught as an echo.
+- At `proven`, the coverage runs from the proof's claim into the sentence, so a sentence that
+  adds a clause of its own to the proved statement passes. That residual sits with the
+  formalization gap, shown beside the claim (the formal statement), not closed.
+- Self-written text goes unmarked when `Bash` or `RunCommand` wrote it (a command line names no
+  place), when a write names the place in a form that cannot be mapped to the cited one (a
+  repository path behind a raw-file URL), or when an earlier session wrote it. An MCP store row
+  written before the evidence begins is read as the server's own text.
+- A reflection a tool rewrites on the way out (an SQL function applied to the literal, an address
+  printed re-spelled) still grounds; closing that needs the spec to say whose hits are trusted.
+  A search hit that repeats a phrase the model searched for, word for word, is refused instead,
+  and fetching the page grounds it.
+- A forged hit header after a blank line inside the last Retrieve hit's preview still reads as
+  one more hit; closing that needs Retrieve to print a delimiter a preview cannot contain.
+
+## Tests
+
+```sh
+bun test src
+```
+
+No test needs a network, a model or Docker. The suites pin each rung's reasons, the evidence
+readers against the formats the real tools print, the self-written trace (`verify-provenance`,
+`verify-self-written`), the backstop (`vacuous`), the renderers and the store's file rules.
+
+Design: [design/0.8.0-provability-release.md](../../design/0.8.0-provability-release.md) §2, §5,
+§6, §10.

package/dist/artifact.d.ts

@@ -0,0 +1,9 @@
+import type { ProofArtifact } from "./types";
+/** `prf_` and 16 lowercase hex: the only shape a proof id has. */
+export declare const PROOF_ID_REGEX: RegExp;
+/** Is `id` a proof id (`prf_<16 lowercase hex>`)? */
+export declare function isProofId(id: unknown): id is string;
+/** A fresh proof id: `prf_` and 16 hex characters from the system CSPRNG. */
+export declare function newProofId(): string;
+/** Does `value` have every field a `ProofArtifact` must, with the right types? */
+export declare function isProofArtifact(value: unknown): value is ProofArtifact;

package/dist/artifact.js

@@ -0,0 +1,38 @@
+/**
+ * Proof-artifact identity and shape (design §6.4): one definition of what a
+ * `prf_` id is and what a stored artifact must hold, shared by the store
+ * (which refuses anything else before a path is built from it) and the
+ * `proven` verifier (which refuses anything else before reading a field).
+ */
+import { randomBytes } from "node:crypto";
+/** `prf_` and 16 lowercase hex: the only shape a proof id has. */
+export const PROOF_ID_REGEX = /^prf_[0-9a-f]{16}$/;
+/** Is `id` a proof id (`prf_<16 lowercase hex>`)? */
+export function isProofId(id) {
+    return typeof id === "string" && PROOF_ID_REGEX.test(id);
+}
+/** A fresh proof id: `prf_` and 16 hex characters from the system CSPRNG. */
+export function newProofId() {
+    return `prf_${randomBytes(8).toString("hex")}`;
+}
+const strings = (value) => Array.isArray(value) && value.every((item) => typeof item === "string");
+/** Does `value` have every field a `ProofArtifact` must, with the right types? */
+export function isProofArtifact(value) {
+    if (typeof value !== "object" || value === null)
+        return false;
+    const v = value;
+    return (v["version"] === 1 &&
+        isProofId(v["id"]) &&
+        typeof v["checker"] === "string" &&
+        typeof v["claim"] === "string" &&
+        typeof v["statement"] === "string" &&
+        typeof v["proof"] === "string" &&
+        typeof v["ok"] === "boolean" &&
+        strings(v["assumptions"]) &&
+        strings(v["axioms"]) &&
+        typeof v["output"] === "string" &&
+        typeof v["outputDigest"] === "string" &&
+        typeof v["checkedAt"] === "string" &&
+        (v["counterexample"] === undefined || typeof v["counterexample"] === "string") &&
+        (v["reason"] === undefined || typeof v["reason"] === "string"));
+}

package/dist/builtins.d.ts

@@ -0,0 +1,32 @@
+/** tool-web's fetcher: its result opens with a `URL: …\nStatus: …` header. */
+export declare const WEB_FETCH = "WebFetch";
+/** tool-fetch's fetcher: its result opens with `HTTP <status>` and the headers. */
+export declare const FETCH = "Fetch";
+/** tool-fs's reader: its result is the file's text. */
+export declare const READ = "Read";
+/** tool-retrieve's tool (pipeline and every `knowledge:` corpus): `[n] id=… doc=… score=…` hits. */
+export declare const RETRIEVE = "Retrieve";
+/** tool-proof's checker tool: its result is the JSON report of the artifact it wrote. */
+export declare const PROOF_CHECK = "ProofCheck";
+/** tool-verify's claim checker: its result is a JSON verdict per claim, with the source text behind it. */
+export declare const FACT_CROSS_CHECK = "FactCrossCheck";
+/** tool-proof's preflight: its input is the model's draft, which it only reads (design §7). */
+export declare const PROVABILITY_PREFLIGHT = "ProvabilityPreflight";
+/**
+ * Is `name` a tool an MCP server serves (`mcp__<server>__<tool>`, tool-mcp's
+ * naming)? Its result is what that server returned — classified at the
+ * `mcp` origin on the way in, and not the runtime's own model output: the
+ * runtime's other non-builtin tools are that — `Task` returns a sub-agent's
+ * final message, `Consult`/`Escalate` a sibling model's reply, `Skill` a
+ * skill body, the memory, plan and goal tools what a model stored.
+ *
+ * Not "never text the model wrote", though: a server with storage (SQL, a
+ * key-value or notes store, a vector index) hands back what was written to
+ * it, by this run or an earlier one. The verifier refuses a claim whose
+ * words an MCP result carries only as text the run itself sent earlier
+ * (verify.ts `echoCheck`); rows stored by an earlier SESSION are the
+ * residual the README names.
+ */
+export declare function isMcpTool(name: string): boolean;
+/** The builtin spec key a recorded tool name maps to (case-insensitively), or `undefined`. */
+export declare function builtinKey(name: string): string | undefined;

package/dist/builtins.js

@@ -0,0 +1,46 @@
+/**
+ * Which recorded tool names are builtins (`@crewhaus/tool-categories`), which
+ * an MCP server serves, and the ones whose result formats the evidence index
+ * and the verifiers read.
+ */
+import { builtinKeyForName } from "@crewhaus/tool-categories";
+/** tool-web's fetcher: its result opens with a `URL: …\nStatus: …` header. */
+export const WEB_FETCH = "WebFetch";
+/** tool-fetch's fetcher: its result opens with `HTTP <status>` and the headers. */
+export const FETCH = "Fetch";
+/** tool-fs's reader: its result is the file's text. */
+export const READ = "Read";
+/** tool-retrieve's tool (pipeline and every `knowledge:` corpus): `[n] id=… doc=… score=…` hits. */
+export const RETRIEVE = "Retrieve";
+/** tool-proof's checker tool: its result is the JSON report of the artifact it wrote. */
+export const PROOF_CHECK = "ProofCheck";
+/** tool-verify's claim checker: its result is a JSON verdict per claim, with the source text behind it. */
+export const FACT_CROSS_CHECK = "FactCrossCheck";
+/** tool-proof's preflight: its input is the model's draft, which it only reads (design §7). */
+export const PROVABILITY_PREFLIGHT = "ProvabilityPreflight";
+/**
+ * Is `name` a tool an MCP server serves (`mcp__<server>__<tool>`, tool-mcp's
+ * naming)? Its result is what that server returned — classified at the
+ * `mcp` origin on the way in, and not the runtime's own model output: the
+ * runtime's other non-builtin tools are that — `Task` returns a sub-agent's
+ * final message, `Consult`/`Escalate` a sibling model's reply, `Skill` a
+ * skill body, the memory, plan and goal tools what a model stored.
+ *
+ * Not "never text the model wrote", though: a server with storage (SQL, a
+ * key-value or notes store, a vector index) hands back what was written to
+ * it, by this run or an earlier one. The verifier refuses a claim whose
+ * words an MCP result carries only as text the run itself sent earlier
+ * (verify.ts `echoCheck`); rows stored by an earlier SESSION are the
+ * residual the README names.
+ */
+export function isMcpTool(name) {
+    return name.startsWith("mcp__");
+}
+/** Builtin keys by recorded name, memoized: the table is 500+ entries long. */
+const keys = new Map();
+/** The builtin spec key a recorded tool name maps to (case-insensitively), or `undefined`. */
+export function builtinKey(name) {
+    if (!keys.has(name))
+        keys.set(name, builtinKeyForName(name));
+    return keys.get(name);
+}