causal-continuity-engine 0.2.0__tar.gz → 0.4.0__tar.gz

{causal_continuity_engine-0.2.0 → causal_continuity_engine-0.4.0}/.github/CONTRIBUTING.md

@@ -53,7 +53,7 @@ do. `just setup` installs every direct and transitive tool from
 `requirements-dev.lock` in pip's all-or-nothing hash-checking mode, refuses
 source distributions, then installs `causal-continuity-engine` editable without dependency
 resolution or an isolated build environment. The reviewed direct inputs are
-`pytest` 9.1.1, `ruff` 0.16.7, `build` 1.6.1, `jsonschema` 4.26.0,
+`pytest` 9.1.1, `ruff` 0.16.10, `build` 1.6.1, `jsonschema` 4.26.0,
 `setuptools` 84.0.0, and `pip` 26.2.1. None is imported by the engine; jsonschema validates
 emitted receipt instances only in the test/release gate.
 
@@ -258,7 +258,11 @@ outcome, say what distinguishes it from "nothing happened".
 
 **No authorship attribution to tooling.** The `.githooks/` hooks and the CI
 attribution guard both reject it, in commit messages and in committed files.
-Do not work around them.
+Do not work around them. This is a rule about credit, not secrecy: AI coding
+assistants are used in this project's development, and every change is
+reviewed, tested, signed off, and committed by the maintainer, who is
+responsible for it. If AI tools did significant work on your patch, say so in
+plain words in the pull request; never list a tool as an author or co-author.
 
 ## Security
 

{causal_continuity_engine-0.2.0 → causal_continuity_engine-0.4.0}/.github/scripts/check_attribution.py

@@ -76,6 +76,13 @@ _ARTIFACT_AI_CREDIT = re.compile(
 _TOOL_BYLINE = re.compile(
     rf"{_AI_ALIAS.pattern}\s*\((?:anthropic|openai)\)[.!]?"
 )
+# An AI coding session appends provenance rather than credit: a trailer whose key
+# names the tool ("<tool>-Session: <link>") and a bare link to the session. Neither
+# names an actor, so the credit grammar above never reached them.
+_AI_SESSION_KEY = re.compile(
+    rf"{_AI_ALIAS.pattern}[- ]session(?:[- ](?:id|link|url))?"
+)
+_AI_SESSION_LINK = re.compile(r"<?https?://claude\.ai/code/session_[a-z0-9_-]+/?>?")
 _TRAILER_KEYS = frozenset({
     "assisted-by",
     "authored-by",
@@ -233,6 +240,11 @@ def scan_prose(
         key, separator, value = line.partition(":")
         if separator and key.replace("_", "-") in _TRAILER_KEYS:
             reason = _actor_reason(value)
+        if reason is None and separator and _AI_SESSION_KEY.fullmatch(
+                key.replace("_", "-")):
+            reason = "tool/model session trailer"
+        if reason is None and _AI_SESSION_LINK.fullmatch(line):
+            reason = "tool/model session link"
         if reason is None:
             match = _CREDIT.fullmatch(line) or _ROLE.fullmatch(line)
             if match is not None:

{causal_continuity_engine-0.2.0 → causal_continuity_engine-0.4.0}/.pre-commit-config.yaml

@@ -43,7 +43,7 @@ repos:
       - id: check-merge-conflict
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: e8f7d69b957a30acb2c875078d5fd012de24b64c  # v0.16.7
+    rev: f12be1ebaa5351c1fc76472de98db2c3446c8253  # v0.16.10
     hooks:
       # Lint only, matching `just lint` and CI. Formatting is advisory in this
       # project and deliberately not a gate: the tree was hand-formatted before

{causal_continuity_engine-0.2.0 → causal_continuity_engine-0.4.0}/AGENTS.md

@@ -14,9 +14,13 @@ history and a bi-temporal causal graph. Event-derived projections are
 rebuildable only while source payloads remain inside the retention window;
 runtime records are authenticated/audited rather than replay-derived. Memory
 tiers L0–L4 and token-budgeted Resume Packets are what
-an agent actually receives on resume. When a requirement changes or evidence
-contradicts an assumption, causal invalidation computes the blast radius
-over typed edges, bounds it, and classifies it deterministically. When policy
+an agent actually receives on resume. When a requirement changes, when the
+source that stated an assumption withdraws it, or when an integration records
+contradicting evidence through `engine.invalidation.fire(...)`, causal
+invalidation computes the blast radius over typed edges, bounds it, and
+classifies it deterministically. It does not read prose for meaning; a trusted
+explicit negation of an active assumption's own wording raises a human-gated
+notice (ADR-133). When policy
 requires proof, a task cannot be marked complete without a signed proof envelope
 whose required verifiers actually ran — `Engine.complete_task` has twenty-two
 independently instrumented rejection gates, including current confirmed authority,
@@ -325,8 +329,11 @@ below what it is about to do (ADR-048).
 
 Never add AI authorship credit to a commit message, PR title or body, code
 comment, documentation file, changelog, or example. No co-author trailers,
-no "generated with" lines, no tool bylines, no mention of AI involvement in
-this codebase anywhere.
+no "generated with" lines, no tool bylines. This is a rule about credit, not
+secrecy: `.github/CONTRIBUTING.md` states plainly that AI coding assistants
+are used in development and that the maintainer reviews, tests, and commits
+every change. Disclosure in plain words is allowed; crediting a tool as an
+author is not.
 
 This is enforced, not requested. `.githooks/commit-msg` and
 `.githooks/pre-commit` block it locally (`core.hooksPath` is set to

{causal_continuity_engine-0.2.0 → causal_continuity_engine-0.4.0}/CHANGELOG.md

@@ -9,6 +9,61 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 No unreleased changes.
 
+## 0.4.0 — 2026-10-17
+
+### Changed
+
+- **Breaking: processor compatibility advances to `cce-processor/1.11.0`;
+  extraction stays `cce-deterministic/1.5.0`.** Preserve older projected stores
+  and their SQLite sidecars after stopping writers. Re-ingest retained sources
+  into a separate new store/project and review its proposals before confirming
+  them. Older projections do not upgrade in place; erased source payloads
+  cannot be reconstructed by replay (ADR-114, ADR-134).
+
+### Fixed
+
+- **A GitHub author association is trusted only in GitHub's exact spelling.**
+  Text authority upper-cased the association before comparing it, so a payload
+  carrying `owner` or `Member` was read as maintainer intent and could raise an
+  ADR-133 notice, while command authorization compared exactly and refused the
+  same value. Both now accept only `OWNER`, `MEMBER` and `COLLABORATOR`; any
+  other spelling is untrusted content. GitHub itself sends only the upper-case
+  values, so webhook data is unaffected (ADR-134).
+
+## 0.3.0 — 2026-10-16
+
+### Added
+
+- **A trusted explicit negation of an active assumption raises a human-gated
+  notice.** An OWNER, MEMBER or COLLABORATOR sentence that restates an active
+  assumption with one negator ("the feed is NOT ordered by timestamp"), negates
+  its cue ("we no longer assume …") or calls it wrong fires a
+  `contradictory_evidence` invalidation in `pending_confirmation`: nothing
+  changes until the owner records `confirm_invalidation` or
+  `reject_invalidation` through `authority --request`, with operands from
+  `Engine.authority_invalidation`. The notice appears in the resume packet and
+  makes `check` report `action_required`. The rule is closed and lexical; other
+  wordings are outside it (ADR-133).
+
+### Changed
+
+- **Breaking: processor compatibility advances to `cce-processor/1.10.0` and
+  extraction to `cce-deterministic/1.5.0`.** Preserve older projected stores and
+  their SQLite sidecars after stopping writers. Re-ingest retained sources into
+  a separate new store/project and review its proposals before confirming them.
+  Older projections do not upgrade in place; erased source payloads cannot be
+  reconstructed by replay (ADR-114, ADR-132).
+
+### Fixed
+
+- **A negated assumption cue no longer proposes the assumption it denies.**
+  "We no longer assume X", "We do not assume X" and "nothing relies on X"
+  extracted the same positive proposal as "We assume X", so a sentence written
+  to retire an assumption offered it for confirmation. The extractor now
+  abstains when a negator stands immediately before the cue. A prohibition or
+  decision form still proposes the constraint or decision the author wrote;
+  other wordings of a denial are outside this closed rule (ADR-132).
+
 ## 0.2.0 — 2026-09-23
 
 ### Changed

{causal_continuity_engine-0.2.0 → causal_continuity_engine-0.4.0}/CITATION.cff

@@ -8,8 +8,8 @@ authors:
   - given-names: Thomas
     family-names: Albrecht
     alias: The Quantum Falcon
-version: 0.2.0
-date-released: 2026-09-23
+version: 0.4.0
+date-released: 2026-10-17
 license: Apache-2.0
 repository-code: https://github.com/thequantumfalcon/causal-continuity-engine
 url: https://github.com/thequantumfalcon/causal-continuity-engine

{causal_continuity_engine-0.2.0 → causal_continuity_engine-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: causal-continuity-engine
-Version: 0.2.0
+Version: 0.4.0
 Summary: Causal Continuity Engine - continuity, causal invalidation, and proof for long-running coding agents
 Author: Thomas Albrecht
 License-Expression: Apache-2.0
@@ -27,7 +27,7 @@ License-File: LICENSE.txt
 License-File: NOTICE
 Provides-Extra: dev
 Requires-Dist: pytest==9.1.1; extra == "dev"
-Requires-Dist: ruff==0.16.7; extra == "dev"
+Requires-Dist: ruff==0.16.10; extra == "dev"
 Requires-Dist: build==1.6.1; extra == "dev"
 Requires-Dist: jsonschema==4.26.0; extra == "dev"
 Requires-Dist: pip==26.2.1; extra == "dev"
@@ -71,9 +71,16 @@ and neither is an unexamined green check.
 CCE addresses each with a mechanism rather than a better prompt. Decisions,
 assumptions, requirements and evidence become nodes in a bi-temporal causal
 graph over an application-enforced append-only, hash-chained event log, so
-*why* is a queryable edge and not a memory. When a requirement changes or evidence contradicts an assumption, the
+*why* is a queryable edge and not a memory. When a requirement changes, when the
+source that stated an assumption withdraws it, or when an integration records
+contradicting evidence through `engine.invalidation.fire(...)`, the
 blast radius is computed over typed edges, bounded, and classified
-deterministically. When policy requires proof, a task cannot be marked complete
+deterministically. CCE does not read prose for meaning. It does recognize one
+closed form: a trusted comment that restates an active assumption with a
+negator, like the one above, raises a human-gated notice that changes nothing
+until the owner confirms or rejects it; a differently worded dispute proposes
+nothing and fires nothing.
+When policy requires proof, a task cannot be marked complete
 without a signed envelope whose required verifiers actually ran — the rejection gates are
 mechanically enumerated and mutation-tested below.
 
@@ -97,7 +104,9 @@ command are already used by an unrelated published project.
 ## Use it from an editor
 
 `cce-engine mcp` speaks the Model Context Protocol over stdio, so an MCP client
-can read a project's control state directly. Four read-only tools:
+can read a project's control state directly. [docs/MCP.md](docs/MCP.md) walks the
+whole path on this repository, from install to a connected agent, and shows how
+the continuity check runs in CI on a runner that holds the store. Four read-only tools:
 `resume_packet`, `list_assumptions`, `list_invalidations`, `continuity_check`.
 
 The subcommand first ships in 0.1.6. The published 0.1.0, 0.1.2 and 0.1.3
@@ -136,9 +145,11 @@ than the official SDK, which would pull in more than twenty packages.
 
 ## Quickstart
 
-The installation commands below install the package PyPI currently serves; to
-run this exact checkout, including its authority and v2 packet examples, use
-the editable installation that follows.
+Install either the published package (the next blocks) or this checkout (the
+editable installation further down), not both: one virtual environment is
+enough. The published package runs this quickstart when this checkout is the
+release PyPI serves; for `main` between releases, or your own changes, use the
+editable installation.
 
 Python 3.11 or newer is required; the engine has no third-party runtime
 dependencies. Choose the block for your platform and run it in a new working
@@ -165,15 +176,15 @@ with `.\.venv\Scripts\python.exe -m pip install causal-continuity-engine`
 and invoke the environment's `cce-engine.exe` by its absolute path in the
 commands below. After activation, `cce-engine --help` checks the installation.
 
-**Upgrading an existing project:** this release uses processor 1.9.0 and
-extractor 1.4.0; projections written by 0.1.6's processor 1.8.0 or earlier
+**Upgrading an existing project:** this release uses processor 1.11.0 and
+extractor 1.5.0; projections written by 0.3.0's processor 1.10.0 or earlier
 processors are refused. There is no in-place projection or authority upgrade. Stop writers,
 preserve the complete old project state (including its database and any SQLite
 sidecars), and re-ingest retained sources into a separate new project/store.
 Review and explicitly confirm the new proposals there; old prose is not approval.
 Do not remove markers or sidecars to bypass refusal. Payloads already removed
 by retention cannot be reconstructed from their digests. New users can simply
-start a fresh project. See [the upgrade notes for this release](CHANGELOG.md#020--2026-09-23);
+start a fresh project. See [the upgrade notes for this release](CHANGELOG.md#040--2026-10-17);
 the [0.1.6 notes](CHANGELOG.md#016--2026-09-20) remain historical.
 
 To run this exact checkout instead of the published package, or to work on the
@@ -189,7 +200,20 @@ python -m pip install -e .
 ```
 
 On Windows, use the PowerShell environment-creation and activation commands
-above, then `python -m pip install -e .` from the checkout.
+above, then `python -m pip install -e .` from the checkout. Keep that
+environment active and `cd` back to your new working directory before the steps
+below, so `cce-demo` and its secrets are not created inside the checkout.
+
+The editable install is enough for the engine but not for the test suite,
+which fails at collection without the hash-locked development tools. Install
+them with `python -m pip install -r requirements-dev.lock` (or `just setup`),
+then run `python -m pytest tests` from the checkout. The project configuration
+already adds `-q`; adding another hides the closing pass/skip summary. The full
+suite can take around fifteen minutes. A few tests skip when the environment
+lacks what they exercise; `-rs` lists each skip with its reason. To see the
+completion gates described below exercised, run
+`python -m pytest tests/test_instrument_validation.py`: it plants each defect a
+gate exists to catch and checks that the named gate catches it.
 
 Global flags come **before** the subcommand: `cce-engine --dir <path> <subcommand>`.
 
@@ -391,6 +415,8 @@ Response: cli-markdown | max bytes: 131072
 The full human view renders each schema field or declares it as metadata kept
 in canonical JSON. Use `resume --format json` for those complete canonical
 field values, including signatures, digests and the state-basis object.
+Recent context lists retrieved nodes by their text; an event node has none, so
+it appears by its id alone (`-  (evt_...)`).
 
 The named verification gap is expected on a fresh project: proof is required
 by default, but CCE cannot safely invent a project-specific verifier
@@ -707,10 +733,17 @@ Published v1 schemas remain historical contracts; v1 proofs are not current
 completion evidence. The v2 schema URLs are bound to the `v0.2.0` release tag,
 and the release workflow verifies that each serves the checked-out bytes.
 
+`verify_proof.py` is a single file. Installing the published package places a
+copy under the environment's `share/causal-continuity-engine/audit/verifiers/`;
+with an editable install, use the checkout's `verifiers/verify_proof.py`. From
+the project directory:
+
 ```bash
 # The complete configured verifier set must already be pinned in project policy.
 cce-engine --dir . --json verify > proof.json
-python verifiers/verify_proof.py proof.json
+VERIFY_PROOF="$VIRTUAL_ENV/share/causal-continuity-engine/audit/verifiers/verify_proof.py"
+# editable install: VERIFY_PROOF=/absolute/path/to/this/local/checkout/verifiers/verify_proof.py
+python "$VERIFY_PROOF" proof.json
 ```
 
 ```text
@@ -718,6 +751,8 @@ UNVERIFIED means nothing was established, not that nothing is wrong.
 An hmac-sha256 envelope needs --hmac-key-hex; a lamport envelope needs --fingerprint
 from a source other than the envelope itself.
 UNVERIFIED C1:P C2:S C3:S C4:P C5:S  proof.json
+
+1 UNVERIFIED
 ```
 
 **That verdict is the honest one, and it is the point.** `hmac-sha256` is not
@@ -725,10 +760,15 @@ a third-party-verifiable scheme at all. Without the secret, an untouched
 envelope and a resealed forgery are indistinguishable, so the verifier reports
 `UNVERIFIED` — not `VALID` — and exits 2. A verifier that returned `VALID`
 there would be telling a stranger that a body a forger rewrote and resealed is
-sound. Supply the key and the same envelope verifies:
+sound. Supply the key and the same envelope verifies. For `hmac-sha256` the key
+is the project's local signing key, `.cce/secrets/signing.key`, given as hex.
+Anyone holding it can also seal a forgery, so this is the operator's own check,
+not a stranger's. The command below puts the key on the command line; run it
+only on a machine you control:
 
 ```bash
-python verifiers/verify_proof.py proof.json --hmac-key-hex <hex>
+python "$VERIFY_PROOF" proof.json \
+  --hmac-key-hex "$(python -c "import pathlib; print(pathlib.Path('.cce/secrets/signing.key').read_bytes().hex())")"
 # VALID      C1:P C2:P C3:P C4:P C5:S  proof.json
 ```
 
@@ -738,7 +778,7 @@ Reaching `VALID` as a stranger requires the fingerprint from somewhere other
 than the artifact:
 
 ```bash
-python verifiers/verify_proof.py proof.json --fingerprint sha256:...
+python "$VERIFY_PROOF" proof.json --fingerprint sha256:...
 ```
 
 The verifier's input boundary is deliberately finite. One invocation accepts
@@ -763,8 +803,9 @@ and `INVALID` dominates — a batch containing both an `INVALID` and an
 Check that the reference implementation still agrees with the corpus:
 
 ```bash
+# from the checkout
 python vectors/generate.py --check
-# reference agrees with every committed vector
+# reference agrees with all 50 committed vectors
 ```
 
 ## What this does not do
@@ -899,6 +940,9 @@ condition.
 - [docs/PUBLIC-FLIP.md](docs/PUBLIC-FLIP.md) — the register of every control
   that could not be enabled while this repository was private, with the exact
   command that turned each one on at the flip.
+- [docs/MCP.md](docs/MCP.md) — connecting an agent over MCP, walked end to end on
+  this repository: install, init, ingest real events, confirm, verify, packet, check,
+  client configuration, and the freshness rule.
 - [docs/RELEASE.md](docs/RELEASE.md) — the reproducible, signed, immutable-ready
   release procedure and its continuity-bound attestation roadmap.
 - [docs/RESEARCH-ROADMAP.md](docs/RESEARCH-ROADMAP.md) — standards-backed

{causal_continuity_engine-0.2.0 → causal_continuity_engine-0.4.0}/README.md

@@ -36,9 +36,16 @@ and neither is an unexamined green check.
 CCE addresses each with a mechanism rather than a better prompt. Decisions,
 assumptions, requirements and evidence become nodes in a bi-temporal causal
 graph over an application-enforced append-only, hash-chained event log, so
-*why* is a queryable edge and not a memory. When a requirement changes or evidence contradicts an assumption, the
+*why* is a queryable edge and not a memory. When a requirement changes, when the
+source that stated an assumption withdraws it, or when an integration records
+contradicting evidence through `engine.invalidation.fire(...)`, the
 blast radius is computed over typed edges, bounded, and classified
-deterministically. When policy requires proof, a task cannot be marked complete
+deterministically. CCE does not read prose for meaning. It does recognize one
+closed form: a trusted comment that restates an active assumption with a
+negator, like the one above, raises a human-gated notice that changes nothing
+until the owner confirms or rejects it; a differently worded dispute proposes
+nothing and fires nothing.
+When policy requires proof, a task cannot be marked complete
 without a signed envelope whose required verifiers actually ran — the rejection gates are
 mechanically enumerated and mutation-tested below.
 
@@ -62,7 +69,9 @@ command are already used by an unrelated published project.
 ## Use it from an editor
 
 `cce-engine mcp` speaks the Model Context Protocol over stdio, so an MCP client
-can read a project's control state directly. Four read-only tools:
+can read a project's control state directly. [docs/MCP.md](docs/MCP.md) walks the
+whole path on this repository, from install to a connected agent, and shows how
+the continuity check runs in CI on a runner that holds the store. Four read-only tools:
 `resume_packet`, `list_assumptions`, `list_invalidations`, `continuity_check`.
 
 The subcommand first ships in 0.1.6. The published 0.1.0, 0.1.2 and 0.1.3
@@ -101,9 +110,11 @@ than the official SDK, which would pull in more than twenty packages.
 
 ## Quickstart
 
-The installation commands below install the package PyPI currently serves; to
-run this exact checkout, including its authority and v2 packet examples, use
-the editable installation that follows.
+Install either the published package (the next blocks) or this checkout (the
+editable installation further down), not both: one virtual environment is
+enough. The published package runs this quickstart when this checkout is the
+release PyPI serves; for `main` between releases, or your own changes, use the
+editable installation.
 
 Python 3.11 or newer is required; the engine has no third-party runtime
 dependencies. Choose the block for your platform and run it in a new working
@@ -130,15 +141,15 @@ with `.\.venv\Scripts\python.exe -m pip install causal-continuity-engine`
 and invoke the environment's `cce-engine.exe` by its absolute path in the
 commands below. After activation, `cce-engine --help` checks the installation.
 
-**Upgrading an existing project:** this release uses processor 1.9.0 and
-extractor 1.4.0; projections written by 0.1.6's processor 1.8.0 or earlier
+**Upgrading an existing project:** this release uses processor 1.11.0 and
+extractor 1.5.0; projections written by 0.3.0's processor 1.10.0 or earlier
 processors are refused. There is no in-place projection or authority upgrade. Stop writers,
 preserve the complete old project state (including its database and any SQLite
 sidecars), and re-ingest retained sources into a separate new project/store.
 Review and explicitly confirm the new proposals there; old prose is not approval.
 Do not remove markers or sidecars to bypass refusal. Payloads already removed
 by retention cannot be reconstructed from their digests. New users can simply
-start a fresh project. See [the upgrade notes for this release](CHANGELOG.md#020--2026-09-23);
+start a fresh project. See [the upgrade notes for this release](CHANGELOG.md#040--2026-10-17);
 the [0.1.6 notes](CHANGELOG.md#016--2026-09-20) remain historical.
 
 To run this exact checkout instead of the published package, or to work on the
@@ -154,7 +165,20 @@ python -m pip install -e .
 ```
 
 On Windows, use the PowerShell environment-creation and activation commands
-above, then `python -m pip install -e .` from the checkout.
+above, then `python -m pip install -e .` from the checkout. Keep that
+environment active and `cd` back to your new working directory before the steps
+below, so `cce-demo` and its secrets are not created inside the checkout.
+
+The editable install is enough for the engine but not for the test suite,
+which fails at collection without the hash-locked development tools. Install
+them with `python -m pip install -r requirements-dev.lock` (or `just setup`),
+then run `python -m pytest tests` from the checkout. The project configuration
+already adds `-q`; adding another hides the closing pass/skip summary. The full
+suite can take around fifteen minutes. A few tests skip when the environment
+lacks what they exercise; `-rs` lists each skip with its reason. To see the
+completion gates described below exercised, run
+`python -m pytest tests/test_instrument_validation.py`: it plants each defect a
+gate exists to catch and checks that the named gate catches it.
 
 Global flags come **before** the subcommand: `cce-engine --dir <path> <subcommand>`.
 
@@ -356,6 +380,8 @@ Response: cli-markdown | max bytes: 131072
 The full human view renders each schema field or declares it as metadata kept
 in canonical JSON. Use `resume --format json` for those complete canonical
 field values, including signatures, digests and the state-basis object.
+Recent context lists retrieved nodes by their text; an event node has none, so
+it appears by its id alone (`-  (evt_...)`).
 
 The named verification gap is expected on a fresh project: proof is required
 by default, but CCE cannot safely invent a project-specific verifier
@@ -672,10 +698,17 @@ Published v1 schemas remain historical contracts; v1 proofs are not current
 completion evidence. The v2 schema URLs are bound to the `v0.2.0` release tag,
 and the release workflow verifies that each serves the checked-out bytes.
 
+`verify_proof.py` is a single file. Installing the published package places a
+copy under the environment's `share/causal-continuity-engine/audit/verifiers/`;
+with an editable install, use the checkout's `verifiers/verify_proof.py`. From
+the project directory:
+
 ```bash
 # The complete configured verifier set must already be pinned in project policy.
 cce-engine --dir . --json verify > proof.json
-python verifiers/verify_proof.py proof.json
+VERIFY_PROOF="$VIRTUAL_ENV/share/causal-continuity-engine/audit/verifiers/verify_proof.py"
+# editable install: VERIFY_PROOF=/absolute/path/to/this/local/checkout/verifiers/verify_proof.py
+python "$VERIFY_PROOF" proof.json
 ```
 
 ```text
@@ -683,6 +716,8 @@ UNVERIFIED means nothing was established, not that nothing is wrong.
 An hmac-sha256 envelope needs --hmac-key-hex; a lamport envelope needs --fingerprint
 from a source other than the envelope itself.
 UNVERIFIED C1:P C2:S C3:S C4:P C5:S  proof.json
+
+1 UNVERIFIED
 ```
 
 **That verdict is the honest one, and it is the point.** `hmac-sha256` is not
@@ -690,10 +725,15 @@ a third-party-verifiable scheme at all. Without the secret, an untouched
 envelope and a resealed forgery are indistinguishable, so the verifier reports
 `UNVERIFIED` — not `VALID` — and exits 2. A verifier that returned `VALID`
 there would be telling a stranger that a body a forger rewrote and resealed is
-sound. Supply the key and the same envelope verifies:
+sound. Supply the key and the same envelope verifies. For `hmac-sha256` the key
+is the project's local signing key, `.cce/secrets/signing.key`, given as hex.
+Anyone holding it can also seal a forgery, so this is the operator's own check,
+not a stranger's. The command below puts the key on the command line; run it
+only on a machine you control:
 
 ```bash
-python verifiers/verify_proof.py proof.json --hmac-key-hex <hex>
+python "$VERIFY_PROOF" proof.json \
+  --hmac-key-hex "$(python -c "import pathlib; print(pathlib.Path('.cce/secrets/signing.key').read_bytes().hex())")"
 # VALID      C1:P C2:P C3:P C4:P C5:S  proof.json
 ```
 
@@ -703,7 +743,7 @@ Reaching `VALID` as a stranger requires the fingerprint from somewhere other
 than the artifact:
 
 ```bash
-python verifiers/verify_proof.py proof.json --fingerprint sha256:...
+python "$VERIFY_PROOF" proof.json --fingerprint sha256:...
 ```
 
 The verifier's input boundary is deliberately finite. One invocation accepts
@@ -728,8 +768,9 @@ and `INVALID` dominates — a batch containing both an `INVALID` and an
 Check that the reference implementation still agrees with the corpus:
 
 ```bash
+# from the checkout
 python vectors/generate.py --check
-# reference agrees with every committed vector
+# reference agrees with all 50 committed vectors
 ```
 
 ## What this does not do
@@ -864,6 +905,9 @@ condition.
 - [docs/PUBLIC-FLIP.md](docs/PUBLIC-FLIP.md) — the register of every control
   that could not be enabled while this repository was private, with the exact
   command that turned each one on at the flip.
+- [docs/MCP.md](docs/MCP.md) — connecting an agent over MCP, walked end to end on
+  this repository: install, init, ingest real events, confirm, verify, packet, check,
+  client configuration, and the freshness rule.
 - [docs/RELEASE.md](docs/RELEASE.md) — the reproducible, signed, immutable-ready
   release procedure and its continuity-bound attestation roadmap.
 - [docs/RESEARCH-ROADMAP.md](docs/RESEARCH-ROADMAP.md) — standards-backed

{causal_continuity_engine-0.2.0 → causal_continuity_engine-0.4.0}/causal_continuity_engine/__init__.py

@@ -7,7 +7,7 @@ authenticated runtime records are not replay-derived (ADR-001). Storage is
 SQLite implementing the PostgreSQL-first relational adjacency design (ADR-011).
 """
 
-__version__ = "0.2.0"
+__version__ = "0.4.0"
 
 SCHEMA_VERSIONS = {
     "anchor": "cce.anchor.v1",