@cejel/cejel 0.1.8 → 0.1.9

package/README.md

@@ -70,7 +70,7 @@ The npm package is scoped as `@cejel/cejel`; its executable remains the short co
 **GitHub Action** — score every PR and publish the badge:
 
 ```yaml
-- uses: BargLabs/cejel/action@v0.1.8
+- uses: BargLabs/cejel/action@v0.1.9
   with:
     min-score: "2.5"   # optional: fail the build below this
 ```
@@ -99,21 +99,21 @@ Microsoft Authenticode code-signing.
 **Docker / OCI.** The same MCP server is published as a non-root, multi-platform image:
 
 ```bash
-docker run --rm -i -v "$PWD:/workspace:ro" ghcr.io/barglabs/cejel:0.1.8
+docker run --rm -i -v "$PWD:/workspace:ro" ghcr.io/barglabs/cejel:0.1.9
 ```
 
 The image defaults to `cejel-mcp` over stdio. To use the CLI instead:
 
 ```bash
-docker run --rm -v "$PWD:/workspace:ro" --entrypoint cejel ghcr.io/barglabs/cejel:0.1.8 .
+docker run --rm -v "$PWD:/workspace:ro" --entrypoint cejel ghcr.io/barglabs/cejel:0.1.9 .
 ```
 
 Beginning with v0.1.7, the standard CLI identity flags are also available directly, while an
 invocation without arguments continues to start the MCP server:
 
 ```bash
-docker run --rm ghcr.io/barglabs/cejel:0.1.8 --version
-docker run --rm ghcr.io/barglabs/cejel:0.1.8 --help
+docker run --rm ghcr.io/barglabs/cejel:0.1.9 --version
+docker run --rm ghcr.io/barglabs/cejel:0.1.9 --help
 ```
 
 The OCI image carries an SBOM, maximum-mode build provenance, and a signed registry
@@ -199,6 +199,40 @@ This verifies the report schema and the report-to-attestation digest, repository
 timestamp, and outcome binding. It does not verify a signature or signer identity; the command
 prints that boundary on every successful verification.
 
+### Experimental Free LLM Pack
+
+The Free LLM Pack is an opt-in, pre-release static check for observable application-integrity
+and evaluation-hygiene weaknesses in supported LLM application code. It is still being calibrated;
+its results are experimental until the frozen untouched-cohort release gate reaches GO.
+
+```bash
+./cejel scan . --pack llm
+```
+
+The command writes the ordinary Cejel artifacts unchanged and adds `llm-report.json`,
+`llm-attestation.json`, and `llm-certificate.html`. The pack runs locally and offline: it does
+not call a model, execute the target application, or send source, prompts, or findings over the
+network. Its findings and status never alter the base Cejel score or verdict.
+
+The alpha recognizes fixture-backed JavaScript/TypeScript patterns around OpenAI, Anthropic,
+Vercel AI SDK, and named model-call shapes. It records LangChain imports as integration metadata,
+but a LangChain import alone does not establish v1 applicability or framework-specific coverage.
+Python coverage is narrower: it recognizes official OpenAI and Anthropic SDK call/response shapes
+for direct unsafe sinks, narrowly fixture-backed action validation and configured self-judge
+checks, unconditional model loops, and secret-like environment values in model requests. See
+[Free LLM Pack usage and limitations](./docs/packs/free-llm-usage.md) for the eight rule IDs,
+artifact verification, exact boundaries, and current limitations.
+
+The pack does **not** measure a model's hallucination rate, prove that an application is safe, or
+provide runtime enforcement. Its attestation is self-generated and unsigned unless an external
+signer binds an identity to it.
+
+The first frozen 24+24-repository calibration cycle ended in an
+[evidence-integrity NO-GO](./calibration/llm/reviews/v1.2-integrity-no-go-2026-07-23.md). The
+required actual-run compatibility record was not retained, so precision, recall, and
+false-positive rate are not estimable and Cejel makes no measured performance claim from that
+cycle.
+
 ### Flags
 
 - `<path>` — repo to score (default: current directory)
@@ -210,6 +244,9 @@ prints that boundary on every successful verification.
 - `--ingest <file|glob>` — fold another scanner's output into the score (repeatable). Accepts
   SARIF, OpenSSF Scorecard JSON, or the generic Cejel external-signal shape — format is
   auto-detected. See "Aggregate your scanners" below.
+- `--pack llm` — run the opt-in experimental Free LLM Pack and write its separate evidence,
+  attestation, and HTML certificate artifacts. Repeatable pack syntax is supported; selecting the
+  same pack more than once still runs it once.
 - `--quiet` — suppress the terminal certificate (files are still written)
 - `-h`, `--help` — print usage and exit successfully
 - `-v`, `--version` — print the version derived from the package manifest and exit successfully
@@ -324,7 +361,7 @@ a configurable `min-score` threshold. The scoring step makes no network calls an
 secrets.
 
 ```yaml
-- uses: BargLabs/cejel/action@v0.1.8
+- uses: BargLabs/cejel/action@v0.1.9
   with:
     min-score: '2.5' # optional; omit to never fail the check
 ```
@@ -485,6 +522,13 @@ scoring tool that has not been checked:
   detected even though `test.js` was present at the repository root. That was a detection gap
   in Cejel, not evidence about the repository. Rubric v6 recognizes AVA's `test.js` and
   `test-*.js` conventions; a regression fixture preserves the correction.
+- **A category with no measured dimensions looked like a real failing `0.0`.** Version 0.1.9
+  labels that state `not measured` in the HTML certificate and prints the measured-coverage
+  context. A genuine measured zero remains `0.0`; JSON report, summary, and badge contracts are
+  unchanged.
+- **A saturated metric rendered a broken-looking numerator larger than its denominator.**
+  Version 0.1.9 leads with the scoring cap and keeps the observation explicit — for example,
+  `2 signals (capped; 4 raw)` — instead of printing `4/2 signals`.
 
 Every one was a trust failure produced by *us* — false alarms about other people's code,
 silent omissions, inconsistent presentation, or a home-only scoring path — and we would