Keep the self-judged plan suite off the public site

CLAUDE.md

@@ -115,6 +115,24 @@ Co-evolution tools are integrated into GSD workflows:
 - Convergence is honest, never forced-silent. A `co-evolve-bouncer.sh` run ends in one of three recorded states (`state.json.convergence_status`): **converged** (markers hit 0 within the configured passes), **adjudicated** (markers survived, so one forced-adjudication pass resolved each one and wrote `adjudication-report.md` mapping every stripped marker to the chosen text + a one-line rationale), or **stuck** (adjudication could not defensibly resolve every marker — the working document is preserved WITH its markers, labeled NOT-final, and fails the scorer gate). A naturally-converging run is byte-identical to the pre-adjudication bouncer; the adjudication pass fires only when markers survive. Exit semantics by design: a stuck document run still exits 0 — the `CO-EVOLVE:STUCK` label and the failing scorer gate carry the signal, not the exit code; only `--execute` refuses a stuck plan, with exit 1.
 - Agent-bouncer overwrites the input file in place; orchestrators should back up first
 
+## What May Be Published as a Result
+
+The public results site carries **standardized benchmarks scored by their own
+official evaluators** and nothing else. Today that means SWE-bench Verified
+through the pinned harness in `benchmarks/code/`.
+
+The plan-composition suite (`benchmarks/`, batches `b1` and successors) is
+**internal only**. Every condition in it is composed by Fable and its primary
+judge is Fable, so it scores Co-Evolution with the model under evaluation. Run
+it, read it, act on it — and do not put it on the site, because publishing it
+beside official-evaluator numbers lends it their credibility. Its exporter
+writes to the ignored `benchmarks/results/site-export/` for that reason.
+
+One site: `benchmarks/site/`, deployed by `.github/workflows/pages.yml` from
+`benchmarks/site/public/`. Before adding a second, check whether that workflow
+would actually serve it — a previous results site under `docs/` documented a
+deploy trigger that did not exist and was never live.
+
 ## Status
 
 - v1.4 npm/MCP publish pending — a human gate, not yet shipped.

benchmarks/README.md

@@ -119,27 +119,34 @@ Writes `results/b1/REPORT.md` and a committed copy under
 verdicts, then per-judge matrices, judge-integrity stats, and the exploratory
 sections — in that order, matching `PREREGISTRATION.md` section 1.
 
-**6. Publish.**
+**6. Export, for internal review only.**
 
 ```bash
 bash benchmarks/export-site-data.sh --batch b1
 ```
 
-Writes `docs/data/b1.json` and regenerates `docs/data/index.json`, which is
-everything the results site at
-<https://alanshurafa.github.io/co-evolution/plans/> renders. Commit those two
-JSON files and push to `master`; the `pages` workflow redeploys on any change
-under `docs/index.html`, `docs/assets/` or `docs/data/`. That is the whole
-republish path — one command, then a commit.
+Writes `benchmarks/results/site-export/b1.json` and an index beside it, under
+the ignored results tree.
+
+**These numbers are not publishable and this suite is not a lane on the results
+site.** Every condition here is composed by Fable and the primary judge is
+Fable, so the suite scores Co-Evolution using the model it is evaluating. That
+is a useful internal signal and an indefensible public claim, and placing it
+next to official-evaluator numbers would lend it their credibility. The public
+site carries standardized benchmarks with official evaluators only — see
+`benchmarks/code/README.md`.
+
+An earlier version of this document described committing the export to `docs/`
+and a `pages` workflow that redeployed on changes there. No such trigger
+existed: `pages.yml` watches `benchmarks/site/public/**` and nothing else.
 
 The export reads the same artifacts `report.sh` does and runs the same
 win-matrix and Bradley-Terry awk, so the site and the report cannot drift
-apart. It carries only publishable fields — scores, costs, model and condition
-names, judge tallies, task ids and difficulty — and fails rather than emit a
-prompt, a plan, a judge's reasoning or an evidence quote.
-`benchmarks/tests/test-site-export.sh` holds it to that against a fixture
-batch, and `docs/README.md` covers local preview and the rest of the site's
-layout.
+apart. It carries only scores, costs, model and condition names, judge tallies,
+task ids and difficulty, and fails rather than emit a prompt, a plan, a judge's
+reasoning or an evidence quote. `benchmarks/tests/test-site-export.sh` holds it
+to that against a fixture batch. That redaction is what makes the export safe to
+share internally; it does not make a self-judged win rate safe to publish.
 
 Pass `--batch-dir` when a batch's results live outside `benchmarks/results/`,
 for instance in another checkout:

benchmarks/export-site-data.sh

@@ -37,7 +37,12 @@ command -v jq >/dev/null 2>&1 || die "jq is required for export-site-data.sh"
 BATCH_ID=""
 BATCH_DIR=""
 RESULTS_ROOT="$SCRIPT_DIR/results"
-DOCS_DIR="$REPO_ROOT/docs"
+# The plan-composition suite scores Co-Evolution against itself: every condition
+# is composed by Fable and the primary judge is Fable. That is a legitimate
+# internal signal and not a publishable result, so the export lands in the
+# ignored results tree rather than in a directory staged for the public site.
+# Point --docs-dir somewhere else deliberately if you have a reason to.
+DOCS_DIR="${CODE_BENCH_SITE_EXPORT_DIR:-$REPO_ROOT/benchmarks/results/site-export}"
 CORPUS_DIR="$SCRIPT_DIR/corpus"
 CONDITIONS_FILE="$SCRIPT_DIR/conditions.yaml"
 JUDGES_CSV="fable,codex,glm"