Publish frozen comparative baselines

CONTRIBUTING.md

@@ -7,6 +7,7 @@ Gloss improves when a benchmark claim becomes easier to inspect, reproduce, or c
 - Add a PowerPoint construct the corpus does not cover.
 - Reproduce a grader false positive or false negative.
 - Improve a candidate assertion and its mutation fixture.
+- Improve a deterministic comparative generation path and prove the score change.
 - Review a prompt requirement, evidence record, or release gate.
 - Make the documentation easier to run from a clean checkout.
 
@@ -29,7 +30,7 @@ Install [uv](https://docs.astral.sh/uv/), Python 3.12, LibreOffice Impress, and
 git clone https://github.com/aronchick/gloss.git
 cd gloss/acidslide-v1/grader
 uv sync --extra dev --locked
-uv run python ../benchmark/validate_corpus.py
+uv run ../benchmark/validate_corpus.py
 uv run pytest tests/test_mutation_fixtures.py -q
 ```
 
@@ -84,7 +85,10 @@ CI must pass. Maintainers may ask for a smaller fixture, clearer provenance, or
 
 ## Launch surface
 
-The static site lives in `site/`. Its public counts come from `site/evidence/preview-v1.json`, which is verified against the generated mutation fixture reports by:
+The static site lives in `site/`. Harness counts come from
+`site/evidence/preview-v1.json`; comparative bars come from the byte-equivalent
+public copy of `acidslide-v1/benchmark/comparative-v1/summary.json`. Verify both
+sources, local assets, copy, and media metadata with:
 
 ```bash
 node launch/verify-launch.mjs
@@ -96,7 +100,8 @@ Rebuild the silent 21-second launch film with:
 node launch/render-video.mjs
 ```
 
-The renderer requires `rsvg-convert` and `ffmpeg`. Do not hand-edit a published number in the site or video without updating the evidence bundle and its verification source.
+The renderer requires `rsvg-convert` and `ffmpeg`. It reads the comparative
+summary directly. Do not hand-edit a published number in the site or video.
 
 ## Conduct and security
 

README.md

@@ -23,6 +23,24 @@ Every launch count is recorded in [`site/evidence/preview-v1.json`](site/evidenc
 node launch/verify-launch.mjs
 ```
 
+## Frozen comparative bundle
+
+[`acidslide-v1/benchmark/comparative-v1`](acidslide-v1/benchmark/comparative-v1)
+contains four repository-owned generation paths, three public seeds per path,
+and 12 editable 20-slide decks. The canonical Linux/amd64 grader completed all
+240 slide renders.
+
+The current local artifact scores are 67.68% for the native paths and 62.32%
+for the visual paths. These are reproducible workflow baselines, not model
+rankings. Every report says `local artifact score; self-reported` and carries no
+model attribution.
+
+Reproduce every deck, report, hash, and public bar:
+
+```bash
+./acidslide-v1/benchmark/comparative-v1/reproduce.sh
+```
+
 Release mode intentionally fails until independent prompt convergence, assertion provenance and evidence, reviewer approvals, baselines, environment manifests, and signed release indexes are complete.
 
 ## Score provenance

acidslide-v1/.dockerignore

@@ -22,3 +22,7 @@ benchmark/release-index.json
 benchmark/release-index-chain.json
 benchmark/baselines/*.json
 benchmark/control-handoffs/*.json
+
+# Comparative decks and local self-reported grading outputs are mounted inputs,
+# never part of the renderer image they identify.
+benchmark/comparative-v1/

acidslide-v1/benchmark/comparative-v1/README.md

@@ -0,0 +1,52 @@
+# Gloss comparative v1
+
+This frozen bundle compares four deterministic, repository-owned PowerPoint
+generation paths:
+
+- `native-precise`
+- `native-fast`
+- `visual-precise`
+- `visual-fast`
+
+Each path has three seeded runs (`1103`, `2207`, `3301`). These are reproducible
+workflow baselines, not model rankings. Every result is labeled:
+
+> local artifact score; self-reported
+
+Every generation record is labeled:
+
+> repository-owned path; no model attribution
+
+## Reproduce every bar
+
+From the repository root:
+
+```bash
+./acidslide-v1/benchmark/comparative-v1/reproduce.sh
+```
+
+The command regenerates all 12 editable 20-slide decks, builds the pinned
+Linux/amd64 grader image, grades all 240 slides, freezes artifact hashes, and
+recomputes every published metric.
+
+## Published artifacts
+
+- `cohort.json` binds the local scoring cohort to the exact grader source,
+  container, prompts, checklist, schemas, fonts, and assets.
+- `manifest.json` freezes the hashes and metrics for all 12 runs.
+- `summary.json` is the only source used by the public chart and launch video.
+- `runs/<path>/run-<n>/` contains `deck.pptx`, `generation.json`,
+  `artifact-context.json`, `report.json`, and `artifact-sha256.json`.
+- `verify_bundle.py` independently checks archive structure, all hashes,
+  disclosure labels, report completion, cohort identity, and summary math.
+
+Metrics:
+
+- **Local fidelity** — AcidSlide’s severity-weighted artifact score.
+- **Visual SSIM** — mean full-slide similarity to the gold renders.
+- **Native pass** — severity-weighted checklist pass rate with every
+  `visual_ssim` check excluded.
+
+Local reports are intentionally ineligible for an official leaderboard. They do
+not assert that a hosted service verified the score or that any model generated
+the artifact.

acidslide-v1/benchmark/comparative-v1/cohort.json

@@ -0,0 +1,33 @@
+{
+  "disclosure": {
+    "attribution": "repository-owned path; no model attribution",
+    "comparison_scope": "reproducible workflow baselines, not model rankings",
+    "verification_label": "local artifact score; self-reported"
+  },
+  "grader_revision": "unavailable",
+  "provenance": {
+    "environment_attestation_sha256": "sha256:62695d02ec683aa3680f2db72ff6c17fcce14fccfb0b5367c75704d92db04602",
+    "grader_source_tree_sha256": "sha256:685ce0505346cf3e844d49ab25c51e906ec917d2bf6b6ea6bdb95e65d4e23d28",
+    "scoring_cohort_id": "sha256:13b3015b9e346d8eb4cd7f26fd13f3c551941a8b64a5bbd7e483063567bfff1f",
+    "scoring_manifest_sha256": "sha256:97a4df0aa45d198e0a1494d6c2501d8c9fb375c617aff1012588d843da2b20b8"
+  },
+  "scoring_manifest": {
+    "bundle_hashes": {
+      "asset_manifest_sha256": "sha256:813a36b7e34fec686863e0d94ec4c040344b1c1a27270cdb103e424f988f4e8c",
+      "checklist_bundle_sha256": "sha256:bcfe2c7c3647742eeb90edbef74c1de09651c05b4cdbc1b5cc4430b67f5e9d83",
+      "font_manifest_sha256": "sha256:a770fe1f2a127a532da82dd8f9063c43dcc51523b91575c676db4fac80bc7c55",
+      "grader_source_tree_sha256": "sha256:685ce0505346cf3e844d49ab25c51e906ec917d2bf6b6ea6bdb95e65d4e23d28",
+      "prompt_bundle_sha256": "sha256:1579cd0439040042b091e276b3d67d0fbb9dfbf7087bea80651fd40835c7b818",
+      "schema_bundle_sha256": "sha256:f9a55826487a67023137ef2319e9a99724469e38cf9ff5e78932801c4d9616a5",
+      "scored_assertion_inventory_sha256": "sha256:62be3978e59e177c819b6900964f0d543a4e39fec0b89fd5c45404ccf214fc47"
+    },
+    "disclosure": {
+      "attribution": "repository-owned path; no model attribution",
+      "comparison_scope": "reproducible workflow baselines, not model rankings",
+      "verification_label": "local artifact score; self-reported"
+    },
+    "environment_attestation_sha256": "sha256:62695d02ec683aa3680f2db72ff6c17fcce14fccfb0b5367c75704d92db04602",
+    "image_digest": "sha256:29f5fe733d762d2633f28824a18807c3273d02c040db56c0cda12affd246732d",
+    "schema_version": "1.0"
+  }
+}