@@ -10,6 +10,32 @@ GitHub Releases.
 
 Use this section for changes that are merged but not released yet.
 
+### Added
+- Continuous integration. Quoto was the only bot in the family with no
+  `.github/` at all: the suite in `tests/` ran only when somebody remembered,
+  and nothing checked that a migration could be applied or undone before it
+  reached the host. `ci.yml` runs the tests, compiles every source file, checks
+  that each package in `requirements.in` is pinned in `requirements.txt`,
+  applies the migrations against a real PostgreSQL 17 and steps one back down
+  again, refuses a branched alembic history, and builds the image.
+- `release.yml` publishes the image to GHCR and creates the GitHub Release from
+  the matching `CHANGELOG.md` section on a `v*` tag, refusing a tag that is not
+  an ancestor of `main` or that has no changelog section. That is steps 6-8 of
+  `docs/releases.md`, which were done by hand.
+- `deploy/ws04/compose.yaml`, the production stack, pulling the released image
+  from GHCR. The stack on the host built its own image from a working copy, so
+  what served users was not a published artifact and nothing on the host could
+  say which commit it came from. `QUOTO_IMAGE` has no default: an unset one
+  stops the stack rather than quietly starting something else.
+
+### Fixed
+- The Dockerfile no longer defaults `QUOTO_VERSION` to `0.10.2`. It stayed at
+  that value through the whole of v0.10.3, so an image built without the
+  argument labelled itself as a release it was not.
+- A second `## Unreleased` section, left behind by the v0.10.2 release, sat
+  between v0.10.2 and v0.10.1. Release preparation renames `## Unreleased` to
+  the new version, and with two of them the wrong one could be renamed.
+
 ### Changed
 - Media description prompts now cast the model as the eyes and ears of someone
   who can't see or hear the file: on-screen text and speech must be quoted
@@ -96,55 +122,6 @@ Use this section for changes that are merged but not released yet.
 - AI audit retention is intentionally unchanged and remains available for debug
   and quality verification.
 
-## Unreleased
-
-Use this section for changes that are merged but not released yet.
-
-### Removed
-- Legacy `OPENROUTER_MODEL` setting. The quote evaluation model is now read only
-  from `OPENROUTER_EVAL_MODEL`, whose default is `poolside/laguna-s-2.1:free`;
-  a blank value falls back to that default. Previously the two settings aliased
-  each other at startup and whatever was written into `OPENROUTER_MODEL` was
-  silently overwritten.
-
-### Added
-- `scripts/bench_web.py` + `scripts/bench_web_ui.html`: a local web hub for
-  picking an eval model. Lists the days recorded in `logs/ai_audit.jsonl`, shows
-  the day's messages and what production picked, runs any set of models against
-  that day with a chosen reasoning effort, streams each model's reasoning and
-  answer live, then compares the picked quotes with tokens, cost and timing.
-  Runs are archived under `logs/bench_runs/`. Read-only, and it sends nothing
-  until Run is pressed. Sampling temperature is never set, matching production.
-  Each selected model can carry its own reasoning effort (or follow the global
-  one), and the catalog sorts by release date, alphabetically or by price, with
-  filters for structured-output support and free models.
-  A second, fully separate tab benchmarks media description: upload an image,
-  video or audio file, watch it go through the bot's own normalization pipeline
-  (resize, re-encode), and compare how each model describes it and at what cost.
-  Each tab keeps its own model selection, efforts, settings and results, since
-  the two jobs need different models. The media tab narrows the catalog to the
-  input modality the uploaded file needs; image-generating models (anything
-  that answers with pictures) and `:batch` tier ids (refused by
-  `/chat/completions`) are dropped from the catalog entirely. The effort picker
-  offers only the efforts a model actually accepts, read from the catalog's
-  `reasoning` block, and flags models where reasoning is mandatory or where the
-  effort is ignored. When a model asks for context, the card shows the whole
-  context block in the exact order production would publish it, reusing
-  `scoring._valid_context_messages` so the two can't drift apart. The summary
-  table sorts by price ascending out of the box, and every column header is a
-  sort toggle; rows without a value (failures) always sink to the bottom.
-- `scripts/compare_eval_models.py`: replays days saved in `logs/ai_audit.jsonl`
-  against several eval models in parallel and prints, per model, the quote it
-  picked, the day verdict, token usage and the dollar cost. Read-only — nothing
-  is written to the database.
-- Token usage and cost of the winning eval call are now requested from
-  OpenRouter (`usage.include`) and stored on `ai_evaluation_runs`
-  (`prompt_tokens`, `completion_tokens`, `reasoning_tokens`, `total_tokens`,
-  `cost_usd`); migration `20260907_01`.
-- Startup warning for renamed env keys (`OPENROUTER_MODEL`,
-  `OPENROUTER_REASONING_EFFORT`). `extra="ignore"` used to swallow them, so a
-  stale `.env` kept settings that had stopped applying.
-
 ## v0.10.1 - 2026-07-21
 
 ### Changed