What the station's ratings of its own breaks are for

[robert-dean]

Status (2026-10-10)

The place to hold an opinion is built, and so is the first thing that reads it.

• The store. deadair.script_ratings (migration 0022) is a table on the write ATTEMPT, not a
  column, with the catalog's liked / neutral / disliked spellings over 1 / 0 / -1.
  render/script.ratings.repository.ts reads and writes it, and GET /scripts carries each
  attempt's rating (script.history.repository.ts).
• The review surface. /scripts has a rating control on every attempt that has words
  (apps/web/src/components/scripts/script.rating.control.tsx), and the Android app's What it said
  list rates them too (2ed8bfc).
• Reader 1, the notebook's distil selection. ScriptHistoryRepository.writtenBy leaves out an
  attempt rated -1 (rated.rating is distinct from -1), so a note is never distilled from a break
  the operator thumbed down.

What is left is everything that reads the ratings after that, one more place to write them, and one
reader that was missed:

• Reader 2, a persona performance page. Nothing in the console or the contracts reads ratings
  per persona.
• Reader 3, proposed sheet edits.
• Rating from the on-air panel. It rates the record that is playing and never the break.
• Stories do not read ratings. persona.stories.repository.ts (saidFor, around line 222)
  still carries the comment saying the clause is one left join through segment_id to
  deadair.script_ratings, "once that table is read here". It is read everywhere else now, so this
  is the same predicate as reader 1 at a second door.

Why the rubric is not enough, and why the readers below lean on the operator

deadair.script_history is one row per write ATTEMPT with the writer, the model, the template, the
neighbours, the token counts and the duration. CharacterFault
(persona.sheet.ts) is a closed vocabulary of
four ways a script is not the character speaking, and BreakPromptShape.mustNameRecord adds a fifth.
So "this persona fell to the floor 40% of the time and 31 of those were out-of-character" is
already a query rather than a feature.

That is a rubric, not an opinion, and optimising against it is a trap the tree already documents
one level down. overusedWords asks and never refuses, on the stated grounds that dictionMarkers
are named in the prompt and counted in the answer, so the cheapest way to pass the character check is
to say the marker list again: a check that declined over it would be refusing the character for
being itself. Any pass that tuned a sheet to reduce declines would be steering toward exactly that:
the decline rate improves, and the station gets worse.

The other candidate signals do not exist and cannot be manufactured:

• The audience is a count. AudienceWatch polls Icecast and follows /admin/eventfeed, and
  neither says anything about a sentence. A listener who leaves during a break and a listener whose
  train went into a tunnel are the same number.
• The operator's recorded opinions are all about music. artists, albums and tracks each
  carry a rating and CandidatesRepository.effectiveRating collapses the three. Nothing anywhere
  holds an opinion about a break.

Where it is written

Two surfaces, both writing the same row. The first is built; the second is what is left:

• /scripts already exists and paginates the whole history newest first
  (scripts.page.tsx, reached as the Scripts
  tab of /voice rather than a route of its own, and GET /scripts in
  render.ck). This is the review surface: an evening
  spent reading back what the station said. The page already filters by kind, writer, outcome and
  segment, and each row carries its rating. Built.
• The on-air panel is where rating in the moment belongs, because the judgement worth having is
  the one made while the break is still in the room. It needs the id of the attempt that produced
  what is playing, which the segment already reaches through
  ScriptHistoryPageQuery.segmentId.

The route exists already, PUT /scripts/{id}/rating under platform.manage (render.ck), mirroring the three catalog rating
verbs including their override of the file's read floor: reading what the station wrote is a view,
having an opinion about it is an operator action.

What reads it, in the order the value arrives

1. The persona notebook's distil selection. The notebook pass
(personas §2) reads a character's own written scripts and distils notes it will
build on later. A note distilled from a break the operator thumbed down is the station reinforcing
exactly what did not work. The predicate is one clause, and rating is distinct from -1. Built in
ScriptHistoryRepository.writtenBy; the stories repository still lacks it (see the status above).

2. A persona performance read. The decline rate by fault kind beside the liked and disliked share,
per persona and per kind. This is where a sheet actually gets tuned by hand, and it needs no model:
it is a page over two tables. It also answers the question the rubric cannot: whether the breaks that
PASSED every check were any good.

3. The deferred half: proposed sheet edits. A rated corpus is what makes "this character would be
better if diction said X" a claim with evidence rather than a guess. Three rules, and each has a
precedent in the tree:

• It proposes and never writes. deadair.pronunciations makes rejected a state rather than a
  deletion for exactly this reason, and CONFIDENCE_BAR is the line between saying a thing unasked
  and offering it. A persona is the operator's voice; a station that quietly rewrote its own character
  overnight is a change they can neither see nor undo.
• Every proposal carries its evidence, on deadair.facts' rule that a claim with no source must
  not be expressible: a proposed diction marker arrives attached to the rated rows that motivated it.
• It is judged by REHEARSAL, not by a diff. PersonaRehearsalService runs against a deliberately
  FIXED pair of invented records so two readings are comparable, which is precisely an A/B harness for
  a candidate sheet against the current one. It exists and nothing uses it this way.

What this does not cover, deliberately

• Listener feedback. There is none. The mount is a mount, AUDIENCE_LINGER_MS is a lease, and
  nothing about a connection is an opinion. Do not build an inference from listener churn: a break is
  seconds long and a departure has a dozen causes.
• Any path that changes a sheet without the operator accepting it. See the three rules above.
• Rating what the station PLAYED. That is artists / albums / tracks and it is built. This is
  about the words.
• A model rating its own breaks. A model asked to check its own work in the same breath approves
  it, which is why fact.model.ts's verification is a separate conversation that has never seen the
  article. There is no equivalent trick here, because "was that any good" has no quote to be entailed
  by. The operator is the judge or there is no judge.

Related

personas for the notebook this feeds and for what is left after the character sheet.
dj-voice for the writers themselves.
fact-enrichment for the propose-with-evidence pattern in its original home.

[robert-dean]

Status check, 2026-09-22 (main at 23b983e)

Done: the ratings table and repository, the rating control on /scripts, and reader 1: writtenBy skips a break rated -1.

Still open: reader 2 (a persona performance page), reader 3 (proposed sheet edits), rating a break from the on-air panel (it rates tracks only), and persona.stories.repository.ts still does not read ratings.

[robert-dean]

Status, 2026-10-23. The stories reader is built (#759). PersonaStoriesRepository.saidFor, which feeds both a later break and the nightly recap of a running bit, now leaves out a telling whose segment's newest written attempt is rated -1. That is the same predicate writtenBy applies to the notebook, judged on the attempt that aired.

Still open: reader 2 (a persona performance read with ratings beside the decline rate by fault kind), reader 3 (proposed sheet edits), and rating a break from the on-air panel. Reader 3 also has a standing obstacle: script_history is pruned after render.scriptHistoryDays and ratings cascade with it, so the rated evidence it would rest on does not last.