# Chapter 1 · A stop the start button never reads · brief · 11 September 2026

The first chapter of the _Tara · continue_ track, rewritten in the Eve edition
on 11 September 2026. It follows the plan in `EVE-V1-REDESIGN-TODOS.md` and the
track brief in [tara-continue-track.md](tara-continue-track.md). The receipt is
under `verification/`.

## The job

Before a sitting starts, Tara is meant to ask one question: may a practice begin
now, or should it give way to crisis-aware support? The handoff bridge answers
it. It is one pure function over three inputs — a mood with an optional distress
level, an optional classifier signal, and optional typed text with context
strings — that chooses at most one of Lilith's thirteen crisis rules and returns
it with a permission flag and the mood's untouched list of suggestions.

The chapter teaches how that one type is chosen, what the answer carries, and
what would have to read it. Nothing does yet: the bridge has no application
caller, the web sit asks for a mood only after the practice has ended, and its
play button records a start without asking anyone.

**Running example.** A member reports grieving and types _I feel unreal_. The
bridge selects dissociation, not grief, returns practice allowed false, and
still returns the three grief suggestions — one of them a guided body scan that
the dissociation rule prohibits. Its reason sentence describes the grieving
mood. Executed against the pinned bridge on this host.

## Three learning promises (the cover)

1. **One type from three inputs** — a classifier's type, then the first
   detection, then the mood.
2. **Fields that disagree** — reason, triggers and matches can each describe
   another input.
3. **A stop with no reader** — the chosen rule comes back as data; no caller
   applies it.

## Three things to remember (the close)

1. **Read practice allowed first.** A slate comes back even when the practice is
   refused.
2. **First answer, not the gravest.** Classifier type, first detection, then
   mood; risk never re-sorts.
3. **A refusal needs a reader.** The chosen rule returns as data; no start
   control reads it.

**Bridge.** The next chapter moves to the context a practice is chosen in: how
long the member has, the hour, and the lineage behind the practice.

## Sections

Three sections, of six, six and six.

| Section                   | Slides                                                                                       |
| ------------------------- | -------------------------------------------------------------------------------------------- |
| Three inputs, one answer  | the running example, eight terms, the three inputs, the classifier, the mood, the text gate  |
| Choosing one type         | two detector passes, the order, the gravest loses, no trigger, four fields, the rule summary |
| Thirteen rules, no reader | the catalog, two pairs read in full, the resource planner, the finish screen, pressing play  |

## Treatment decisions

- **The six rule instalments became one registry and two comparisons.** The
  inherited guide read the thirteen rules as six slides of two or three. The
  catalog is now one thirteen-row registry, `tara-handoff-rules-immediate` —
  risk, response frame, how the session stops, what a restart needs, and the
  cooldown — and `tara-handoff-rules-trauma`, `-violence` and `-specialist` are
  registered as merged away into it, their narration tracks pruned. The two
  pairs that share a frame and differ on everything else stay as compare panels:
  panic and dissociation, which holds the running example's rule, and grief and
  psychosis-adjacent, which holds the mood's fallback. The registry's notes name
  the most specific prohibitions of the other rules; their full lists remain in
  the source.
- **The priority ladder became a layer stack.** A decision tree admits two
  questions and the selection has three; four ordered bands, each boundary the
  condition that falls through, show the order and where the running example
  stops.
- **The finish screen is the chapter's capture.** The bridge has no surface of
  its own. The pinned web sit's finish screen is the only place the sit asks for
  a mood, and apart from the Report dialog the only place it takes a sentence,
  and it does both after the practice has ended; three callouts on the existing
  capture say what each collects and where it goes.
- **The boundary became sequence lanes.** The inherited close described the
  enforcement chain in the abstract. It now traces the path a member actually
  takes — play, the local clock, the start POST, the store — and none of its six
  messages carries a mood or reads the permission.
- **A glossary.** `tara-handoff-vocabulary` defines the eight terms the chapter
  uses exactly, which is the glossary slide this track's setup asks for in
  chapter one.

## Corrections to the inherited guide

- The detector keeps the first matching **signal** of each rule, with every
  phrase of that signal that matched — not the first matching pattern. Executed:
  _I feel unreal, not in my body_ gives one dissociation detection carrying both
  phrases.
- The context-only case is decided in the detector's **second** pass: the panic
  rule's own context signal is _hyperventilation audible on voice channel_, and
  a bare _hyperventilation_ matches only the taxonomy entry.
- The resource planner's missing region is not only "not DEFAULT": the coverage
  helper reads the same missing region **as** DEFAULT and reports complete
  coverage, so the check passes for the call that returns placeholders.

## What this chapter does not establish

No application caller, no enforcement of the start, no audio or session
suspension, no delivered resource, no cooldown clock, check-in, review or
follow-up. No clinical approval of the policy and no claim that any resource
contact is current or reachable. The detector's phrases are exact; how often
they are false alarms or misses in real speech is not measured.
