# Tara taxonomy selection · source discovery · 8 September 2026

This prepares a required subject of the [full V1 program](v1-completion.md). The
[catalog guide](v1-tara-practice-catalog.md) explains content labels and object
construction; member-input selection and the actual safety-policy consumer are a
distinct explanation. This note records source inspection, not newly executed
selection or application acceptance. Source pin:
`1ceae52c7896a59ac74a4429d410f3218f0b91c6`.

## Three separate feature models

The complete feature mood, modality and theme modules were reviewed, along with
the complete domain Tara–Lilith mood/crisis bridge. They are separate from the
contract-exported taxonomy datasets. The catalog source fixture already loads
these complete feature modules, but its twenty catalog cases do not verify all
selection behavior. Full selection cases and the owning UI/service caller still
need review before this subject is closed.

Mood entries cover twelve named moods. Each entry contains a display name, a
base distress level, a slate and a handoff declaration. Each slate item carries
a theme, modality, duration bucket and rationale; the rationales are authored
policy text, not clinical efficacy evidence. Anxious has four entries; the other
moods have three. The high-distress base entries are agitated, grieving and
fearful. Their stored handoff defaults require the Lilith handoff route,
suppressPracticeStart true and resourceLevel crisis.

`resolveMoodRecommendationSlate` uses the supplied distressSignal when present,
otherwise the entry's base distress. Its helper named highestDistress simply
returns this selected value; it does not compute a maximum over both. Handoff is
still required if the base entry requires it, a classifier flag is true or the
effective signal is high/crisis. The resolver returns the original slate even
when handoff is required. A consumer must honor the separate handoff state;
merely displaying a returned suggestion does not establish permission to start.

The resolved handoff is rebuilt when required. Its reason prefers the entry's
stored reason, then a classifier reason, then a message about the effective
signal. Its resource level is crisis for an effective crisis signal or a true
classifier, otherwise support. Therefore a high-distress entry's stored crisis
resource level can become support in the resolution. A lower explicit signal can
change the reported effective level without clearing a base-required handoff.
These distinctions need executable examples, not a generic red/green mood table.

The fifteen theme entries map to compatible mood arrays and default modality
arrays. `resolveThemesForMood` filters this array in its declared order;
`defaultModalitiesForTheme` returns that theme's actual default array. These
helpers do not automatically intersect the mood slate, theme defaults and
modality constraints into a final ranked practice. No such integration should be
inferred from the shared identifiers.

The fourteen modalities include silent, guided, four breathwork kinds, three
sound kinds, movement, posture, visualization, prayer and devotional reading,
across seven families. They carry audio/movement needs, sensory load,
accessibility fallback, contraindication metadata and optional four-phase
cadence. `resolveModalitiesForConstraints` only excludes audio-required entries
when audioAvailable is explicitly false, movement-space-required entries when
movementSpaceAvailable is explicitly false, and movement when
reducedMotionRequired is true. Omitted availability flags leave those entries
eligible. The helper does not itself apply all contraindication metadata, check
an actual device or produce a working accessibility fallback.

## Validation boundaries

Mood catalog validation checks complete/unique mood coverage, supported moods, a
nonempty slate, unique themes and required handoff/route consistency for high
distress. Theme validation additionally checks nonempty/unique supported mood
and modality lists. Modality validation checks complete/unique supported values,
nonempty fallback text, family/audio/movement consistency and counted-breathing
cadence existence with a positive summed duration. These procedural validators
have a different boundary from full Zod schemas; a list of intended TypeScript
fields is not equivalent to validating every incoming numeric or string value.

## The Tara–Lilith policy bridge

`libs/oshun/domain-tara/src/lilith-mood-crisis-handoff.ts` consumes the feature
mood resolver and Lilith persona-policy functions. It normalizes an optional
classifier signal, resolves the mood slate, and invokes utterance detection only
for a nonempty trimmed utterance. A classifier object defaults triggered to true
if omitted; its confidence is copied by normalization but not used as a
threshold by this bridge.

Crisis-type precedence is explicit classifier type, first Lilith detection, then
the mood-derived type. A grieving mood-derived handoff maps to acute grief; the
remaining mood-derived handoffs map to acute panic. These are implementation
policy mappings, not a diagnostic assessment of a person. An explicitly supplied
classifier type can be selected independently of a false triggered property. The
existing owning tests and complete persona-policy detector/rule catalog must be
read and exercised before teaching that edge case as a validated behavior.

If no type is selected, the bridge returns practiceAllowed true and no rule.
Otherwise it returns required true, suppressPracticeStart true, practiceAllowed
false, matched patterns and a summary of the selected Lilith rule. That summary
includes required resources/statements, prohibited behaviors, post-crisis
actions, governance-review status and resume cooldown. This function returns
policy data; it does not by itself navigate, contact a person, open resources,
record governance review or enforce a button. The actual consumer still needs a
complete trace. A bounded pinned-source search found this feature resolver
consumed here; that search does not prove deployment-wide exclusivity.

## Teaching plan

Visual thesis: show a member-input decision with clearly separated mood,
constraints, candidate metadata and policy handoff. Use a decision tree, an
explicit precedence diagram and one worked slate specimen. Labels should show
which source/consumer made each choice.

Content plan: distinguish content-inferred labels from member inputs; inspect a
slate; explain distress overrides and base-required handoff; follow themes and
modality constraints; show the bridge and its selected rule; finish at the
actual surface enforcement and evidence needed to return to practice. Retain all
twelve moods, fifteen themes and fourteen modalities through an organized
reference or deepening route rather than placing the whole taxonomy on one
crowded slide. Trace the independent contract datasets and native callers.

Interaction thesis: native diagrams retain complete states in PDF; source
reveals expose the exact policy module and example inputs. Real source UI is
useful only once the owning consumer is found and its seams are explicit.

## Follow-up consumer and detector trace

During the web-search delivery, read the complete five-case owning bridge test.
It covers high-distress anxious, grieving/fearful mapping, a classifier
override, utterance override and peaceful input. These tests were read, not
rerun in that delivery. A pinned `git grep` across apps and libs for
`resolveTaraLilithMoodCrisisHandoff` found its definition and this test only. No
application call was located by that symbol search; this is not proof against an
indirect consumer, different implementation or deployment-specific wiring. Do
not draw a completed UI enforcement arrow from the bridge without a caller.

Read the complete `getLilithCrisisBehaviorRule`, `detectLilithCrises`,
`detectLilithCrisisSignals` and their matching/normalization helpers in the
1,957-line `crisis-behavior-policy.ts`. Matching lowercases utterances/context
and uses substring inclusion against declared phrases and context strings. The
detector visits the declared crisis-type order, retains only the first matching
signal within each rule, then appends taxonomy-only crisis types not already
present. It does not numerically sort the merged result by risk level. The
bridge takes its first detection. The comment calling the declaration order
severity order must not be expanded into a guarantee that every combined result
is sorted by measured severity. Signal minimumRiskLevel is metadata in this
matching path; no threshold comparison is performed here. Full policy catalog
and taxonomy-entry review, including competing matches, remain required before
source cases or comprehensive teaching acceptance.

The rule getter validates the crisis-type identifier and returns the catalog
object. The bridge’s returned summary is a declaration of required resources,
statements and actions; the getter has no IO or enforcement. Reviewed the full
acute-grief rule and initial parts of the broader catalog and resource registry.
The larger module also owns regional resources, response/follow-up plans,
reentry controls and no-bypass policy. Those additional bodies are not yet fully
reviewed in this selection trace and belong in the wider Lilith/platform depth.

## Additional policy-owner reading during service-discovery delivery

Read the complete regional normalization/registry resolvers, response-plan and
follow-up-plan builders, frame operation evaluator, elevated-authorization
predicate and no-bypass result builder in
`libs/oshun/persona-policy-lilith/src/crisis-behavior-policy.ts`. The module is
byte-identical at the primary pin and
`6830fb7b036c002a6453589593ac825645049538`; SHA-256:
`b9212aefb24e401012474d086fa99aefde889c4ab9a11690e51f385115a0b531`. This is
source reading for the next required subject, with no new executed policy cases
or application acceptance.

The regional normalizer accepts explicit aliases for US, GB, EU and BR and
otherwise returns DEFAULT. The public registry resolver merges universal,
regional and caller overrides in that order. The response-plan helper uses a
different entry condition: without a truthy resourceRegion it uses only the
supplied resourceRegistry or an empty object. It does not automatically choose
the universal default. A missing required kind becomes a synthetic resource with
priority zero, a missing-kind name and null phone/text. The coverage helper
checks only that each resolved resource kind has a trimmed name; that predicate
is not evidence that a contact is current, reachable or appropriate to a member.
Do not turn these source constants into current crisis-contact guidance.

The response builder validates crisis/signal identifiers, reads the selected
rule and returns plain-operator, synthesis-stop and memory-suspension flags,
statements, required resources, prohibited behavior, cooldown and a safety
incident record routed to crisis-trained reviewers. It does not itself send a
resource, stop a runtime, write an incident or arrange review.

The follow-up builder validates the consent enum and positive-integer delay and
retention values. Check-in status is scheduled only after opt-in; otherwise it
is awaiting-opt-in with no scheduled time. Event memory has a record only for
granted consent. A separate audit-retention record is returned even without that
event-memory consent. Defaults are 86,400 seconds for the check-in delay and
2,555 retention days; they are source policy defaults, not a legal retention
assessment. Date.parse plus ISO conversion calculates timestamps and rejects
unparseable input. The returned cooldown is always active and takes its duration
from the response plan. Scheduling, writing, retaining and stopping
recommendations still require an actual consumer.

Frame evaluation rejects unsupported operations, generative summarization and
practice recommendation. Prior contemplative-script recall is blocked only when
the explicit wouldReenterCrisisState flag is true or the supplied content
contains a listed case-insensitive substring. Other operation kinds are not
scanned for those reentry phrases by this function. The evaluator returns a
plain-operator fallback when blocked; it does not dispatch that fallback.

The elevated-authorization predicate checks approved status, nonempty ID and
rationale, supported scope/role, parseable grant/expiry and expiry after grant.
Expiry against the current time is checked only when now is supplied. It does
not check grant time against now or fetch/verify an authorization authority.
These edge cases need source execution and consumer review before being taught
as an accepted authorization mechanism.

The no-bypass helper maps every supplied competing directive into a rejection
record and returns fixed plain-operator, stop-synthesis and memory-suspension
controls. It preserves resources from the input response plan while using the
constant frame controls. The result's allCompetingDirectivesRejected is a
constructed declaration, including for an empty input list. It is not a record
of a runtime having enforced each instruction. The full rule/resource catalog,
competing detector matches, owning tests and application enforcement trace
remain open; these helper readings do not close that larger policy review.
