A background job opens one wide event for its whole run, and the per-request
memo cached the operator switches, module availability, module preferences,
provider presence and consent kinds on it. The nightly status passes loop
over every candidate inside that one event, so a switch turned off mid-run
did not stop the remaining users, and the wire re-check read the same cached
answer.
memoizePerRequest takes a freshInBackground option; every input that decides
whether AI work may run sets it, so a background caller reads them on every
call while a request still reads them once.
The Coach correlations tool, the per-metric card and the period
narrative scanned every channel, so a switched-off module still fed
their statistics and the text a model was grounded on. The matrix's
modules option is now required, each caller resolves the record's map,
and the Coach skips the lab pass with labs off. A registry guard holds
every caller to resolving the map itself.
The mask's doc comment now states what it does to the false discovery
correction: fewer channels is a smaller family, so q-values of the
remaining pairs move and a borderline pair can cross the threshold.
Brings in the jobs, stored-text and consent package and the extraction
routes. Conflicts resolved:
- The digest keeps its capability-driven reaction line and check-in,
now through the shared `aiCapabilityForRecord`; the local copy of it
is removed.
- The route inventories classify the extraction routes (documents,
labs OCR, medication text, fenced chat and attachments) as actions;
a capability call counts as the gate for the module inventory.
- With every caller moved, `requireAssistantSurface` and
`AssistantDisabledError` are deleted. The `assistant.disabled.*`
codes stay on the capability refusal, and the guard now asserts the
retired gate is gone.
- `aiCapabilityToServe` answers whether stored model text may be shown
from the record's own state, apart from who is reading.
Refs #975
The status chokepoint, the comprehensive briefing, reaction lines and
AI-composed nudges now ask `aiEgressRefusal` about exactly the chain they
are about to send to, instead of each pairing the operator-key check
with a surface receipt lookup of its own. The consent kinds come from the
capability table, so the wire and the published capability answer from
the same rule.
The document summary, its catch-up pass and the index backfill stop on a
closed switch or module before anything is picked, and leave "no
provider" and "consent needed" to the document pick, which re-checks the
wire for the provider it chose. A closed switch still only heals PENDING
to NONE; a withheld pick records UNAVAILABLE. The per-pick document
consent helper is removed, since nothing calls it any more.
The auto-read toggle mints the shared `AI_EXTRACTION_CONSENT_KIND`, and
the web grant helper goes back to minting the master grant only.
Refs #975
Brings in the extraction routes and the capability re-check at the wire
(`aiEgressRefusal`), so the jobs and stored-text work can share it.
Conflicts in the document routes, the fenced chat, the auto-read toggle
and its contract text resolve to the extraction branch: those files no
longer call the per-pick document consent helper.
The weight status note's snapshot now carries the person's own target
band and which way is progress, and its comparison signal is framed as
higher-better below the band and lower-better above it instead of the
neutral target-band framing. The weight prompt tells the model to judge
changes against that target and keeps the loss pace and milestone
guidance for when losing is the goal. The target is folded into the
input fingerprint, so editing it regenerates the note even when no
weight reading moved.
Refs #1006
Data never depends on AI. The Insights data reads no longer ask an AI
switch or the `insights` module (now only the AI analysis opt-out): ECG
list, detail and upload, rhythm events, derived scores, rule cards,
correlations and patterns, the baseline strip, health status, intraday
pulse, and the Today digest with its dismiss. Breathing screening moves
to the sleep module and lab changes to the labs module. Stored Coach
facts, conversations, message ratings and the seen stamp stay readable
and erasable while the Coach is unavailable. Cards now say
`provider: "rules"`.
Reads that mix data with model text answer 200, null the model text
and publish an `ai` state from the capability resolver, without warming
anything: comprehensive (`ai.briefing`, `ai.statusText`, `hasProvider`
from the presence probe), the derived assessment override, the period
narrative, nudge status, the seeded opener, about-me questions, the
workout paragraph, the pregenerate warm (`queued: false`), the Today
digest (lead, reaction line, check-in) and the dashboard briefing,
which is now applied per read after the snapshot cache. Hiding the
Coach no longer hides the briefing.
`insights/generate` GET and POST ask the `briefing` capability and
`insights/chat` POST the `coach` capability, with the typed refusal
envelope; a missing provider keeps the chat's `coach.provider.none`
frame and adds `reason`. The /coach page, the MCP ECG and intraday
tools and `aiAvailable` on the provider route follow the same answer.
An inventory of AI and data routes replaces the Coach gate inventory:
every route that asks a capability names it, and no data route asks.
Refs #975
Every AI job now asks its capability for the user before it builds a
snapshot, and skips with a `.skipped` annotation naming the reason: the
nightly pregenerate (briefing and statusText, each half on its own), the
status batch and the per-metric crons, the on-demand status worker, the
narrative cron, Coach nudges and their AI composer, the Coach memory
refresh, reaction lines, workout paragraphs, and the document summary,
catch-up and index jobs. Seven of these read no assistant switch at all
before, so the master switch now stops them too.
The cron candidate queries drop people who switched AI analysis off in
SQL, and no longer read `disableCoach` outside the Coach: hiding the
Coach hides the Coach. Coach nudges use the one presence probe, so a
central Codex account is no longer skipped; a missing consent still gets
the deterministic template. The arrival spine asks only the
provider-free half (switches and modules) before it enqueues, keeping
its module graph free of provider code.
Document reading needs an extraction receipt (ai_extraction or ai_full)
for any external provider; the auto-read toggle is the trigger, no longer
a consent shortcut, so a revoked receipt wins over a toggle left on. Lab
report scans get the same rule for a whole chain.
Two structural guards: every file that calls a provider is named with
the reason its call is admitted, and the chokepoints re-check a
capability; every AI job references its capability gate. An integration
matrix runs each worker under the refusing states with a provider spy.
Refs #975
Every status note reader and writer moves to InsightStatusCache: the
generators, the unchanged-data gates, the stale-while-revalidate read,
the ingest debounce, the previous-note prompt block and the
medication-compliance card with its per-medication lines. A stall now
opens a negative-cache window next to the note instead of a stub row
that shadowed it.
Stored model text is never served while its capability is unavailable,
whatever the reason: the status readers and gates check statusText, the
period narrative hides a model-written row while periodNarrative is
unavailable (the deterministic narrative is data and always shown), and
the daily digest drops reaction lines while reactionLines is
unavailable. A read-only miss now reports why (`unavailable` with the
capability's reason) and keeps hasProvider meaning provider presence, so
a withdrawn consent no longer reads as "no provider".
runStatusCompletion takes the capability as a required argument and
re-checks it before the provider chain is resolved, so a job queued
before a switch flipped still stops at the wire. The comprehensive
briefing re-checks briefing the same way. hasUsableStatusProvider is the
one presence probe.
Refs #975
requireAiCapability answers at the top of a route, before a provider is
picked, and a job enqueued before an operator switch flipped runs long
after any route answered. aiEgressRefusal asks again about the providers a
caller is about to send to: the capability for the record (the request's
own resolution, or the job's for the record it names) and, for exactly
those providers, whether a consent receipt of the kinds the capability
table declares is active.
A pickDecides gate option leaves no_provider and consent_required to that
pick, for routes that read the chain in their own order.
The coach's cycle snapshot read a year of raw readings for the phase
crosstab, including steps, heart-rate variability and CGM glucose. On a
CGM account that is over 100 000 objects per snapshot.
The crosstab only uses each day's sum and count after picking one
source and device per day, and the pick depends only on which sources
and devices are present. readSourceDayAggregates returns one row per
type, day, source and device, and the day map now carries each row's
reading count, so the day values are the ones the raw readings give.
The 30-day clinical glucose reads stay raw: time in range, SD, J-index
and the risk indices are functions of individual readings, and a
sensor's rate bounds the window below 45 000 rows.
Refs #1023
GET /api/auth/me gains ai: the state of every capability and the
provider block, resolved for the active record and masked to the
grant like moduleAccess. The authority comes from the switch the
payload already resolved, so a delegate reads not_permitted_for_record
and a guardian is limited to the operator's key.
useAiCapability is the web reader and fails closed until the payload
arrives. GET /api/feature-flags becomes a deprecated projection of the
operator switches with a Deprecation header and a link to its
successor.
loadAiCapabilityInputs reads the switches, module access, provider
presence and active consent kinds for one record in parallel, memoised
per request, and fails closed to check_failed. requireAiCapability,
getAiCapability and aiCapabilityForJob resolve for the request's
record, masked to the active grant like moduleAccess, or for a job's
record under the worker's authority.
apiHandler renders AiUnavailableError as one envelope with errorCode,
capability, reason and module. Existing codes stay: the operator
family keeps assistant.disabled.<switch>, modules keep module.disabled,
consent keeps consent.ai.required. New codes are ai.record.notPermitted,
ai.provider.none and ai.unavailable.
probeProviderChain describes the chain that would serve a record
without decrypting, building a client or refreshing a token: which
entries hold a credential, in chain order, whether each reads images,
the in-browser OCR opt-in, and where the credential comes from. It
applies the provider-work authority, appends the central Codex only
behind the opt-in, and falls back to the legacy provider only for an
empty chain. hasAnyConfiguredProvider and resolveProviderAvailability
now answer through it.
A pure resolver turns operator switches, module access, provider-work
authority, provider presence and consent receipts into one state per
capability, reporting the outermost reason. The document provider rank
moves to its own module so the resolver and the document routes read
the same order.
One frozen table names every AI capability with the operator switch,
owning modules, consent rule, modality and provider order that decide
it. The reason enum is listed in precedence order. No I/O, so the
resolver, gates, contract and guards can all depend on it.
Background insights can no longer use up the day before somebody asks a question; a fifth of the daily limit is held back for what a person does by hand.
The background ceiling applied only when the operator funded the chain, so on a
chain the operator does not pay for the automatic generators and the chat shared
one undivided ceiling and the first could still take all of it. That is the
configuration the recommended remedy for the original lockout produces — put the
personal Codex account first — so the guarantee was missing on exactly the setup
it was written for.
The ceiling is now expressed as what it is meant to protect. Background work may
run until the day's spend would leave less than a fixed reserve for whoever is
waiting on an answer, on both cost owners. A reserve and a quota look alike at
the boundary and mean opposite things in the middle: a quota stops useful
generation on a day nobody was ever going to interrupt, while a reserve costs it
nothing until the moment it would take the last of the day. An account
generating 1.45 million tokens of insights on its own plan keeps every one of
them and the chat still holds 400 000.
INTERACTIVE_RESERVE_SHARE replaces JOB_SURFACE_SHARE. The old name described a
quota over the jobs rather than a floor under the person waiting, and a name
that states the opposite of the intent is what put the wrong comparison in this
ledger twice already.
The fraction is a default an operator will want to change. The setting and its
documentation are separate work.
CLAUDE.md forbids planning shorthand in committed artefacts, and this branch
introduced 74 lines of it across source, tests, a schema field and two
migrations. Someone reading the migration SQL in a year has no way to resolve a
wave letter or a round number, and a comment that points at something outside
the repository names no reason at all.
Three passes. A version followed by a parenthetical keeps the version and loses
the parenthetical, because the version says since when something holds. A
sentence that leaned on the vocabulary to carry its meaning is rewritten to
state the reason directly: the arrival reservation now says it was refused over
money the operator never spent, rather than naming the defect a wave existed to
remove. Three test titles carried the marker in the name itself and are renamed
to what they pin.
Comments only. No behaviour, no renames beyond those three titles.
The owner split left an operator-funded chain with no ceiling at all. Its arm
compared operator_tokens, and that counter returns to roughly zero on every
reconcile the user's own plan settled: on the 2026-09-11 evidence the shared
key answered 500 and codex served, so the counter never grew, every job
reservation was admitted forever, and a runaway client loop could write an
unbounded row. A pre-existing concurrency contract caught it from the other
side, two reservations that must not both pass one cap both passing.
reserveBudget now refuses when the owner's counter is over its cap OR the day's
mixed total is over the abuse ceiling, and a background surface gets its share
of that ceiling too, so background spend is bounded on the total and not only
under a counter the operator's key never fills. The surface is a required
argument, so the type checker enumerated every reservation site.
The hop guard learns the same lesson: an operator-funded fallback hop is now
rationed at the calling surface's ceiling rather than the full operator cap. A
job admitted under the job share could previously spend the whole 200 000 once
its chain fell back onto the shared key, which is the original defect reached
by the fallback path.
The job share applies only where the operator pays for the chain. Multiplying
whichever ceiling the chain resolved halved background generation for a
self-hoster on a local model, who costs the operator nothing.
Refusals now name the ceiling they hit and whose it was. A refusal used to emit
the day's mixed total and nothing else, which on an operator refusal is not the
counter that tripped.
Thirty-one files carried internal build vocabulary and anchors pointing at
notes that are not in the repository: a reader of `checklist.ts` in a year
would find its rationale keyed to a research id they cannot open, inside a
comment that also names an internal build stage. Two test titles carried the
same anchors.
The version tags and the whole of the reasoning stay; only the vocabulary and
the dangling references go, with the few sentences that were built around an
anchor rewritten to stand on their own.
The done screen folded the `unknown` provider verdict into `unhealthy` and
painted "This instance's provider is not answering right now." `unknown` is
an empty `provider_health` table, which is the state of every instance until
its first AI call — so the most common path through the screen made a claim
about something nobody had measured. It now falls to the neutral panel, which
says nothing either way; the offer is withheld in both cases as before.
The projection itself drew the same unearned conclusion from staleness alone:
a quiet instance whose last success was two days old read `unhealthy` without
any failure behind it. A verdict now needs an observed one — a row whose last
outcome was a failure — and a merely old success reads `unknown`.
The automatic generators — status and reference cards, the briefing, the
arrival reaction, the nudge, document summaries and the index backfill —
reserved against the same ceiling as the chat. On the operator's account
that meant 150-330 ledger rows a day, 172 metric generations before
06:42Z, and a ceiling already spent when he opened the chat.
Every background surface now reserves through
`resolveDailyCapFor("job", chain)`: half the cost owner's ceiling, all
job surfaces together. The interactive routes keep the whole ceiling, so
a day of background work can no longer take the user's own away.
The reservation compared the day's TOTAL against whichever ceiling the
chain resolved, and the chain walker's hop guard did the same. On an
account whose background jobs run on the user's own ChatGPT plan, both
saw a total the operator had not paid for: the chat refused at 06:42Z
with 513 539 tokens against the 200 000 operator ceiling, and the
fallback hop to the shared key was closed on the same arithmetic.
An operator-funded chain is now measured against `operator_tokens` and a
user-plan chain keeps its abuse ceiling on the total; the hop guard reads
the operator share too. A day of turns the user's own plan paid for can
no longer lock anyone out of the operator's ceiling.
`coach_usage` counted one number per (user, day). A turn served by the
user's own ChatGPT plan after the operator's shared key answered 500 was
therefore charged to the operator-cost cap, and the operator's chat was
locked out every morning by background work he pays for himself.
The ledger gains `operator_tokens`: the share of `total_tokens` served by
an operator-funded provider. `reserveBudget` books a reservation to the
owner of the chain's primary and `reconcileSpend` settles it against the
provider that actually served, both inside the same statements as the
total so concurrent requests can never see one counter without the other.
Every call site names its cost owner and its serving hop; the argument is
required, so a new surface cannot inherit the old blind accounting.
The done screen promised working insights on the strength of a presence
read. A fresh account's own provider-health ledger is empty by
construction (the table is unique on user + provider), so the evidence has
to be an instance-wide fold. It folds across accounts and returns three
words: no counts, no timestamps, no account can be inferred from it.
Fails closed against the grain of the ledger it reads. There the fail-open
protects a generation already running; here the output is a promise made to
a person, so an empty table, a timeout or a read error is 'unknown', and
'unknown' buys nothing.
Both report paths dropped them: the raw path iterated the named contexts,
and the dense path skipped any row without a context outright, so a report
built for a meter that records no meal time carried an empty glucose panel
while the readings sat in the database. The FHIR export dropped them for
the third time, at the LOINC lookup — its own comment already said an
unspecified reading codes as 2339-0, but the table had no entry to hit.
The insights context breakdown names the bucket, and the Coach's glucose
block now resolves through the shared helper rather than its own copy.
Refs #943
A gateway can answer a failed generation with HTTP 200 and put the reason
in the body: a top-level `error` object and no `choices`, or
`choices[0].error` beside `finish_reason: "error"` when a generation dies
mid-flight. Both fell through to the empty-content throw, so a dead credit
balance, a data-policy block, a rate limit and a model that is down all
recorded the same hop: status 0, reason "returned empty content", no
excerpt. Nothing on the provider health card told them apart.
Raise the shape a non-2xx reply raises instead, with the gateway's own
message leading and its `error.code` as the status when that code is an
HTTP error status. The chain classifier, the health ledger and the card
then see 402 / 403 / 429 / 503 and treat each of them the way they treat
the same status on the wire. Message and excerpt are bounded and
key-redacted by the same helper the non-2xx path uses, so remote text is
no more trusted here than it is there.
Gated on the gateway tag: api.openai.com signals every failure with a
status, so the pinned tags parse exactly as before, and a genuine empty
reply stays the empty-content throw it was. The local-endpoint provider
had the same hole on its buffered path, its buffered-reply fallback and
its stream, and gets the same treatment. The body classifier learned the
two account-side refusals a gateway words in its own vocabulary rather
than the vendor's.
The operator's global provider always built an OpenAI client, so an admin
who saved an Anthropic key and https://api.anthropic.com/v1 (a host the
settings allowlist accepts) had every AI call fail: an OpenAI-shaped body
posted at Anthropic's Messages API, with nothing in the error pointing at
the mismatch.
Route by hostname instead. api.anthropic.com and any subdomain of
anthropic.com get the same AnthropicClient the per-user ANTHROPIC arm
builds, with the same Claude model default; every other host keeps the
OpenAI client, which is what an OpenAI-compatible endpoint needs. The
match parses the URL rather than searching the string, so a lookalike
host cannot claim the Anthropic wire, and an unparseable value keeps the
OpenAI path it had before.
The chain tag stays admin-openai: it names the slot, not the wire, and it
is persisted in user chains and pinned by the OpenAPI enum. The wire is
already visible in the instance type the wide events read.
The Anthropic vision check listed the Claude 3 and 4 families by prefix,
so a key pinned to claude-sonnet-5, claude-opus-5 or a Fable model was
told the provider cannot read images, and lab and document OCR refused
the upload even though the model handles image and PDF blocks.
Replace the per-release prefix list with a rule: any claude-<tier>-<major>
slug from the 4 family upward, plus every claude-fable-* model, counts as
vision-capable. Claude 3 handling is unchanged. The Settings preset list
gains claude-sonnet-5 and claude-opus-5 alongside the existing entries.
The Anthropic client steered JSON replies by appending an assistant turn
containing an opening brace and gluing the brace back onto the reply.
Current Claude models (the 4.6 / 4.7 / 4.8 family, Opus 5, Sonnet 5,
Fable 5 and 5.1) reject an assistant prefill with HTTP 400, so every
JSON surface on the provider (briefing, status cards, document and lab
extraction) failed for a self-hoster on a current model while plain chat
kept working.
Drop the prefill on every model. The request now carries a plain
instruction on the last user turn to reply with a single JSON object and
nothing else, and the reply is narrowed to the outermost object body via
a shared extractJsonObject helper, so a fenced or prefixed answer still
parses. The insights stripJsonFences helper delegates to the same
function rather than keeping a second copy. The tool-use path is
unchanged.
Reported independently by @sweidinger.
The global admin provider's base URL is operator-typed in the admin
settings, but the client pinned requirePublicHost unconditionally for
its tag — so an operator pointing it at a private OpenAI-compatible
proxy (an OAuth sidecar on the instance's own Docker network) got a
safeFetch refusal on every call, indistinguishable from an outage. The
admin-key tag now consults the same ALLOW_LOCAL_AI_PRIVATE_HOSTS
policy as the user-configured gateway: a public URL never consults the
allowlist, a private host without an entry is refused exactly as
before, and both ends of the grant are the operator's own. The codex
tag stays hard-pinned; its base URL is a repository constant. Every
non-operator admin-key construction pins the public api.openai.com
constant, so the personal-key posture is unchanged.
The test asserting the old blanket pin is split into the three cases
that actually differ: allowlisted private host passes, unlisted
private host still refuses, codex refuses even a blanket allowlist.
Registering `ko` in the `Locale` union left seven locale-keyed tables
without an entry, so the build could not type-check and the safety
screens, the citation-coverage grader and the reference-range parser
read nothing at all for a Korean reader.
Three failures were not missing keys but wrong assumptions:
- The shared fold decomposes with NFD and strips U+0300-U+036F. Latin
accents become plain letters, but Hangul decomposes into jamo, which
are not in that range and survived — so folded Korean text was a jamo
sequence no precomposed table entry could ever match. It recomposes to
NFC now; every Latin, Polish and German fold is byte-identical.
- `normaliseLabKey` kept `[a-z0-9]`, which reduced every Korean analyte
name to the empty string, so the derived index held none of them and a
Korean cholesterol result reached a FHIR export uncoded.
- The reference-range parser had the hyphen and the dashes but not the
tilde a Korean lab actually prints (`3.5~5.0`).
The Korean safety banks carry no `\b` next to a Hangul token — `\b` is
an ASCII `\w` boundary and Hangul is not `\w`, so `/\b증량/` matches
nothing — and the clause order is the mirror of the Latin banks, since
Korean closes with the verb. The dose patterns require a target-marking
particle for the same reason the Latin banks require "to"/"by": without
it the permitted restatement "7.5mg을 복용 중이고, 체중은 내려가고
있어요" tripped the lowering pattern.
The Coach and Insights prompts route `ko` down the reviewed English body
plus a Korean reply directive; there is no `safety-contracts.ko.yaml`, so
the native builder has no Korean ground rules to compose from.
Register `ko` in the locale list, language names, and labels, and ship
messages/ko.json so the whole UI — dashboard, settings, share views,
clinician PDF, notifications, and AI prompts — renders in Korean.
Extend the locale-parity, English-leak, plural-completeness, and copy
guard tests to cover the new locale, and update README/CONTRIBUTING to
list Korean among the supported languages.
CadenceKind and WeekdayToken were imported only to be re-exported. The
re-export went in the previous commit but the import survived, because
the doc comment above it still named both types at the time.
Eighteen modules carried a convenience re-export of a symbol defined
elsewhere. In each case the callers the comment describes now import the
concrete module directly, so the line had no importer left. Where the
whole re-export block went, the comment explaining it went with it —
doctor-report-pdf-core, trend-card and medication-extract-prompt each
claimed to be a single import surface that nothing was using.
Definitions are untouched; imports that existed only to feed a removed
re-export go too.
Same shape as the derived and MCP barrels: the line in the barrel had no
importer because the consumer imports the concrete module. Affected:
cycle (15), coach tools (9), record-settings (8), links (4), data-list,
workout-detail and analytics/score (1 each).
analytics/score also carried CompositeValue on an explicit export line
above an export * of the same module, so the name stays exported either
way.
Definitions are untouched.
The grounding tokenizer read English unit words and English or German
month names, while the reply is written in the reader's language. Two
silent failures came out of that, pointing in opposite directions.
A missing unit loosens the guard. "7,4 Stunden", "45 minuti" and "8000
pasos" came back unkinded, and an unkinded token clears the kind gate
against every ledger entry, so a fabricated night's sleep could ground
itself on a glucose reading that happened to share its magnitude. Kind
scoping is the thing that stops a weight grounding a blood-pressure claim;
for five languages it was not running.
A missing month corrupts correct prose. "Le 21 juillet" left the day
untyped, the plain-magnitude pass graded it, nothing matched, and the
soft-strip replaced it — the reader was shown "Le […] juillet". Spanish
and Italian also put a particle between day and month, so the pass admits
one.
The unit vocabulary had been written out three times and the copies had
already drifted; only one of them knew about "%". One list now feeds the
classifier, the range pass and the measurement pass. Polish months are
listed in the genitive, which is what a date uses and also the only form a
trailing word boundary can match, since that boundary is defined over
ASCII and "ń" is not part of it.
The gate is "looks off-topic AND does not look health", and both halves
were English/German word lists. Two things followed. An off-topic ask in
French, Spanish, Italian or Polish was never recognised. And a genuine
health question in those languages that happened to brush an English deny
word was refused outright: "Ho una serie di misurazioni strane" trips
"serie", and nothing on the allow side knew "misurazioni", so a question
about the reader's own data came back as a refusal.
The allow side is no longer transcribed. Every metric label key and the
health nav labels already name these concepts in all six bundles, and
those labels are the nouns people type, so the vocabulary is read from
`messages/<locale>.json` and a seventh language is understood the day its
bundle lands. Matching is whole-word over a shared fold applied to both
sides. A subtractive list keeps label scaffolding out; French "bien-être"
alone would have contributed "bien" and "etre" and marked nearly every
French sentence as health.
The deny side stays hand-written — these are not app labels — and its new
arms match the folded message. A pattern anchored with \b cannot match a
word starting with "ż", because neither the preceding space nor the letter
is an ASCII word character, so there is no boundary between them.
The prompt-injection bank is untouched. It reads the person's own message
and the foreign-text path is fenced elsewhere; this is defence in depth,
not an open door.
The normative-claim bank held English and German words. Insight
recommendations are generated in the reader's locale, so for French,
Spanish, Italian and Polish every recommendation scored non-normative:
the annotation reported zero normative claims and zero uncited ones, and
the admin quality dashboard read that as full coverage rather than as no
measurement at all. A check that cannot fail is indistinguishable from a
check that passes.
The bank now has one arm per shipped locale and the caller passes the
generation locale; the reader's arm and the English one both run, because
a fallback provider answers a non-English reader in English. Matching
folds both sides through the shared label fold and anchors the leading
word boundary, so inflections still hit while the everyday Italian
"soprattutto" no longer counts as a comparative.
These words are the model's prose, not app labels, so they are not in the
message bundles and cannot be derived from them. The structural test
refuses an empty arm instead: a seventh language fails the suite rather
than quietly scoring nothing. An empty arm also used to compile to an
alternation that matched every string, which would have flipped the same
blindness to the opposite extreme.
Marc-André Bombeck