This repository has been archived on 2026-11-08. You can view files and clone it. You cannot open issues or pull requests or push a commit.
Files
carl/docs/CLIENT-INTEGRATION.md
james.maesandClaude Opus 5.5 654a33a778 docs: reconcile current state, counts and evidence paths (#11, #12)
Add one dated current-state section (566 tests at 9bd782e against
foundation d0a11bf, owner chat failing per #34), mark closed preview URLs
as not live, align data-policy statements with AGENTS.md, label older
counts historical, and replace local machine paths in the handoff and
session log with neutral descriptions.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-02 13:07:49 -05:00

30 KiB
Raw Permalink Blame History

Carl family client integration

Carl uses the foundation /api/agent/v1 protocol. Foundation stable 0.4.1 is remotely qualified at source b23f89a44647ddc8fd0f3f0894b159ae823638b7, with signed publication and remote-consumer evidence from delivery run 36703216841. This release corrects the provider-continuation/completion defects found in 0.4.0. Follow the matching foundation docs/CLIENT-API.md and OpenAPI contract for tokens, conversations, polling, limits and native iPhone login.

Historical (September 30; see the current count): the development parent was 0.5.0-SNAPSHOT, with qualified Next UI1.0.0-RC.7 and the owner-approved exact QQQ4.1 development closure. That source passed 377 tests, zero failures/errors/skips,59 matching packaged browser checks and fresh source/artifact scans; see current qualification. Foundation PR15 merged at 0b6f437; main0.5.0-SNAPSHOT publication is independently qualified in run36811134831; see signed development evidence. Immutable0.4.1 remains the signed remote release baseline in config/release/foundation.json; recovered local artifacts are not fresh credentialed resolution. Carl hosted delivery and live model/identity/iPhone qualification remain blocked on their separate prerequisites.

Set kof22.agent.client-api.enabled=true, .issuer, and .audience only after selecting and qualifying the identity provider. Set the public HTTPS origin through kof22.agent.qqq.public-origin. The native API audience must differ from the QQQ operator audience. The foundation validates dedicated access tokens; prompt text, JSON caller fields and QQQ cookies cannot substitute for them.

An operator explicitly provisions carl_identity(issuer,subject,member_id) and current carl_member/domain permissions. There is no default family member or mapping from a display name. CarlFamilyAccess resolves the exact verified issuer and subject to the current durable household permission epoch. Do not reuse member IDs. Permission changes invalidate old conversations and stored artifacts conservatively; clients clear protected caches and start a fresh conversation. Existing reports may need regeneration even when only another household permission changed. V7 intentionally denies older artifacts without an epoch snapshot.

Ordinary private chat exposes only Carl's six caller-aware read tools (bills, finances, records, category budget variance, explicit presentation preferences and possible appointment windows). Shared conversations have no private read tools. Explicit reports, drafts, financial comparisons and plan changes use a separate user workflow intent with the server's immutable participant set, never model-supplied identities or audience.

Workflow kind input fields Result
report from, through: ISO dates, at most two years Persisted source-linked household facts for the audience's intersecting access
debt-comparison asOf: ISO date; currency: explicit currency; monthlyBudget: positive exact decimal string; horizonMonths: 1–600; budgetEvidence: explicit assumption/evidence Persisted comparison of authorized debt inputs; supplied budget is not proof of affordability
purchase-assessment cashPlan: positive permitted record ID; purchaseDate: ISO date; allInPrice: exact decimal string; purpose: bounded text; allInCostsKnown: boolean Conditional cash budget from the saved dated forecast, selected reserves and goals; financing offers require separate qualification
offer-comparison offerIds: one to ten distinct permitted record IDs; asOf: ISO date Scoped equal-principal/currency financing costs, residuals and dated source limitations; no offer acceptance
rental-report propertyIds: one to 100 distinct permitted records; from, through: ISO dates spanning at most 366 days; asOf: explicit date Source-linked property cash figures, explicit ownership and partial source coverage
tax-packet propertyIds: one to 100 distinct permitted records; taxYear: explicit 1900–2200; asOf: ISO date, interpreted at UTC midnight Dated preparation evidence and missing-document checklist; tax calculation remains unqualified
draft workId: positive integer; purpose: FOLLOW_UP, QUOTE_REQUEST, SCHEDULING_INQUIRY or SERVICE_QUESTION Persisted local vendor draft, explicitly not sent
plan-create sourceArtifact: current financial comparison ID; title, reason: bounded text New draft plan; comparison audience must exactly match immutable conversation participants
plan-step plan, expectedVersion, canonical step UUID, title, assignee member ID, ISO due, location, nullable canonical dependency UUID, reason Versioned task; assignee must have current plan access; task revisions require renewed agreement
plan-agree plan, expectedVersion, reason Explicit agreement after current complete assumptions and assigned tasks are checked
plan-check-in plan, expectedVersion, canonical step UUID, status: TODO/REPORTED_COMPLETE/VERIFIED_COMPLETE/BLOCKED, note, nullable evidence record ID Reported work remains distinct from human evidence-reviewed completion; every recipient must see typed evidence and its dependencies
plan-rebase plan, expectedVersion, sourceArtifact, reason Replaces assumptions with a current comparison for the same audience; preserves prior versions and tasks, returns to draft for review
plan-export plan Protected PDF bytes as contentBase64, mediaType, filename, exportedPlanVersion, exportedAt; plan is the snapshot at workflow completion and may be newer than the dated PDF

Use PUT /api/agent/v1/conversations/{conversationId}/workflows/{kind}/{requestId} with {"input":{...}}, then GET the same URL. Generate one UUID per logical request and retain it across retries. Reusing it for different input, requester, conversation or audience fails. PENDING means work is still running; UNKNOWN requires reconciliation and must never trigger a new automatic request. Results return protected artifact content directly, without public download URLs. Plan mutation and its saved result commit atomically; retries return the original result after current audience/source checks. Exact keys are required, including explicit nulls for nullable task fields. Plan inputs never accept caller identity or an audience override.

Carl owns two workflow workers and a sixteen-request queue. Nonconversational workflows have a thirty-second domain execution budget. Conversational generation uses the configured model turn timeout, capped at 270 seconds, plus a bounded thirty-second SQL completion window; planning and narration share the same model budget. SQL execution is bounded by the remaining budget (rounded up to one second) and normal query bounds. Interrupted startup work becomes UNKNOWN; it is not replayed. Earlier controlled SDK/HTTP acceptance used identified development artifacts. The integrated report-wave source also passed the 319-test stable-parent gate; final product qualification remains pending and hosted/live model/identity/iPhone qualification remains BLOCKED.

Secret names only: KOF22_AGENT_ANTHROPIC_API_KEY, KOF22_AGENT_DB_PASSWORD, KOF22_AGENT_QQQ_DB_PASSWORD, KOF22_AGENT_QQQ_PASSWORD. Identity-provider secrets belong to the selected deployment's protected configuration; no client secret is embedded in an iPhone binary. Synology credentials and collection identifiers are not yet configured or connected.

Portfolio summaries and protected details

portfolio-comparison accepts exactly accounts (1–100 unique positive IDs), moves (0–20 unique reviewed allocation IDs), asOf, currency, decimal-string monthlyBudget, integer horizonMonths (1–600), rollover (AVALANCHE, SNOWBALL, or MINIMUM_ONLY), and budgetEvidence. The service additionally limits the combined projection to 25,000 debt-period rows. The response contains a protected artifact ID and bounded summary, including currency, estimated totals, payoff date, comparison qualifications, and counts. PARTIAL can mean a strategy cannot pay off within the horizon. Available credit and an assumed monthly budget do not establish affordability.

To retrieve the preserved details, start a portfolio-detail workflow with exactly artifact, section, strategy, offset, and limit. Sections are assumptions, sources, gaps, months, payoffDates, and cashDifferences. Supply strategy: null for the first three; other sections take CURRENT_PAYMENT, MINIMUM_ONLY, AVALANCHE, SNOWBALL, or USER_DIRECTED (cash differences compare a candidate against current, so select a candidate). Pages have 1–10 items and a 128 KiB bound, with totalItems and nextOffset; use a smaller page if needed. Immutable saved schedules are never silently truncated. A missing strategy schedule requires inspecting its gaps, not treating it as zero debt. The configured conversation audience must retain permission to every source on every retrieval. Native administration also exposes the full protected artifact.

Purchase payment choices

purchase-options accepts exactly cashPlan, purchaseDate, allInPrice (decimal string), purpose, allInCostsKnown, card, and offers (zero to five distinct permitted PURCHASE_FINANCE offer IDs). card must be explicit null or exactly {account, payoffDate, graceConfirmed, balanceReviewed, evidence}; payoffDate may be null. Grace claims require coherent current dated zero-balance evidence and reject contradictory newer imported balances. Ordinary financing with a first payment beyond one modeled month remains unqualified because the initial accrual/grace period is not implemented. Cash, full-pay card and supported actual store-financing terms share the same dated reserve/cash constraints; credit limits are never an affordability input.

The protected result includes currency, all-in price, conditional maximum cash budget, cost/liquidity qualifications and per-option payment counts. Retrieve immutable repayment pages through purchase-detail with exactly artifact, option, offset and limit (1–24 payments). Each page includes currency, total and nextOffset. Unsupported terms or incomplete forecast coverage return unknown qualification, not fabricated savings. No order, application or payment occurs.

Shared calendar and reminder intents

calendar-plan accepts exactly plan, canonical step UUID, positive expectedVersion, collection (events or reminders), operation (PUBLISH, RETIRE, RECONCILE, SYNCHRONIZE, STATUS) and priorRequest. priorRequest is a canonical UUID only for RECONCILE and must otherwise be null. Only operator-configured fixed collections are eligible. The existing plan audience must exactly match the conversation participants; the separately configured shared-calendar audience must also pass current source checks. STATUS returns at most 51 recent local operations/observations and does not refresh the provider. Unknown writes require read-only reconciliation of the original operation ID.

SYNCHRONIZE may retain a typed remote reminder observation. It does not change task completion. reminder-review accepts exactly plan, observation, expectedVersion, decision (ACCEPT_REPORTED_COMPLETE or DISMISS) and bounded note. Current source permissions, plan version and latest synchronized resource identity/hash are rechecked. Acceptance records human-reviewed REPORTED_COMPLETE, never verified financial execution. The provider does not prove which family member edited a shared checkbox. Conflicting, inaccessible or changed observations require renewed review. Calendar endpoints, arbitrary content, caller IDs and audience overrides are never input fields.

Tax source packets and bounded saved-fact navigation

tax-planning-packet accepts exactly taxYear (1900–2200 or explicit null), asOf (nonfuture RFC3339 instant), properties (1–20 unique permitted IDs), alternatives (0–40), and references (0–100). It combines supplied property facts, conditional human-proposed structures and dated primary-source reference identities. Reference applicability is evaluated as of the requested instant; supplied domain facts remain current records, not reconstructed history. No tax calculation, entity recommendation or professional approval is inferred. The result is a protected artifact ID and bounded summary, with full data available through section pages.

artifact-section accepts exactly artifact, path (RFC6901 JSON pointer, root is ""), nonnegative offset, and positive limit. The root contains facts, narrative and limitations. Object/array pages allow at most 25 entries; each descriptor includes its exact next path and type. Large nested values are navigated explicitly, not included unboundedly. Text leaves allow at most 4096 Unicode code points per page; totalItems, offsetUnit and nextOffset allow complete reconstruction without cutting surrogate pairs. Decimal values remain exact. Missing paths and oversized pages fail explicitly. Every page rechecks the original conversation participants against all current source permissions; this is not a public artifact URL.

The carl_read_availability ordinary read capability takes exact from/through instants and integer minimumMinutes (1–1440); requested interval is at most seven days. It merges only common authorized busy intervals, excludes canceled/transparent events, returns at most 100 candidate windows and includes coverage/freshness limitations. No private event prose enters its result. Incomplete/private calendar coverage can conflict with every suggestion. Native administration also provides household-zone date/time entry and rejects ambiguous/skipped DST endpoints rather than guessing which occurrence was intended.

Focused reports and operational recovery

Use explicit focused-report input {focus: BILLS|CALENDAR|VENDORS, from, through} or bill-period-comparison input {beforeFrom, beforeThrough, afterFrom, afterThrough} with ISO dates. Results contain protected artifact summaries; artifact-section retrieves complete facts/narrative in bounded pages. Comparisons require comparable source coverage and preserve hypotheses separately from observed differences.

report-status accepts only {reportRequest: UUID}. report-reconciliation accepts {reportRequest: UUID, reason: string} and uses the workflow request UUID as its new reconciliation identity. Both require the original actor and exact current conversation audience. Reconciliation can recover a provably committed artifact or fence an abandoned generation; it never replays narration. Retrieval rechecks current authority. FAILED means a new explicit generation needs a new request UUID; UNKNOWN needs review, not blind repetition.

Selected balance sheets and rental stress

balance-sheet accepts {asOf: YYYY-MM-DD, maximumAgeDays: 0..3660, accounts: [IDs], estimatedProperties: [IDs], linkedProperties: [IDs]}. Select at least one account/property, at most 100 accounts and 100 unique properties; a property may occur in only one valuation list. These are selected permitted balances, not proof of complete household coverage. Linked asset/debt identities are counted once.

rental-stress-report accepts only {scenario: ID} for an existing human-reviewed native scenario. It replays the authorized historical rental interval with explicit vacancy/repair/simple-interest assumptions; it does not predict mortgage payments, taxes or future affordability. Both return a protected summary with artifact-section navigation. Creating scenario assumptions is currently a separate explicit native human workflow.

Protected report downloads

report-export accepts {artifact: ID, format: PDF|TEXT} and returns a small immutable export manifest: export request UUID, artifact ID, media type, total byte count and SHA-256. report-download accepts {exportRequest: UUID, format: PDF|TEXT, offset: integer, limit: integer}. The limit is1–65,536bytes; follow nextOffset until null, concatenate decoded contentBase64 bytes in order and verify the manifest length/hash. Text uses UTF-8: decode after byte reassembly. Maximum stored export is4MB; oversized reports fail explicitly. The app may offer copy/save after complete reassembly; no public URL or outbound send is created.

Every page rechecks export ownership, current artifact/source permissions and every immutable conversation recipient. Source or permission changes can invalidate an earlier manifest. Do not treat downloaded client caches as automatically revoked; clear private cached results when /me.permissionRevision changes and follow the approved device-storage policy. No actor or audience fields are accepted.

Expected and observed plan effects

plan-expectation accepts plan, current version, step UUID, kind (CASH_PAYMENT, TRANSFER, PRINCIPAL_REDUCTION), account, exact decimal-string amount, date-only from/through, and reason. The verified conversation participants must exactly match the plan audience. The result is a protected immutable expectation ID; no payment or task completion occurs.

plan-effect-comparison accepts that expectation, an explicitly selected array of at most100 transaction IDs, and evidence. It returns a bounded artifact summary; artifact-section retrieves protected details. The result preserves expected/observed/difference amounts, currency, observation date, source selection and MATCH/DIFFERENCE/UNDETERMINED. Missing observations remain unknown. Current permission checks apply to creation and every retrieval, and results never automatically verify financial execution or complete a task.

Conversational reports, purchases and financial plans

Use the existing workflow PUT/GET protocol with kind conversation. Input accepts only message (nonblank, at most 16,000 characters and the configured input limit) and optional replyTo (a canonical UUID of a prior conversational workflow request). Retain the logical request UUID across retries and retain the returned request chain for follow-ups. Do not send caller, audience, operation, SQL or provider-target fields. This domain workflow history is separate from ordinary /turns history; ordinary read questions continue to use the existing read-tool interface.

Carl interprets the human request using the configured model, validates a closed proposal, then invokes its own deterministic services under the verified caller and immutable conversation audience. Report scopes are HOUSEHOLD, BILLS, CALENDAR and VENDORS with explicit date intervals; draft purposes match the existing draft workflow. The protected result payload exposes these kind values:

Result kind Meaning and details
CLARIFICATION Missing or unsupported inputs require a focused human answer; follow up with replyTo. No artifact or plan is created.
BOUNDARY The requested external action is unavailable. No send, booking, purchase, application, payment or commitment occurs.
CAPACITY Configured inference capacity was unavailable; status is FAILED and no artifact was generated.
ARTIFACT A saved report or local vendor draft, with protected artifact-section navigation.
PURCHASE Conditional cash/card/actual-offer comparison, with a bounded purchase summary and purchase-detail repayment pages.
PURCHASE_BUDGET Conditional cash ceiling before a price is known. allInPrice and preferredOption remain null; payment comparison remains PARTIAL.
FINANCIAL_COMPARISON Saved selected-debt strategy comparison, bounded portfolio totals and portfolio-detail schedules/sources; no plan is created.
FINANCIAL_PLAN The comparison plus an explicitly requested DRAFT plan v1. createdPlanId and createdPlanVersion identify the original creation; plan is reloaded as the currently authorized snapshot.
RENTAL_REPORT Selected-property cash/rent figures by currency, with source coverage limitations.
RENTAL_STRESS Replay of one named saved scenario; its reviewed assumptions are not changed.
TAX_PACKET Supplied property evidence, alternatives and reference identities; tax calculations and structure recommendations remain unqualified.
BALANCE_SHEET Selected current permitted valuations/balances, separated by currency, with explicit age and valuation choices.
EXPENSE_FORECAST Selected scheduled/actual expenses and separate reserve earmarks, with the saved calculation currency.
INVESTMENT_SCENARIO Educational projection from explicit assumptions and saved goal context; no inferred readiness, suitability or affordable contribution.

For a purchase or budget question, identify the permitted saved cash plan, explicit currency and ISO purchase date. A known price must come from the human's quoted evidence; taxes/delivery/fees must be explicitly included to assert all-in costs. A budget-only question does not need a fabricated price. Existing cards remain unqualified for grace/full-payment treatment without reviewed evidence; model selection cannot assert those terms. Only actual permitted purchase-financing offers can enter the comparison. Missing offers or incomplete cash/reserve/source evidence remain explicit limitations, and available credit never increases the cash budget.

For a financial comparison, supply the named permitted debt accounts, as-of ISO date, currency, total monthly debt-service budget, horizon in months and rollover assumption (AVALANCHE, SNOWBALL or MINIMUM_ONLY). The budget is a human-supplied scenario assumption, not verified available cash or an extra payment inferred from income. Carl selects at most twenty debts and five existing reviewed moves, subject to the service's projection bounds. It cannot invent debt terms, offers or allocations. The current evidence guard requires an explicit currency followed by the complete canonical decimal amount, and a complete month-count token; unsupported grouped/signed representations require clarification rather than matching a numeric suffix. Explicitly request and name a draft financial plan to create one. Creation requires current FINANCE management permission and the exact comparison audience; it does not agree a plan, add tasks or perform financial actions. Subsequent agreement, task changes and rebase remain separate explicit typed workflows.

The six report kinds above return PARTIAL conditional results with a calculation object bounded to 32 KiB. Larger summaries return summaryState: DETAIL_REQUIRED and detailWorkflow: artifact-section; retrieve the complete authorized saved facts through those pages. Their narrationState is NOT_REQUESTED, not a narration failure. Missing figures are not zero, and selected sources do not establish complete household coverage.

Name each permitted source explicitly; ambiguous names require clarification. State date roles as from YYYY-MM-DD, through YYYY-MM-DD and as of YYYY-MM-DD; tax packets also require a tax year, and balance sheets require a maximum age in days and a property valuation choice. Expense forecasts require an explicit currency. Investment examples require a named saved goal, starting month, horizon in months, currency and separately labeled initial capital, monthly contribution, initial fee and monthly fee. Rates require explicit decimal-fraction units—for example, monthly return decimal fraction 0.01 and asset fee rate decimal fraction 0.00. Missing units, percentages, basis points or contradictory units trigger clarification; Carl does not convert them implicitly or fill missing assumptions with zero. Tax as-of dates mean midnight UTC for reference applicability; they do not reconstruct historical records or authorize fetching a source URL.

The reviewed and integrated conversational lifecycle adds PLAN_CREATE, PLAN_STEP, PLAN_AGREE, PLAN_CHECK_IN, PLAN_REBASE, PLAN_PROGRESS, PLAN_EXPECTATION, PLAN_EFFECTS, PLAN_CALENDAR, REMINDER_REVIEW, GOAL_PRIORITY and GOAL_TRADEOFF. Independent source review is clear and its frozen clean gate passed347 tests; the original-source gate is running. Final packaged/browser/security and live qualification remain separate. These use the same conversation PUT route and message/replyTo input; clients never send the internal proposal operation or authority fields.

Name the plan and its current version in each human request. Creation and replanning explicitly select a saved comparison. For example, “Create a draft plan titled Debt freedom from comparison 123”; “Add a step to plan Debt freedom version 1: Review payment checklist; assign to Alice; due 2026-10-01; location Home”; and “I agree to plan Debt freedom version 2”. New task and child-request UUIDs are server derived. Agreement retains the existing incomplete-source, stale-source and assigned-step guards. Editing or rebasing returns the plan to DRAFT for renewed review. Negated, hypothetical, conditional, quoted third-party and earlier requests cannot authorize a new mutation. Quoted titles, assignees and locations are ordinary arguments; a negative rationale such as “because I don’t want more debt” does not negate an explicit current action. Missing or ambiguous selections yield CLARIFICATION with a closed, non-executable proposed operation and a request for explicit confirmation.

Check-ins name the task and current version and preserve human REPORTED_COMPLETE, TODO or BLOCKED. Model interpretation and a shared checkbox never produce VERIFIED_COMPLETE. Progress reloads the currently authorized plan and steps. Explicitly request PDF to receive a protected, bounded export with application/pdf, contentBase64, exportedPlanVersion and exportedAt. The PDF remains a dated immutable version while the returned plan is current; protect client caches as described above. PDF bytes are excluded from subsequent model history.

Expected effects require a selected current agreed plan and task, account, exact currency/amount, effect kind and explicit from/through dates. Observed-effect comparisons name the saved expectation and every selected transaction ID. The resulting artifact retains unknown coverage and distinguishes amount matches from verified execution; it does not complete tasks. Configured calendar requests explicitly select events or reminders, the task and current version, and PUBLISH, RETIRE, RECONCILE, SYNCHRONIZE or STATUS. Provider endpoints and credentials are never model/client fields. Reconciliation names the original server-derived operation ID; synchronization can retain an observation awaiting human review. Review explicitly names the observation and chooses accept reported completion or dismiss. Publication, reads and PDF rendering occur outside workflow completion locks; local plan, expectation, review and goal mutations commit with the pending workflow result in one transaction.

Goal priority requests name a currently permitted saved goal, its exact revision and a new priority: “Set priority for goal Debt freedom revision 1 to priority 2. Human priority choice.” The output includes the immutable priorityRevision receipt and freshly authorized current goal. Changing priority never selects an investment stage. Educational debt/investment tradeoffs explicitly select debt comparison and investment scenario IDs, a saved debt strategy and every goal, plus human evidence. Results keep known debt costs separate from uncertain growth, retain UNQUALIFIED_ASSUMPTION/EDUCATIONAL_ONLY qualifications, and label incomplete or incompatible source assumptions INCOMPARABLE. Protected artifact-section pages and native QQQ expose the same facts. Existing explicitly typed workflows retain their contracts; controlled fixture success is not live-provider or full-product qualification.

Shared results use the intersection of participants' permissions. Every history/result/detail retrieval rechecks the current permission epoch and source access; plan retrieval also rechecks its exact audience. Source changes mark saved comparisons and plans stale instead of silently recalculating or rebasing them. Preserve the distinction between the original created plan version and its current snapshot. Clear protected caches when permission revision changes; old history cannot restore revoked access. Full facts, narrative and limitations remain available through bounded artifact-section pages, with specialized portfolio/purchase pages described above. Do not infer missing amounts from absent summary fields.

PENDING remains in progress; UNKNOWN must not be automatically replayed. Repeating the same logical request does not repeat model generation or create another plan. Comparison persistence precedes draft creation; plan creation and workflow completion commit atomically, and interruptions retain a truthful unresolved state. Truncated planning creates no artifact; narration failure preserves deterministic facts with failed narration and a PARTIAL result. Financial comparisons remain PARTIAL because cash affordability is unqualified even when their requested projections and narration succeed. Native QQQ exposes the same saved artifacts and versioned plans.

Planning and narration share context/output/deadline controls and recorded usage. Conversational configuration explicitly supports turn timeouts up to 270 seconds, with a further bounded 30-second SQL completion window for truthful partial results. Model inference shares kof22.agent.limits.max-concurrent-turns across ordinary chat and workflows; the existing two-worker/sixteen-request workflow admission remains separately bounded. The reviewed controlled SDK/HTTP, PostgreSQL and native QQQ scenarios establish the implemented request paths; they do not establish live model interpretation quality. The current377-test consumer gate and59 matching packaged browser checks pass. Hosted delivery and live identity/model/iPhone qualification remain BLOCKED as described above; controlled SDK responses and offline UI do not establish live-model interpretation.