This repository has been archived on 2026-11-08.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/ACCEPTANCE.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

21 KiB
Raw Permalink Blame History

Acceptance evidence

Current state (2026-10-02): source 9bd782e passes 566 tests against foundation main d0a11bf (see current state); the earlier failing 566-test run is superseded. The owner's local runtime still runs the older 563-test package. No preview URL below is live. Counts, packages and evidence files below are dated, point-in-time records; many cite logs in the private owner-held bundle, not in Git. Paused for owner review; see the takeover handoff.

Owner ordinary chat workflow: FAIL (Carl#34; fixes in local branches pending push). The real monthly-spending question returned UNKNOWN after nine successful reads. Check Status is separately under investigation. Prior specific backend cases do not establish that the owner receives replies through the native chat window. Bounded aggregation and complete UI-to-answer acceptance are required before calling chat ready.

October 2 local scope (historical): the sealed563-test package runs with retained FlywayV62/coreV8, actual owner OIDC, ten protected historical originals, independently field-checked selected/all23,610 transaction CSVs, a source-linked readiness draft, explicit dashboard defaults and eight actual browser checks. The prior559-package exact “Look at…” document answer passes independent current-access/read-audit/usage/citation checks. The optional “now” prefix repair passes regression tests in563; a fresh document answer is COMPLETE but its complex follow-up remains UNKNOWN. At the current32-call bound all24 actual reads succeed; a separate read-only diagnostic reproduces ten-iteration exhaustion. Completion behavior remains under repair, and every failed request is preserved. Earlier two source-grounded document/plan conversations retain their553-package provenance. Exact sanitized evidence separates these results and the outstanding date-control repair. Whole-product financial acceptance remains incomplete; no Liquibase runtime/release/image/production qualification is implied.

Current real-family application acceptance: NOT COMPLETE. Actual private imports, verified local owner identity, live Anthropic and real search/ESB services are working in the local QA environment. Owner-confirmed goals and a source-linked readiness DRAFT are saved; protected native records/PDF pass with anonymous denial. This does not supply missing current debt terms, cash/reserve/budget, account identity/ownership or rental facts. Calendar/reminders and Authentik are deferred by the owner. Protected historical-document ingestion/retrieval (Carl#31) and governed table CSV exports pass actual private acceptance in the current563-test/V62 package; other processes retain their prerequisite and acceptance limits. No private data may enter this public repository or CI artifacts.

Running-package repair qualification: PASS for the stated local scope. Full Java21 clean verify passes490 tests with zero failures/errors/skips and unchanged consumer quality gates. The first direct-route full run found four negative-intent compatibility failures; the corrected run passes all28 HTTP conversation cases. The new sealed package runs with actual retained records and verified identity. Live inventory and saved-plan answers are independently checked against permitted SQL facts, supporting report, goals, verified read-only audit and positive provider usage. Actual bulk preview leaves private transactions unchanged; protected records and fresh native PDF retrieval pass with anonymous denial; search and broker are READY. The live transaction answer independently matches exact permission-filtered SQL count/signed USD total and permitted source citations. All three actual conversations pass current access, caller-bound read audit, positive provider usage and ANSWER-only persistence without changing existing plans/reports/goals. These checks do not establish every process, historical-document ingestion or full financial priming. SESSION-STATE.md identifies the exact package and limitations.

Sanitized current qualification records the package, gates, actual checks and remaining limits; detailed private evidence stays outside Git.

Source-account correction: consumer build PASS. Java21 mvn spotless:apply clean verify with the separately qualified dependency cache passes428 tests with zero failures/errors/skips and unchanged quality gates. V60 is applied to retained private PostgreSQL through the supported packaged migration helper. Actual source values/dates/fields/provenance and protected reader agreement are checked privately. Unknown accounts remain excluded from qualified cash/spending/balance totals; no fake zero graph is rendered. This is headless source-load evidence with an authenticated technical importer, not real-family browser/model acceptance. Detailed private evidence must remain outside this repository.

Historical controlled RC8 money presentation: PASS. Whole-product/live qualification and hosted release: BLOCKED. Fresh full Java21 clean verify passes419 tests with zero failures/errors/skips and unchanged quality gates. Its application archive payloads and all libraries match the frozen running preview. Seven final synthetic suites pass117 actual browser checks. Both package/image scans inventory269 Java packages with zero HIGH/CRITICAL findings; all251 application files match. Exact money qualification records commands, source/package hashes and limits. Native currencies, grouped symbols, right-aligned money and PDF figures, exact raw decimals, explicit missing data and separate rates/shares are qualified.

Historical preview localhost:57132 (closed; not live) used frozen app 372f094e (package SHA-256 prefix, not a commit) with an independently fictional household and fixture sign-in. Its replacement at50010 has now been stopped. The model was disconnected. Apps/tables-only navigation, contextual processes, Carl AI branding, lower-left chat, System Operations/ESB/admin diagnostics, saved dashboard selections and typed homes/mortgages passed controlled checks. Foundation numeric alignment PR18 merged through its required protected gate. Carl’s merged implementation PR1 records the development source. Hosted qualification remains blocked by FOUNDATION_PACKAGES_TOKEN and immutable RC/stable prerequisites. Earlier 408-test/113-browser RC8 qualification, 386-test shared-plan/restart qualification and 377-test/59-browser RC7 qualification retain historical package provenance.

The representative evaluation suite now has22 passing deterministic assertions, real PostgreSQL/API/native record agreement, five preserved offline Talk checks and explicit prerequisite/cleanup proof. Controlled evaluation evidence retains exact source identities and the live-only final correction limit. Six live cases remain BLOCKED; human narrative review is PENDING. The optional browser fixture is updated; the original377/59 qualification keeps its original fixture provenance, and the newer bounded-cleanup and shared-plan grammar correction have separate evidence above.

Complete native graceful shutdown: BLOCKED. The selected state-provider inspection confirms no supported close/shutdown API. FND-03 PASS covers retaining the inherited supported contract only; it does not establish complete lifecycle readiness. The earlier proposal was not installed or published; packaged JVM replacement is durability evidence.

Historical live-conversation credential hold: the existing authorized1Password token is unavailable to the CLI after dismissed desktop prompts. Controlled SDK/HTTP tests prove the conversational capabilities, while the running UI truthfully exposes an unavailable model. Real household/model/Synology identity/data-policy acceptance remains unqualified. Current RC8 foundation main0.5.0-SNAPSHOT publication is independently qualified at1464b10/run36878751071: signatures,480 source hashes,66 staged hashes,nine hosted Maven identities and17 fresh hosted consumer tests pass. The prior PR16 publication retains its separate evidence. The prior0b6f437 publication retains its own exact evidence. Carl-specific package access/checks remain blocked. Stable foundation0.4.1 remains the remote release baseline; the current parent0.5.0-SNAPSHOT is only a development closure. Deployment was not requested. The earlier evidence below is retained as historical provenance.

The earlier Carl0.1.0-SNAPSHOT source passed its 299-test integrated gate with no failures/errors/skips, Java21, copyright/Spotless/Checkstyle/ErrorProne/convergence and unchanged80%line/60%branch floors. Exact foundation0.4.0 is pinned. The packaged browser passed 31 checks, including narrow repeated two-file upload, protected PDF/text export, supported native source selection and logout/revoked-session replay. PDFs and main review screenshots were visually inspected. The application dependency scan inventoried 240 Java packages with no HIGH/CRITICAL findings. The original clean container scan is now historical: a refreshed database detects three HIGH OpenSSL package findings. The reviewed container-only repair upgrades those packages; repaired ARM64/AMD64 scans each inventory 240 Java packages with zero HIGH/CRITICAL and Java21/non-root smoke passes. The application JAR and 299-test/31-browser evidence are unchanged; no additional application rerun is implied. Source secret scan found none. See sealed local evidence.

Historical local application qualification and repaired container checks: PASS. Historical unpatched container: FAIL under refreshed vulnerability data. Hosted release qualification: BLOCKED until working push credentials, authorized private-package access and successful hosted checks. Integration qualified: BLOCKED pending real Synology/model/identity and data-policy acceptance. Deployed: NOT APPLICABLE — deployment not requested. This does not label the whole live product complete. Foundation0.4.0/sourcebb25a2a1f3ca562b054400a321773852ae839f39 is independently remotely published. Carl tested recovered exact release bytes matching all nine signed publication hashes; it has not proven a fresh credentialed consumer Maven fetch or hosted package access.

Baseline scenario evidence

PASS below means the stated controlled synthetic application behavior passed. It never means live model/provider acceptance, complete family data, or universal model correctness.

Scenario Status Actual evidence and remaining limits
AT-01 PASS CarlServiceTest.at01at02ImportExactCurrenciesAndDuplicateIdentity verifies exactUSD200 and separateEUR against PostgreSQL.
AT-02 PASS Same test plus BillCsvTest.invalidRowsAreExplicitAndBatchCannotCommit, source identity/retry tests.
AT-03 PASS at03at09at17MissingBillsAndUntrustedTextRemainFactsAfterAttributedCorrection: null amount/date stays missing; paid assertion requires evidence and remains unverified.
AT-04 PASS at04at13SavedReportRechecksSourceAccessAndRevocation, typed account/calendar restrictions, family/native HTTP and protected export negatives.
AT-05 PASS AgendaDecoderTest fixed recurrence/exceptions/cancellation/all-day/DST and private occurrence fixtures; PostgreSQL agenda ingestion.
AT-06 PASS CalendarAgendaServiceTest.expiredReadRetainsRecordsLastSuccessAndTruthfulCoverage, bounded controlledHTTP failures. LiveSynology separately blocked.
AT-07 PASS VendorRecordsTest immutable grounded drafts/versions/current-source export; CarlVendorNativeTest actual authenticated download.
AT-08 PASS CarlReadToolsTest closed read catalog; AgentApplicationTest governed SDK tool loop; provider types expose no financial/vendor writes. Calendar authority is the explicit later exception.
AT-09 BLOCKED Imported instructions remain inert source text and typed authorization denies caller spoofing; representative live-model adversarial evaluation is still unqualified.
AT-10 PASS FocusedReportsTest.calendarFailureAndVendorReportsPersistWithoutClaimingCompleteness, household coverage/freshness and missing-data facts.
AT-11 PASS CarlServiceTest.at11NarrationFailurePreservesFacts and focused report failure fixtures preserve deterministic facts with failed narration.
AT-12 PASS ReportRecoveryTest.interruptedWorkerIsFencedWithoutRepeatingNarration, committed-artifact recovery, durable request retries and conservativeUNKNOWN; no blind replay.
AT-13 PASS Source/member/field/linked-account denial across stored artifacts, family get and native/PDF/vendor exports.
AT-14 PASS workflowDeadlineCancelsSlowDatabaseReadAndLeavesServiceResponsive, bounded workflow queue and controlled provider/recurrence/input limits. Provider timeout adds its own bound; not an exact30s wall-clock guarantee.
AT-15 PASS Original377-test package launches from a separate directory;59 actual browser checks cover native financial/import/export/logout, branded grouped desktop/narrow dashboards, Talk and search/ESB. Synthetic TLS/OIDC only; live identities and deployment are separate.
AT-16 PASS Actual application/read-tool fixtures run withSlack disabled, bounded read catalogs and no prohibited write capability.
AT-17 PASS at03at09at17MissingBillsAndUntrustedTextRemainFactsAfterAttributedCorrection: original evidence retained, attributed correction and matching generated facts.
AT-18 NOT APPLICABLE — deployment not requested Production backup/restore qualification requires approved operating scope.
AT-19 PASS CarlConversationHttpTest.naturalLanguageWorkflowPersistsTheSameAuthorizedReportAndDraftShownByQqq proves the actual controlled SDK/HTTP-to-domain-to-QQQ report/draft path with protected history/retries/revocation in both the original377-test and fresh386-test gates. Five native Talk browser checks prove shared-service UI, saved outcomes and current privacy. Live model acceptance is separately BLOCKED; offline UI evidence does not prove a live answer.

Expanded scope and evidence

PurchaseAssessmentsTest, PortfolioPlansTest, ExpenseRecordsTest, RentalRecordsTest, RentalAllocationReviewsTest, BudgetPreferencesTest, TaxPlanningRecordsTest, BalanceSheetsTest, ReminderObservationsTest, CalendarAvailabilityTest and ReportRecoveryTest contain real PostgreSQL requirement fixtures, including shared/private/revoked boundaries. Their pure calculators additionally assert independently fixed cents, dates and shortfalls. No calculator-only test substitutes for a domain workflow.

The complete individual numbered requirement map records implementation, final-gate and live prerequisites separately. The earlier application source passed its local exact-source gates. The original377-test/59-check handoff and fresh386-test/32-check package gates pass; live and hosted prerequisites remain blocked. Tax rules remain undetermined without tested applicable rules; supplied government metadata is not archival custody or professional approval. All reports label selected/currently accessible scope and missing coverage.

Initial source is signed locally as e1213fa429166d8016f777c5a89016356dee105f; the reviewed container/evidence repair is recorded in the signed follow-up history. HTTPS push was rejected for missing OAuth workflow scope and the SSH agent refused signing. FOUNDATION_PACKAGES_TOKEN is absent; no remote implementation PR or release exists. Source/application scans and the repaired dual-architecture image scans are recorded with their separate hashes; the old container archive must not be released. Release policy tests pass seven cases using an isolated synthetic GPG key. Repository environments/tag immutability are installed; hosted development/RC/main delivery remains blocked until package access and required checks qualify. See CI/CD, session state and remaining work.

Default-model compatibility hold

The preceding local gate counts are retained as controlled evidence. Subsequent real SDK/HTTP regressions contradict broader default-model readiness: the selected foundation adapter loses required thinking/signature content in tool continuations and treats incomplete responses as completed. The correction is now published in independently qualified stable0.4.1. The current development adapter includes the qualified continuation correction, and377 tests plus matching package/browser/security checks pass. Live provider acceptance remains BLOCKED. No live provider acceptance is implied. The model's continuation requirements and stop-reason contract support this correction. See the current session hold.

The missing conversational implementation is integrated. Earlier development evidence used an unpublished corrected dependency; the current plan/goal evidence and 367-test gate use qualified stable0.4.1. These earlier results are superseded for current local package acceptance by the377-test/59-browser RC7 evidence above; they do not qualify live accounts.

October1 — final RC8 dock/System/public household development qualification

The current exact package evidence records408 Java tests with zero failures/errors/skips, unchanged80%line/60%branch floors,113 browser checks in seven distinct controlled suites,10 deterministic public-fixture tests and actual PostgreSQL acceptance of10920 transactions/22365 balances with35 reconciliations. Native/table/tool home profile facts and protected attributed history pass. Administrator diagnostics deny ordinary members. Both final distribution and exact container image inventory269 Java packages with zero HIGH/CRITICAL findings; all251 image application files match the frozen package.

The historical fictional preview ran localhost:63173 (closed; not live) (fixture sign-in), with native RC8 app/table navigation, System nesting, saved dashboard selections, Quick Search/ESB and a lower-left chat dock. Its model is disconnected;48 dock checks establish transport/history/privacy and honest UNKNOWN behavior, not live inference. The final native Talk suite adds5 controlled checks. Six real model evaluations, household/Synology policies, complete production shutdown and deployment remain unqualified. Existing source/fixture/package provenance remains separately retained.

Current HTTPS GitHub access includes workflow scope; earlier failed push evidence is historical. Fresh private Maven metadata returns401 and a matching immutable foundation release is unavailable. The unchanged RC preparation guard rejects the development parent while preserving the POM. No complete product or immutable release is claimed.

October 2 — integrated document, CSV and receipt source

V61 stores immutable supplied documents and original bytes with source identity, nullable dates, explicit audience and unverified status. Native upload, protected review/confirmation, bounded extracted-text reads and original downloads use current domain permissions. V62 stores bounded selected/all-authorized CSV snapshots behind a caller-bound history and native download process. Downloads recheck current source access and permission epochs, including a change after storage load commits. Fixed mutation receipts acknowledge only their own committed operation; they do not expose retrieved private fields or make old process state reusable.

The completed 532-test integrated source gate retains its original provenance. The final expanded gate and matching replacement-package checks are pending at this review. The running local QA package remains the separately qualified 490-test/V60 package; actual historical-document import, CSV acceptance and V61/V62 private migration/startup have not been completed. Isolated Liquibase transition checks do not establish a matching deployed engine or a private rollout.