END_TO_END_MATRIX.md

End to end matrix

Every product surface, and what was actually run against it.

One row per surface. What was exercised is the behaviour, not the intent. Result is the output of the run, or the words NOT VERIFIED where no run produced one. Evidence is a path in this repository or a command whose output is quoted in a path. A row with no evidence path is a row nobody has to believe.

Two sources of results appear below and they are kept apart on purpose. Rows citing RELEASE_GATE.md were run at v7-freeze, commit 942de9e, and the gate file records the command and its output. Rows citing a dated directory under artifacts/verification/ were run earlier and are kept unedited, which means some of their counts are smaller than today's. Where the two disagree, the gate file is the later number and the artifact is the record of when it moved.

Web

Surface	What was exercised	Command or action	Result	Evidence
Landing, production	Root 200 on both spellings of the host, root element present, first paint stylesheet served and black, no-JavaScript recovery state, favicon, deep route survives a refresh, entry module referenced	npm run smoke:web -- https://lacuna-five.vercel.app	9 of 9	RELEASE_GATE.md, scripts/smoke-web.ts, earlier local run at artifacts/verification/2026-08-18-v6/smoke-web.txt, capture at artifacts/screens/live/live-landing-1920x1080.png
/judge	Page served, 200, carries the application root	npm run smoke:demo -- https://lacuna-five.vercel.app	Part of 30 of 30	RELEASE_GATE.md, scripts/smoke-demo.ts, captures at artifacts/screens/live/live-judge-1920x1080.png and live-judge-fullpage.png
The 18 app routes	The route table itself, and every read the demo screens make: 15 /api/demo/* parts answered 200, a write to one returned 404, three deep demo routes plus /judge served with the root element	npm run smoke:demo -- https://lacuna-five.vercel.app	30 of 30. Three of the 18 routes were fetched as pages (/demo/dash, /demo/ask, /demo/hydra); the other 15 are the same shell reading the same 15 endpoints, all of which answered	RELEASE_GATE.md, web/src/app/routes.ts, scripts/smoke-demo.ts, captures at artifacts/screens/live/
Browser console, every route	Console errors, uncaught exceptions, failed and 400-and-above requests, horizontal overflow and whether anything was drawn, at nine viewports from 360px to 4K, then again with prefers-reduced-motion set	npm run audit:routes and npm run audit:routes -- <url> --reduced-motion	198 of 198 in each pass. 0 console errors, 0 exceptions, 0 failed requests, 0 routes scrolling sideways. The first run of this audit failed 18 of 23: every signed-in route scrolled sideways on a phone, which is fixed	artifacts/route-audit/routes.json, routes-reduced-motion.json, scripts/route-audit.ts
Ask	One live question per resolver outcome through /api/ask, the endpoint the screens use, each checked for source_state: live and a measured took_ms; a two hop question answered and cited; at least three distinct outcomes reached	npm run smoke:demo -- https://lacuna-five.vercel.app	Part of 30 of 30	RELEASE_GATE.md, scripts/smoke-demo.ts, artifacts/continuity/one-context.json, capture at artifacts/screens/live/live-ask-1920x1080.png
Workspace, demo	The demo workspace holds the corpus rather than being an empty workspace with a name; it refuses writes	npm run smoke:demo -- https://lacuna-five.vercel.app	Part of 30 of 30, memory row total greater than zero, POST returns 404	RELEASE_GATE.md, scripts/smoke-demo.ts, src/api/workspace.ts
Timeline	The web Timeline screen reads changes and conflicts, both answered 200 in the demo sweep. The revision chain itself (considered, oldest first, with validFrom, txTime, supersededBy, current) is returned by the MCP and CLI timeline paths and is covered by unit tests	npm run smoke:demo, npx vitest run tests/unit	Screen reads answered, part of 30 of 30. Revision chain covered in tests/unit/mcp-server.test.ts and tests/unit/mcp-tools.test.ts. No committed transcript of a live lacuna_timeline or lacuna timeline call exists	RELEASE_GATE.md, src/mcp/result.ts, web/src/app/routes/context.tsx, capture at artifacts/screens/live/live-timeline-1920x1080.png
Contradiction	A question whose sources disagree returns CONFLICT with reason contradicted, identically from three clients against one cloud workspace	npm run continuity	q-contradicted-01, same: true, CONFLICT, one conflict listed	artifacts/continuity/one-context.json, RELEASE_GATE.md
Missing evidence	A retracted value and an out-of-scope subject both return NO_EVIDENCE with their own reason codes, identically from three clients	npm run continuity	q-retracted-01 reason retracted, q-out_of_scope-01 reason out_of_scope, both same: true	artifacts/continuity/one-context.json
Mobile viewport	Landing and /judge captured at 375x812 against production, each checked on write for ground colour and compressed density	npm run screens	Two captures exist and passed their checks. No 375-wide capture exists for the other 16 routes	artifacts/screens/live/live-landing-375x812.png, artifacts/screens/live/live-judge-375x812.png, scripts/screens.ts

Accounts

Surface	What was exercised	Command or action	Result	Evidence
Sign up	A fresh visitor reads as signed out, the session read issues a CSRF cookie, sign up creates an account, a taken address is refused	npm run smoke:auth -- https://lacuna-five.vercel.app	Part of 12 of 12	RELEASE_GATE.md, scripts/smoke-auth.ts
Sign in	Sign in works from a clean cookie jar, a wrong password is refused	same	Part of 12 of 12	RELEASE_GATE.md, scripts/smoke-auth.ts
Session persistence	The session survives another function invocation, and sign out ends it	same	Part of 12 of 12	RELEASE_GATE.md, scripts/smoke-auth.ts
Onboarding	Onboarding names a workspace and the name persists to the next read	same, plus a by-hand pass in a browser from a cleared cookie jar	12 of 12. The by-hand pass reached the dashboard reporting 0 current · 0 historical · 0 conflict, a real empty workspace	RELEASE_GATE.md, scripts/smoke-auth.ts
Workspace, signed in	A signed-in workspace reads empty unless it is the one named as the demo	covered by the account suite and by src/api/router.ts	12 of 12 for the account flow. The signed-in path to the demo corpus is decided in code and asserted in the unit suite, not in a production transcript	RELEASE_GATE.md, src/api/router.ts
Google sign in	A previous production browser pass reached onboarding, but the later security review found an unsafe email-only account merge	held below release acceptance	Provider/subject binding, JWKS/RS256, PKCE and nonce candidate code exists. Router wiring, no-store redirects, negative merge tests and a new production browser pass are still required	docs/FINAL_EXECUTION_STATE.md, tests/unit/google-auth.test.ts, tests/unit/auth-identity.test.ts

Clients

Surface	What was exercised	Command or action	Result	Evidence
CLI	Every gold question asked through lacuna ask and compared field by field against both MCP transports: status, answer, reason code, claim id, superseded claims, evidence, evidence total, source state, and every graph read with its parameters and row counts	npm run parity	SWEEP_IDENTICAL: 64 of 64, ALL_IDENTICAL: True	RELEASE_GATE.md, artifacts/verification/2026-08-18-gates/parity.txt, single-question captures at artifacts/verification/2026-08-14b/cli-ask.json and cli-abstain.json
MCP	The same sweep over stdio and HTTP; four tools advertised over stdio at the time of that capture; a third-party client consumed the documented config block and drove both transports	npm run parity, plus the MCP Inspector CLI	64 of 64 identical. Seven tools are advertised on the public endpoint now: five Lacuna-native reads (lacuna_ask, lacuna_explain, lacuna_timeline, lacuna_read_question, lacuna_health) plus the connector-compatible search and fetch. A live tools/list against deployed /mcp returned all seven on 21 Aug. The older stdio transcript remains a historical four-tool capture	RELEASE_GATE.md, artifacts/verification/2026-08-14b/mcp-stdio.txt, artifacts/verification/2026-08-14e/README.md, timings at artifacts/mcp/stdio-timings.txt, live catalog source at src/mcp/tools.ts
Lacuna surface continuity	The deployed web over HTTPS, the CLI in a local process, and an MCP server started as a subprocess, all reading the same HydraDB Cloud workspace, on six questions covering six outcomes	npm run continuity	ONE_CONTEXT_IDENTICAL: true, six questions, every row same: true. This is not a ChatGPT or Claude proof	RELEASE_GATE.md, artifacts/continuity/one-context.json
ChatGPT custom app	Production public MCP health, ask, timeline, explain, sentence read, search and fetch against HydraDB Cloud	invoked through the installed ChatGPT Lacuna app	VERIFIED READ ONLY; seven of seven public tools accepted	artifacts/verification/2026-08-21-v10/chatgpt-public-connector.json
Claude MCP	Not exercised against Lacuna	not run	NOT VERIFIED	none
Packaged Lacuna SDK	No package exists	package inspection	NOT SHIPPED	package.json

HydraDB

Surface	What was exercised	Command or action	Result	Evidence
Cloud ingest	The corpus and claim graph written to HydraDB Cloud, indexed, and a sampled entity plus the index read back and compared to what was written	npm run ingest:cloud	159 written, 159 accepted, 159 indexed, 0 refused; sample and index both read back identical; 86 entities, 174 claims, 72 sessions	artifacts/hydra/cloud-ingest.json
Cloud reads, against the node	All 64 gold questions asked of the node and of the cloud, compared field by field with wall clock and read log excluded	npm run parity:cloud	identical: true, 64 questions, node 342 reads against cloud 119, median node 108ms against cloud 230ms	artifacts/hydra/cloud-parity.json, RELEASE_GATE.md
Cloud health, from the deployment	A real round trip from the deployed function to the configured database	GET /api/health on production	ok, api.hydradb.com answered in 160ms, database lacuna, collection backend	RELEASE_GATE.md, api/index.ts
Local node, contents	Every stored key read back and compared against the ingest plan, so a missing node and a stray node cannot cancel out	npm run census	graph matches the plan exactly	RELEASE_GATE.md, artifacts/verification/2026-08-18-gates/census.txt
Local node, counts	The loaded graph reported back by the CLI	npm run cli -- status	Session 72, Message 5246, EvidenceSpan 174, Claim 174, Entity 86	RELEASE_GATE.md records the ingest as 5752 vertices and 5908 edges; the per-label counts are the output of this session's run
Local node, query layer	Every query builder run against a live node, skipping rather than passing when none answers	npx vitest run tests/contract	77 passed, 4 files, none skipped	RELEASE_GATE.md, earlier run at artifacts/verification/2026-08-18-gates/contract.txt

Checks

Surface	What was exercised	Command or action	Result	Evidence
Typecheck	The whole tree, no emit	npx tsc --noEmit	exit 0	RELEASE_GATE.md, earlier run at artifacts/verification/2026-08-18-gates/typecheck.txt
Unit tests	The suite that needs no database	npx vitest run tests/unit	1,208 passed, 59 files	RELEASE_GATE.md, recorded in artifacts/verification/2026-08-19/unit.txt. Earlier dated artifacts keep their own counts and are not edited
Contract tests	Query builders against a live node	npx vitest run tests/contract	77 passed, 4 files	RELEASE_GATE.md, artifacts/verification/2026-08-18-gates/contract.txt
Benchmark	Lacuna against five flat retrieval approaches over the same corpus, 51 configurations, embeddings run locally	npm run bench	Lacuna 64/64 with 18 mean context tokens; best baseline hybrid+2hop@50 +conflict at 63/64 with 1843. A one question lead, stated as one	artifacts/bench/report.txt, results.json, read in full in docs/BENCHMARKS.md
Snapshot replay	All 64 gold questions replayed from the recorded snapshot through the real decoder and resolver	npm run snapshot:verify	64 questions, 0 answer mismatches, 0 wrong verdicts	artifacts/verification/2026-08-18-gates/snapshot-verify.txt
Under load	400 requests at concurrency 12 against the deployed ask endpoint, every response compared to the same question asked alone	npm run soak	26.3 a second, p50 387ms, p95 805ms, p99 1423ms, 0 failures, every answer unchanged	artifacts/soak/soak.json, scripts/soak.ts
Blast radius, both stores	A package change walked for what it reaches, on HydraDB Cloud and on the node	npm run proof	13 services at depth 3 for both pact-check and moss-index, the two stores returning the same set, one superseded dependency edge refused	artifacts/proof/proofs.json
False premise, three clients	An unsupported premise beside a real revision and an unknown subject, through the deployed web, the CLI and an MCP subprocess	npm run proof	THREE_CLIENTS_IDENTICAL: true, 1 answered and 2 refused with different reason codes, none guessed	artifacts/proof/proofs.json
Mobile and reduced motion	Nine viewports from 360x800 to 3840x2160, then the same sweep with prefers-reduced-motion set	npm run audit:routes, then with --reduced-motion	198 of 198 in each pass, 0 scrolling sideways	artifacts/route-audit/routes.json, routes-reduced-motion.json
Landing scenes	Every data-scene section measured for drawn text, so no scene is an unexplained empty viewport	Browser, against production	28 scenes, 0 blank. The shortest carries one approved line, No evidence means no answer.	measured in the browser on 2026-08-19
Public copy	Landing, app routes, onboarding, auth, README and the social captions, for filler, em dashes as punctuation, exclamation marks and over-long sentences	npm run copy:lint	47 files, 0 findings	scripts/copy-lint.ts
Secrets	Every live key value searched across tracked files and the whole history	git grep -F, git log --all -S	0 tracked files, 0 commits	docs/CREDENTIAL_ROTATION_CHECKLIST.md
Architecture boundaries	The web importing the resolver, a surface opening its own store, a client re-deciding which claim is current, a route naming a gold answer	npx vitest run tests/unit/architecture.test.ts	8 of 8, and each guard verified by deliberately breaking it	tests/unit/architecture.test.ts, docs/ARCHITECTURE_INVARIANTS.md

What this matrix does not cover

Named here rather than left for a reader to notice.

• Google sign in is not accepted. A redirect and earlier browser success are insufficient because the old callback merged accounts by email. The new provider-bound path must pass integration, negative and production-browser gates before it can be checked above.
• No connector syncs. The Connectors screen lists nothing that pulls. File and document ingest through the web is the largest missing piece of product completeness, and it was left unstarted rather than half started against a frozen release.
• Voice is unconfigured in the deployment, and the screen says so.
• No end-to-end provider audio exists. The route and state-machine tests do not prove microphone, realtime transcription, the selected clone, streamed playback or interruption in production.
• No evidence about two writers. The soak row above covers concurrent readers. Nothing here says what happens when two clients write at once.
• No distributed scheduler atomicity. HydraDB persistence has no CAS or transaction seam in this adapter. Process-local locks, quotas and leases do not establish exactly-once behaviour across Vercel instances.
• Roughly a fifth of the specified harness exists. Counted item by item in docs/HARNESS_CONFORMANCE_MATRIX.md: nothing end to end, four partial, six absent. There is no run object, so everything present is scoped to a single question.
• The corpus is generated by this project. The retrieval comparison is fair between the systems it compares, because they all read the same corpus, and it is not a public benchmark. The same generator wrote the questions, so 64/64 says the pipeline does what the structure says, not that it is right about the world.
• No per-route mobile capture. Two surfaces were captured at 375x812. The other sixteen routes have no narrow-viewport evidence.
• No live transcript of the timeline or explain tools. Both are covered by the unit suite and both go through the same shared projection as ask, which the parity sweep covers on 64 questions. Neither has a committed transcript of its own.
• No editor or agent runtime has held an interactive MCP session. The clients that have connected are a stdio driver from this repository, the SDK's own client in the parity run, and the MCP Inspector CLI. A client run from a terminal is not a host.
• No private MCP issuance proof. The candidate listener fails closed and the capability store persists only digests, but the authenticated issue/use/ revoke path is not yet wired and deployed.
• No ChatGPT/Claude or Supademo evidence. Neither may be inferred from the web/CLI/MCP continuity artifact. A metadata-verified local judges-master video candidate exists separately; it is not owner-approved or published.
• artifacts/deployment/production.json is a record of 2026-08-18 and one of its honestLimits lines is now stale. It says the answer engine still reads the self-hosted node. The deployment now reads HydraDB Cloud, which is what npm run continuity and npm run parity:cloud measure. The file is kept unedited because a transcript edited to agree with a later run has stopped being a transcript.

Related

• RELEASE_GATE.md, the gates and their commands.
• docs/EVIDENCE_INDEX.md, every public number and the file it came out of.
• docs/ARCHITECTURE_FINAL.md, the layers these surfaces sit on.