Updated 2026-08-22. This file describes the accepted V10 product tree and its
current production deployment. The current handoff is
docs/V10_RELEASE_STATUS.md; V8 matrices remain
dated evidence.
The accepted production deployment remains dpl_6f9rhpqCTsdDcyvoR5VN489nP1vF.
Candidate commits b11b471, c48d58d, 1264423, 356190c, 5451bf8,
dc5c343, 35c32cd, 6feb07a, c4f13e8, e3c24ab, 9aa5440,
bf987e2, 750a9f5, 3dafaa9, a9542ed, 292532a, 894c03d, 5e02c65,
cb0e9f4, 2a05cfe, 641d039, 8f8c5ab, 7ff2fcb, 4cfc608, da8fdfc
and 360da39
add durable-read/session routing fixes, GitLab observation persistence,
LongMemEval evidence alignment, bounded DOCX XML extraction, and idempotent
browser agent retries, duplicate-submit auth and Work guards, and remove the
obsolete DOCX parser dependency. They are
pushed to v10-zeus-release but could not be promoted on 2026-08-22 because
Vercel rejected new deployments after the free daily deployment quota was
exhausted.
The latest candidate passes the complete serial unit suite: 117 files and
2,294 tests, with no skips. Commit 641d039 also makes the landing hero
session-aware, so a verified member returns to the private workspace instead
of being sent to the public read-only explore route.
- 117 unit test files and 2,294 tests passed with no skips in the latest serial verification.
- Root and web TypeScript checks pass.
- The production web bundle builds from 139 modules. The entry is 306.97 kB, 100.66 kB gzip.
- Copy lint scans 58 files with zero findings.
- Public and private graph APIs expose an interactive overview and an exact provenance table/DAG with signed cursor pagination.
- Two bounded, no-write agents run Researcher to Reviewer handoffs from a Context Pack. Memory-derived recommendations state their reason, evidence, permissions and budgets before any run or schedule is created.
- The public workspace is read-only in production:
POST /api/explore/agent/runand its legacy/api/demoalias return403 public_preview_read_onlybefore reading a body or calling a provider. Existing public run records remain inspectable. Authenticated, CSRF-protected/api/workspace/agent/runstill creates a real workspace-scoped run. - Daily schedule creation, run-now, retry and cancellation exist. The cron dispatcher discovers private workspaces server-side and authenticates before reading the registry.
- Google OAuth verifies Google's RS256 signature and claims, uses state, PKCE S256 and nonce, and binds an account to the provider's stable subject.
- Password recovery rotates a credential epoch and invalidates every older 30-day session.
- Workspace/profile updates are isolated from credential records, so an in-flight rename cannot restore a pre-recovery password or session epoch.
- Hosted password signup is disabled. HydraDB's current writer cannot provide the atomic unique create needed to make same-email password signup safe.
- Private MCP uses a 256-bit random bearer stored digest-only. It expires 30
days after issue, can be revoked earlier, and issuance returns
createdAtplusexpiresAt. Capabilities fail atexpiresAt; version-1 records fail closed and must be reminted after rollout. The workspace comes from the signed-in account, never from request input. Body, request, tool, write and issuance limits are enforced. - Hydra Cloud configuration accepts only the exact canonical HTTPS origin, so a lookalike host cannot receive the bearer through misconfiguration.
- Production answers use HydraDB Cloud as a collection-scoped addressed-record store. Lacuna fetches those deterministic entity records and applies temporal standing, contradiction, abstention and multi-hop resolution in application code. The native Cypher adapter and 162-query compatibility probe target a separate self-hosted HydraDB node; they are genuine proof, not the deployed answer path.
- One-off file ingestion (TXT/MD/JSON/CSV/PDF/DOCX), public GitHub and GitLab snapshot imports, public HTTPS/API import, and signed bounded webhook delivery are implemented workflows. The production redacted catalogue exposes eight descriptors and marks each available when its server-side support is configured. Linear, Jira, Slack, Notion, Gmail, Confluence, Database source and Spotify remain planned.
- The exact 399-character
package-sessionblast-radius request isNOT_PROVEN. The subject is absent, the sentence exceeds the 300-character sentence API contract, and no Web, CLI or MCP blast-radius command carries that full request. The recorded hostile audit isartifacts/verification/2026-08-21-v10/package-session-proof-audit.json. - The landing, dashboard, agents, tools, graph and mobile shell have been redesigned. Reduced motion disables non-essential movement.
- Voice browser/runtime routes are implemented and fixture-tested. Typed Ask remains available when the provider is absent. Candidate browser playback and connector response readers now cancel stalled streams instead of leaving the UI busy. Candidate Google token/JWKS JSON reads share the provider deadline, and landing overlays are shielded from particle labels.
- LongMemEval has a scoped exact-span personal-memory bridge for explicit first-person facts. The oracle ingest check reads all 500 records with zero adapter failures and zero ground-truth leaks, producing 128 claims across 84 records (16.8% ingest coverage). No official score is claimed.
The stable URL is https://lacuna-five.vercel.app. It points to immutable
deployment dpl_6f9rhpqCTsdDcyvoR5VN489nP1vF, built from release commit
d5f9d02. Post-deploy checks passed HydraDB health, Google OAuth start with
PKCE/state/nonce and Referrer-Policy: no-referrer, malformed callback
fail-closed handling, and private auth CSRF refusal. Public main and the
release branch contain the accepted product source.
The selected narration voice is the verified ElevenLabs Vaibhav Lalwani Professional clone. Raw clone audio stays local and gitignored so a biometric voice asset is not redistributed. A narration render is not proof of the product's live voice route; no provider-enabled STT/TTS/interruption session is accepted in production.
The 179-second V8 MP4 is historical and rejected. The V10 master is machine
accepted at 178.500 seconds, 1920×1080/30 fps with H.264 video, AAC audio and a
full decode pass; its SHA-256 is
e73e6e0bf1de598b3c1c998a43057ac06e8dcb3b492a19d3ac8623c8d9cb9d96.
The owner's uninterrupted watch, Supademo, YouTube upload and form submission
remain open. The seven-tool public ChatGPT MCP proof is accepted; Claude remains
untested. A packaged Lacuna SDK and a Spotify connector are not shipped
capabilities. Nothing has been submitted or uploaded.
- HydraDB's adapter exposes no CAS or transaction. Cloud scheduler leases, agent idempotency, quotas and workspace registry updates are not distributed-atomic across serverless instances. Do not claim exactly once.
- Provider work can begin twice during a multi-instance race before a shared run record collapses the result.
- API budgets are process-local. They are useful guardrails, not distributed billing controls.
- The cron registry scan is sequential and needs sharding or a durable queue at large tenant counts.
- CLI and MCP share canonical memory/evidence reads but do not expose agent lifecycle controls.
- Existing pre-binding Google records deliberately fail closed until the signed in owner completes the verified Settings → Link Google migration flow.
- MCP path-form capabilities may appear in infrastructure logs. Prefer
Authorization: Bearer <capability>at/mcp; use/mcp/w/<capability>only when a client cannot set headers. Revoke the credential when finished even though it expires after 30 days.
Do not submit the hackathon form and do not upload or publish the final video.
Those are owner actions. Draft answers and an acceptance checklist are in
docs/SUBMISSION_FINAL.md and
docs/SUBMISSION_CHECKLIST.md.