This is the canonical owner handoff for the protected V10 release. It is paste-ready except for the YouTube URL, which only the owner adds after the local film passes owner review and is uploaded. The V10 machine gates now pass; nothing in this file means the video was uploaded or the form was submitted.
| Item | Accepted value |
|---|---|
| Track | 03 — Memory and Context Retrieval |
| Product commit | 05fe9d15dea75e4db7f5eb61d17533aa26e6e5a8 |
| Acceptance-doc commit | 695b55e63dd3a28ff6ee2f6f83ec3919bf6f0eaf |
| Immutable deployment | dpl_GZhotqcHc2p3f2AKCeezQKNidjwc |
| Immutable URL | https://lacuna-6fq3hiy9q-vaibhav4046s-projects.vercel.app |
| Stable URL | https://lacuna-five.vercel.app |
| No-account judge path | https://lacuna-five.vercel.app/judge |
| Public repository | https://github.com/vaibhav4046/lacuna |
Public main at release probe |
05fe9d15dea75e4db7f5eb61d17533aa26e6e5a8 |
The machine-readable release probe is
artifacts/submission/v10-exact-release-probe.json.
Later local commits do not inherit this deployment gate.
Lacuna
Lacuna is temporal, provenance-first memory for long-running agents, built on
HydraDB. It keeps corrections, contradictions and exact source evidence as a
queryable graph, returns the current supported answer, and names the reason
when the history cannot support one.
Long-running agents do not only forget. They confidently retrieve facts that
were later corrected, treat proposals as decisions, collapse two disagreeing
sources into one answer, and invent answers for facts that were never stated.
Similarity search can rank relevant passages, but it does not establish which
claim is current, what it replaced, or whether the evidence is missing.
Lacuna gives cross-session memory an explicit temporal and provenance model so
another model receives a small, evidence-bearing Context Pack instead of an
unresolved transcript.
Lacuna ingests conversations as immutable source evidence and claim records.
Corrections create explicit supersession history instead of overwriting old
facts. A deterministic resolver returns current answers, source quotations,
revisions, conflicts, proof paths and machine-readable abstention reasons.
The product includes a deployed no-account judge workspace, plain-English Ask,
a searchable memory table, an interactive overview graph, an exact evidence
DAG, a nine-command CLI, stdio MCP, a live seven-tool read-only HTTP MCP
endpoint, and two bounded no-write agent roles with persisted Work records and
handoffs. Signed-in workspaces and daily schedule definitions exist; the public
preview remains read-only.
The reproducible demo corpus contains 72 generated sessions, 5,246 messages,
174 claims and 86 entities. On its generated 64-question evaluation, Lacuna
answered 64/64 with zero false or unsupported answers and used 18.27 mean
estimated context tokens. The strongest tested flat-retrieval configuration
answered 63/64 with 1,842.57. This is a single generated corpus, not official
LongMemEval or a general accuracy claim. No LongMemEval score was produced.
https://lacuna-five.vercel.app
Optional direct judge link:
https://lacuna-five.vercel.app/judge
HydraDB is Lacuna's persistent context substrate, not a decorative export. In
the accepted production profile, 72 generated conversation sources, 86
addressed entity records and one index record were accepted and indexed in a
workspace-scoped HydraDB Cloud collection. The answer path fetches those
deterministically addressed records; HydraDB's query and relations surfaces
separately expose semantic and graph context. Lacuna then applies temporal
standing, contradiction handling, abstention and bounded relationship
resolution in application code.
The repository also includes a self-hosted adapter pinned to HydraDB v0.1.1.
That adapter stores nodes and edges and executes bounded native Cypher through
NodeSource; 162 compatibility probes are retained. Native Cypher is genuine
self-hosted proof, but it is not the deployed Cloud answer path.
The two adapters returned field-for-field identical outcomes on all 64
generated questions. Without HydraDB, the deployed product loses its durable,
collection-scoped source and claim records, relation inspection, graph context
and shared retrieval seam. Lacuna's policy layer is deliberately separate: it
decides what is current or unsupported after HydraDB returns the evidence.
TypeScript, Node.js 20.11+, HydraDB over HTTP, React 19, React Router, Vite,
the official Model Context Protocol SDK, Vitest, hash-wasm and HyperFrames.
Lacuna code is Apache-2.0. HydraDB is AGPL-3.0 and runs as a separate service.
Vaibhav Lalwani — solo builder: product design, temporal memory model, HydraDB
adapters and evaluation, web product, CLI and MCP surfaces, agent runtime,
testing, deployment, documentation and demo direction.
https://github.com/vaibhav4046/lacuna
[OWNER: PASTE THE VERIFIED UNLISTED YOUTUBE URL HERE]
Do not use an earlier V8 render. The accepted V10 master must be at most 179 seconds, carry the Vaibhav Lalwani Professional narration, contain burned-in captions, and pass the local metadata and full-length review gates. Publication of the clone narration also requires the owner's confirmation that its use is authorized.
Local V10 master: video/hyperframes-v10/renders/lacuna-v10-hack-hydra-final.mp4
(178.500 seconds; SHA-256
e73e6e0bf1de598b3c1c998a43057ac06e8dcb3b492a19d3ac8623c8d9cb9d96).
Machine acceptance is recorded in
artifacts/video/v10-final-metadata.json.
- Production Cloud does not execute the self-hosted native-Cypher answer path.
- The exact 399-character
package-sessionrequest isNOT_PROVEN. - Bounded file ingestion (TXT, Markdown, JSON, CSV, PDF and DOCX), public GitHub/GitLab snapshots, public HTTPS/API import and signed webhook delivery are implemented. Linear, Jira, Slack, Notion, Gmail, Confluence and Database source remain planned account integrations. Spotify is not implemented.
- ChatGPT is proved only for the seven-tool public read-only corpus. Private
rememberis not accepted. Claude was not tested. - No packaged Lacuna SDK is published. The official MCP SDK is an internal dependency.
- Provider-backed voice is bounded and tested; typed Ask remains the explicit fallback whenever microphone, provider or browser playback is unavailable.
- Google OAuth security and an authorized chooser → callback → dashboard round trip are accepted on the stable alias. The latest hosted session-read timeout and cross-browser playback candidate are pushed but await Vercel promotion.
- Hosted schedules are persistent; distributed exactly-once execution is not claimed.
- The evaluation is generated and is not LongMemEval. No official benchmark score exists.
- Open the final local MP4 and its matching V10 metadata/SRT; confirm duration is at most 179 seconds, audio and captions are complete, no secret or private identifier appears, and every claim matches V10_FINAL_EVIDENCE_MAP.md.
- Watch the entire film once without scrubbing. Confirm the Vaibhav clone is authorized for publication.
- Upload that exact MP4 to YouTube as Unlisted. Do not upload the rejected V8 film.
- Open the YouTube URL in a signed-out/private window, confirm it plays without an access request, and verify YouTube displays a runtime below 3:00.
- Paste the URL above, copy the official answers into the form, re-open the production and repository links signed out, then submit and retain the confirmation.
Optional proof improvements such as completing a fresh Google identity callback or reminting a private MCP capability are not required to submit and must not be described as complete unless separately captured.
The repository and product links are public and verified, and the local V10 master passes codec, duration, decode, audio, SRT and sampled-frame gates. The official entry is still incomplete until the owner watches and authorizes the film, uploads it, inserts its viewable URL, and submits the form. No upload or submission was performed while preparing this handoff.