Research Requirements

Productization command lane

Video → Goal-Data Execution System

Night Shift 3 refresh: architecture v1.1, schema/taxonomy contract, execution confidence lanes, dependency action queue, and an action-first promotion control board.

Night Shift 3 plan (current) Previous cycle Requirements contract Execution status: Scope locked · M1 in progress · v1.1 baseline

Release quality gates

Ingestion success
≥ 98%
Label completeness
≥ 97%
Inter-annotator agreement
κ ≥ 0.78
E2E pipeline latency
p95 < 30m / 10m video

Execution control rail

Now
M1 contract lock + ingest fail-fast enforcement
  • Schema/API contracts frozen for v1.1.
  • Consent/privacy validator required before write.
  • Lineage closure remains hard release precondition.
Next decision
Reviewer/adjudicator staffing for M3 lane
  • Need 2+ weekly reviewers + 1 adjudicator lane.
  • Direct impact on M3/M4 confidence and handoff speed.
Blocked risk
Retention policy + legal consent mapping
  • Pilot datasets blocked if unresolved before M2 close.
  • Owner lane: Founder + Legal + Data Ops.

Milestone confidence lanes (modernized)

M1 · W1-2
Contract + ingest integrity

Schema/API lock, validator hard-blocks, immutable storage lineage.

Confidence: HighExit: ingest ≥98%
M2 · W3-4
Temporal reliability

Segmentation/event workers, revision lanes, disagreement routing SLA.

Confidence: MedUnblock: threshold tuning
M3 · W5-6
Label ops stabilization

Dual-review enforcement, adjudication trail, queue SLA instrumentation.

Confidence: MedUnblock: reviewer capacity
M4 · W7-8
Build reproducibility

Deterministic manifests, diff verifier, schema integrity checks.

Confidence: HighExit: repro ≥0.99
M5 · W9-10
Eval + promotion

Failure trend packet, go/hold/reject decision workflow, rollback pointer.

Confidence: MedUnblock: compute scheduler

MVP scope snapshot

In scope
Cycle 3 productization core
Ship now
  • Single-tenant ingestion + immutable lineage
  • Segment correction + disagreement routing
  • Taxonomy-bound labeling with safety dual-review
  • Deterministic dataset builds + evaluation gating
Out of scope
Post-MVP expansion lane
Later
  • Real-time streaming ingestion
  • External self-serve annotation portal
  • Multi-tenant RBAC + billing
  • Full active-learning auto-relabel loop

Architecture layers (v1.1)

  • Capture + ingestion plane with consent/privacy preflight and immutable storage lineage.
  • Temporal structuring plane with disagreement triage and revision audit history.
  • Goal semantics + label ops plane with dual-review for safety-tagged failures.
  • Dataset build plane with deterministic manifests and reproducibility checks.
  • Evaluation + promotion plane with baseline comparison and rollback pointer.
  • Execution observability plane with trace IDs, SLA debt counters, and unblock metadata.
Architecture contract: v1.1 · lineage closure + packet completeness required for promotion

Evaluation loop control board (modernized)

Step 1
Candidate build
Owner: Data Ops · Freshness: 4h · Status: Ready
Step 2
Benchmark + failure buckets
Owner: ML Infra · Freshness: 6h · Status: Running nightly
Step 3
Corrective action queue
Owner: Eng + Research Ops · Freshness: 12h · Status: 3 open actions
Step 4
Decision packet assembly
Owner: Founding Eng · Freshness: 24h · Status: In progress
Step 5
Release council decision
Cadence: Mon/Wed/Fri · SLA: same-day follow-up owner assignment
Cadence: Mon quality review · Wed failure triage · Fri release council

Dependency action queue (modernized)

Technical
Immutable object storage + lifecycle policy
High impact
Owner lane: Platform/Data Ops · ETA: W2 · Freshness: 24h
Unblock action: Finalize retention matrix + run lifecycle policy smoke tests on 20 seed assets.
Technical
Metadata DB + migration guardrails
High impact
Owner lane: Backend · ETA: W2 · Freshness: 12h
Unblock action: Add migration contract tests + rollback verification in CI before M1 exit.
People
Reviewer + adjudicator staffing lane
Schedule risk
Owner lane: Research Ops · ETA: W5 · Freshness: 48h
Unblock action: Lock weekly rota (2 reviewers + 1 adjudicator) and publish escalation backup owner.
Governance
Consent + retention legal policy sign-off
Blocking risk
Owner lane: Founder + Legal + Data Ops · ETA: W3 · Freshness: 72h
Unblock action: Approve deletion/retention matrix and bind exception handling to release packet checklist.

Working notes · notes/dataset-requirements-from-literature.md

Ready