AI Delegation Plan
Status — living document, iterated per dispatch cycle
This page maps every roadmap work item to a delegation tier, an executor, and the data it needs — and it is meant to be re-graded continuously: each completed agent brief either promotes more work into Tier 1 (by building the harness that makes it verifiable) or documents why it stays human. Companion page: Agent Briefs — Batch 1 contains the dispatch-ready briefs. The goal state: everything delegable is delegated, and humans spend their time only on Tier 3.
The delegation rules
- No brief without an acceptance test. Every delegated item ships as brief + hard acceptance criterion (reproduce metric X, pass suite Y, precision ≥ Z on a labeled sample) so the agent can self-verify before a human reviews. The roadmap sub-pages' "done when" lines are the seeds of these criteria.
- Agents never hold production credentials or touch contributor PII. Work happens on samples, staging schemas, and local/dev environments. PII-bearing recordings enter agent workflows only after the PII pass (0.6), or as team-member (operator) recordings with explicit internal consent.
- Irreversible actions are human-gated: payouts, bans, deletions, publishing quests to the fleet, sending anything external (David, labs, studios), git pushes to shared branches, prod DB migrations.
- Every brief names its data. An agent dispatched without its input data identified and reachable is a stalled brief — the Data & inputs column below is as binding as the acceptance test.
- Review is tiered like the work: Tier-1 output gets code review + acceptance-test verification; Tier-2 output gets a named human decision-maker; Tier-3 is human work with agents as researchers/drafters.
Tier definitions
| Tier | Definition | Review |
|---|---|---|
| 1 — AI-led | Verifiable output, contained blast radius, prior art to lean on; agent works to an acceptance test | Code review + test verification |
| 2 — AI-drafted, human-decided | Agent produces candidates/drafts/analyses; a named human makes the call | Decision by owner |
| 3 — Human-led | Relationships, legal sign-off, money, gates; agents assist with research and drafting only | n/a |
Executor legend
- CC — Claude session in this workspace (interactive, cross-repo, good for research + scaffolding + docs).
- OC — OpenClaw / autonomous coding agents against a single repo with CI (good for long-running, test-driven implementation).
- DR — deep-research agent fan-out (web + papers, cited output).
- Dr. R — the standing demand-side researcher agent
(
ec-workspace/.claude/agents/world-model-researcher.md, added 2026-07-08): a skeptical world-model-eval scientist persona that issues Research Requests (RR), Format Specs (FS), Demo Triage verdicts (DT), Pre-registration Reviews (PR), and Literature Watch (LW). Every demo, dataset, quest, and pre-registration passes Dr. R before humans spend money on it. First output:researcher-dt-001-2026-07-08.md— POC graded WOULD INTRIGUE; mandatory third arm (equal-hours public-VPT commodity control) required to reach WOULD CONVINCE. - H — human owner (named where known from the 2026-07-07 debrief).
The map
Phase 0 — Foundations
| Item | Tier | Executor | Data & inputs needed | Acceptance test |
|---|---|---|---|---|
| 0.1 Capture-spec doc + client conformance checks | 1 (spec draft + test code) / 2 (spec decisions) | CC draft → OC tests; H decides | Current capture pipeline docs (pure-v2), sample recordings + _input.csv | Sync-verification tool measures alignment on a sample session; spec ADR merged |
| 0.2 Age verification | 3 (vendor/flow) / 2 (integration code) | H picks vendor; OC integrates | Vendor API docs; payout-flow code (contributor portal) | Staging: unverified account cannot reach payout |
| 0.3 Consent v-next | 3 (language) / 1 (plumbing) | H + counsel; OC extends consent-versioning | Current consent schema; SAG-AFTRA 2025 IMA text | Re-consent flow passes E2E test in staging |
| 0.4 EULA audit | 2 — AI does the reading, counsel decides | DR (Brief B-01) — RETURNED 2026-07-07: 0 green / 4 amber / 15 red; packet in ec-workspace/research/world-model-data-flywheel/b01-eula-audit-2026-07-07.md | Title list from playroll-cpp/res/games.json; live EULA/policy URLs | ✅ 19/19 classified with dated verbatim quotes; pending: counsel sign-off |
| 0.5 Provenance ledger | 1 | OC (Brief B-07) | Supabase schema (data model); consent-versioning tables | Sample export ships a machine-readable manifest; withdrawal propagates in staging |
| 0.6 PII pipeline | 1 (components) / 2 (error budget) | OC | Labeled sample of recordings w/ on-screen PII annotated (team-made, ~200 frames/title) | Precision/recall ≥ agreed bar on the labeled sample |
| 0.7 Anti-cheat vetting | 2 (dossier) / 3 (vendor contact) | DR — dossiers RETURNED 2026-07-07 (0.7-anticheat-dossiers-2026-07-07.md): all 4 Phase-1 titles clear for non-injecting capture; ER online last, after burner protocol; reusable 14-item checklist included | Title list + anti-cheat stack per title | ✅ Dossiers done; pending: EOS Help ticket for written EAC acknowledgment (H) |
| 0.8 Title-set selection | 2 | CC candidate analysis; H decides | 0.4 output + data-domain table | Frozen set with per-title rationale |
Track A0 — Klindt POC
| Item | Tier | Executor | Data & inputs needed | Acceptance test |
|---|---|---|---|---|
| Paper/repo survey | 2 | DR drafts; Antonio decides | David's shared links (posted in team channel per debrief) | Memo: chosen model + harness + gaps, signed off by Antonio |
| Eval harness (base model running + metric reproduced) | 1 | OC on dfgpu (Brief B-02) | GPU box (dfgpu); model weights (MineWorld / open Oasis); one published eval number | Reproduced metric within tolerance; one-command README |
| Benchmark probe candidates | 2 | CC/DR draft (Brief B-04); Antonio picks + pre-registers | Harness from B-02; prototypes/minecraft-spatial-tracer (RCON ground-truth state) | ≥5 candidate probes with metric, failure rationale, implementation sketch |
| Probe implementation | 1 | OC | Pre-registered probe specs; spatial-tracer bridge | Probes run against base model; failure cases reproduced with numbers |
| Quest authoring (2 quests + verifiers) | 2 | CC drafts via novelty-engine patterns; Marco/Vincenzo review | prototypes/novelty-engine (goals-as-code loop); requirements DSL | Verifiers pass on pilot recordings |
| Operator collection (~20–30 h) | 3 | H (Phil + operators) | Locked capture spec; quest definitions | Hours banked, verified, provenance-clean |
| Fine-tune + evaluate | 1 (execution) / 2 (reading results) | OC on dfgpu; Antonio/Marco interpret | Collected hours; harness; frozen recipe | Pre-registered report generated end-to-end by pipeline |
| Writeup + send to David | 3 | H (Antonio content, Alex relationship); CC drafts | All of the above | Sent |
Track A — Founding Experiment
| Item | Tier | Executor | Data & inputs needed | Acceptance test |
|---|---|---|---|---|
| Coverage atlas v1 | 1 | OC (Brief B-03) | Existing corpus sample: ≥50 h video + _input.csv across ≥3 titles, PII-safe subset; GPU for embedding | Atlas artifact + thin-cell report; stability check across embedder seeds |
| Eval-slice pre-registration | 2 | CC drafts; H (Antonio/Marco) freeze | Atlas cells; metric menu from B-02 harness | Dated pre-registration doc merged before collection |
| Directed-arm quest generation | 2 | Quest compiler (C1) + human triage | Thin-cell list; novelty engine | Every published quest passed A–E triage review |
| Arm randomization + collection ops | 3 (ops) / 1 (tooling) | H runs; OC builds assignment tooling | Fleet roster; lobby/quest attribution (recording spine) | Assignment audit: no contributor in both arms per window |
| Matched fine-tunes + evaluation + report | 1 | OC on dfgpu | Both arms' curated data; frozen recipe | Report generated by pipeline from pre-registered config, humans only interpret |
| Naturalness analysis | 1 (stats) / 2 (interpretation) | OC/CC | Directed + organic session sets | Metrics computed per pre-registration; report drafted |
Track B — Generalist-IDM
| Item | Tier | Executor | Data & inputs needed | Acceptance test |
|---|---|---|---|---|
| Unified event format + loaders | 1 | OC (Brief B-06) | Sample recordings + _input.csv per title; D2E's OWA format as reference | Round-trip tests pass; loaders overfit a tiny batch |
| IDM v1 training | 1 | OC on dfgpu | Labeled corpus (Phase-1 titles), contributor-split | VPT-bar accuracy on ≥3 titles, held-out contributors |
| Cross-game protocol + adaptation curve | 1 (runs) / 2 (claims) | OC runs; H approves published claims | Held-out title with withheld ground truth | Protocol report with zero/few-shot curve |
Track C — Engine Plumbing
| Item | Tier | Executor | Data & inputs needed | Acceptance test |
|---|---|---|---|---|
| C1 Quest compiler v1 | 1 (code) / 2 (every published quest) | OC extends prototypes/novelty-engine | Novelty-engine repo; DSL spec; A–E rubric as promptable checklist | Gap→quest pipeline emits valid DSL + verifier; human review queue works |
| C2 Coverage signatures + triggers | 1 (client code) / 2 (what ships in overlay) | OC in playroll-cpp/ui | Overlay/capture codebase; atlas cell definitions | Signatures computed on sample sessions ≤ size budget; triggers remotely versionable in staging |
| C3 Curation pipeline | 1 | OC (Brief B-05) extends data-filtering/pipeline | data-filtering repo; corpus sample; labeled quality examples | Dedup/quality precision-recall on labeled sample; provenance events emitted |
| C4 Fraud detection v1 | 1 (detectors) / 3 (enforcement) | OC | Input-stream corpus; synthetic fraud set (scripted/replayed sessions the team fabricates — Brief B-08) | Red-team session caught; false-positive rate ≤ budget on honest sample |
| C5 Influence scorer + annealing harness | 1 | OC on dfgpu | Track-A harness; per-clip features (atlas, verification, tier) | One calibration cycle: scorer ranking vs annealing verdict reported |
Phases 2–3 — Revenue & Compound
| Item | Tier | Executor | Notes |
|---|---|---|---|
| Pitch decks, one-pagers, data cards | 2 | CC drafts from experiment outputs; H owns the room | Every number traceable to pre-registered docs |
| Target-lab dossiers (who's hiring, what they train, warm paths) | 2 | DR refresh per quarter | Feeds Alex's outreach |
| Contract/indemnity templates | 3 | Counsel; CC produces briefing memos | — |
| Lab/studio/publisher relationships, negotiations | 3 | H (Alex) | — |
| Dashboards (campaign progress, coverage, unit cost) | 1 | OC in playroll-analytics | Real queries against staging views |
| Quarterly EULA re-audit, tripwire monitoring | 1 (scan + diff) / 2 (verdicts) | DR scheduled | Diff report per quarter; changes flagged to counsel |
Data inventory — what each stream of work is waiting on
The single consolidated list of data prerequisites; a brief cannot dispatch until its row here is green.
| Data asset | Needed by | Exists today? | Owner action |
|---|---|---|---|
Corpus sample: ≥50 h video + _input.csv, ≥3 titles, PII-safe | B-03 atlas, B-05 curation, B-06 loaders | Unverified (revised 2026-07-08): ~80–120 gross hours exist program-wide (CS2/OW2/Valorant dominate), but no hours metric exists in analytics, PII-safety is unproven (attacker-sim findings), game-audio tracks may contain voice chat (quarantine), and bug-16-era sessions lack trustworthy consent | Vincenzo: cut the sample with per-title hour counts, consent-era filter, and audio-track handling noted — this is now a mini-brief, not an export |
playroll-cpp/res/games.json title list | B-01 EULA audit, 0.7, 0.8 | Yes | none — readable now |
| David's shared papers/links | A0 survey, B-04 probes | Yes (team channel, per debrief) | Paste into the brief |
| dfgpu access for agent runs | B-02, all training briefs | Yes (data-filtering GPU box) | Provision an agent-scoped account/env |
| Model weights: MineWorld / open Oasis / Matrix-Game | B-02 harness | Public | Download on dfgpu |
prototypes/novelty-engine | C1, quest drafting | Yes | none |
prototypes/minecraft-spatial-tracer (RCON ground truth) | B-04/probes, POC verifiers | Yes (prototype) | Confirm it runs against the POC server version |
| Labeled PII frames (~200/title) | 0.6 PII pipeline | No | Team labeling session (~half day) |
| Synthetic fraud set (scripted/replayed sessions) | B-08 / C4 | No | 2–3 team members fabricate cheating sessions (~1 day) |
Quest requirements DSL live schema | C1, quest briefs | Yes (quests design + Supabase) | none |
| Operator roster + schedule for ~20–30 h | A0 collection | Partially | Phil confirms capacity |
The iteration loop for this page
Weekly (or per dispatch cycle): review returned briefs → check acceptance tests → merge or bounce → re-grade the map (what did this unlock? which Tier-2 item now has a harness that makes it Tier-1?) → cut the next batch of briefs → update the data inventory. The explicit ambition: each cycle moves at least one item down a tier or retires it, until the only standing human work is Tier 3 by nature, not by tooling debt.