The portfolio optimizes for accepted progress per model call, not agent count. Accuracy remains an acceptance gate; duplicated opinions are purchased only when risk or contradictory evidence justifies them.
Run paid agent work in at most two project sessions at once. Run at most one high-assurance, full-forum, or UltraCheck workflow at once. The current primary portfolio is:
| Window | Project | Delivery objective |
|---|---|---|
| 1 | iOS | Automate the snapshot API and prove it on connected devices. |
| 3 | Nur | Give buyers enough verified information to make purchase decisions. |
| 4 | Persona | Build the strongest practical anti-tamper camera-detection stack. |
| 7 | Orchestrator | Improve the native routing and delivery system itself. |
| 9 | DC | Advance its own repository and Beads backlog independently of Nur. |
The remaining projects stay queued in Beads while two delivery slots are active. Rotate a slot after an accepted checkpoint, a typed device/operator blocker, or a no-progress stop. Do not keep five speculative forums running.
- Direct turn: one narrow edit, check, or device action. No forum.
- Quick forum: default for one substantial artifact. One owner gathers the evidence needed to advance it; one independent checker inspects the result criterion by criterion. No standing research panel. One targeted revision is permitted only after an actionable evidence-backed failure.
- Standard forum: explicit
standard forum; adds competing opening and verification perspectives and permits one revision. - Full forum / UltraCheck: explicit only. Use for contradictory evidence, critical security boundaries, or release-grade broad verification.
- Sweep: default 8 queued units and 4 weighted active slots. Explicit
large sweeppermits 24/6; an explicit 64-unit request permits 64/8.
Sweep's default final gate is one independent artifact audit plus a separate judge. Explicit high assurance adds a second coverage audit. Critical and high risk units still receive independent unit review; routine duplicate review is sampled instead of universal.
Every paid wave must produce at least one of:
- a scoped commit or artifact;
- a real passing or decisively failing check;
- a device measurement or captured trace;
- new primary evidence that changes the next action; or
- a typed blocker naming the smallest operator or infrastructure action.
A summary, repeated repository inventory, untested proposal, or returned agent does not satisfy the floor. Stop before another paid wave when the floor is not met. Keep the Bead in progress and checkpoint the exact blocker.
Use the Anthropic subscription for the native conductor and primary Opus work. Use the smaller OpenAI allowance for sampled engineering or independent verification, not every unit. Credit-backed cloud routes are eligible only after a live inference test succeeds; catalog visibility, an active agreement, or public availability is insufficient evidence.