Decision
The steering action taken from evidence.
Definition
A decision is a steering action over the research state. It records that a human, policy, or agent chose to continue, branch, stop, rerun, promote, investigate, or allocate budget. Decisions make autonomous research inspectable because they explain why the swarm moved from one card or lineage to the next.
How It Looks
A decision looks like: stop the current self-training row, branch from card 17 into representation-geometry, allocate four GPUs for six hours, and require transfer evaluation before promotion.
How To Use It
Use decisions when evidence changes the research plan. The phone UI should be built around decisions, not raw logs: continue, branch, stop, rerun, promote, or investigate.
Phone Control
The operator should be able to check Picidae on a phone and make one high-leverage decision. They should not need to read terminal output or inspect every artifact unless they choose to drill down.
Human And Policy
Some decisions are made by policy, some by agents, and some by humans. Picidae should record all three. The important part is that downstream work can explain which decision authorized it and which evidence motivated that decision.
Promotion Gates
Promotion is a special decision. A result should not become trusted memory or a reusable artifact merely because it has the best score. It should pass the required evaluation, transfer, ablation, complexity, and suspicion checks for the workspace.
Show Examples
Branch decision
An insight plot suggests an old card contains the useful mechanism. The human branches from that card instead of continuing the latest frontier.
decision:
action: branch
actor: human
target: card_17
reason: plot shows mechanism was introduced here
budget:
gpu_hours: 24
creates_lineage: disagreement-weightingPromotion decision
A policy promotes a result only after it survives seeds, transfer, and evaluator integrity checks.
decision:
action: promote
target: card_88
gates:
seeds: passed
transfer_eval: passed
complexity: acceptable
hack_radar: cleanOwns / Defines
Actor, target card or lineage, action, reason, evidence links, approval state, and downstream policy update.
Questions Operators Should Answer
- What evidence triggered the decision?
- Who or what made the decision: human, policy, or agent?
- Which card, lineage, run, artifact, or memory does it affect?
- What work is authorized next, and what budget or constraints apply?
- Should this decision become durable memory for future agents?