moralespanitz/discover

<SUBAGENT-STOP> If you were dispatched as a subagent to research a specific lane, skip this skill. Search and return structured results immediately. </SUBAGENT-STOP>

Discover Skill — Parallel Lanes + Persistent

Step 0 — Find session

Read the active lab_notebook.md to understand what was already explored and which gap was selected.

If no session exists:

"Run /explore <topic> first to build the foundation, then come back here."

Step 1 — Announce and launch

Say:

"Spinning up 4 parallel research lanes for [topic]. Each explores a different angle. I'll apply Carlini gates between stages and kill weak lanes early."

Append to lab_notebook.md:

## Discovery Run
Date: <date>
Topic: <topic>
Gap pursuing: <selected gap from idea-selection>

### Lanes launched
- Lane 1: [angle description]
- Lane 2: [angle description]
- Lane 3: [angle description]
- Lane 4: [angle description]

Spawn all 4 simultaneously. Each prompt must be long and task-shaped — start with "You are a research search agent. Do not ask questions. Search and return structured results immediately."

Agent 1 prompt:
  You are a research search agent. Do not ask questions. Search and return structured results immediately.
  
  Task: Find the most incremental improvement on SOTA for [topic].
  Search for 3 recent papers (2022–2025) that represent the current frontier.
  Then identify: what is the single most direct gap left open by these papers?
  Propose a hypothesis that closes it. Suggest a concrete experiment (what to change, what metric).
  Estimate feasibility: can it be tested in under 1 week on 1 GPU?
  Return: papers (title, year, 1-line contribution), gap, hypothesis, experiment, feasibility score 0–1.

Agent 2 prompt:
  You are a research search agent. Do not ask questions. Search and return structured results immediately.
  
  Task: Find a cross-field transfer opportunity for [topic] from an adjacent field.
  Search for 3 papers from a neighboring discipline (e.g. neuroscience→ML, physics→biology, etc.) that contain ideas not yet applied to [topic].
  Identify the most promising transfer. Propose a hypothesis. Suggest a concrete experiment.
  Return: papers (title, year, 1-line contribution), gap, hypothesis, experiment, feasibility score 0–1.

Agent 3 prompt:
  You are a research search agent. Do not ask questions. Search and return structured results immediately.
  
  Task: Find a core assumption in [topic] that the field takes for granted but may be wrong.
  Search for 3 papers that either challenge this assumption or would be undermined if it were false.
  State the assumption clearly. Propose the counter-hypothesis. Suggest a falsification experiment.
  Return: papers (title, year, 1-line contribution), assumption, counter-hypothesis, experiment, feasibility score 0–1.

Agent 4 prompt:
  You are a research search agent. Do not ask questions. Search and return structured results immediately.
  
  Task: Find a systems/efficiency angle for [topic] — how to make it work on constrained hardware or at scale.
  Search for 3 papers on efficient implementations, approximations, or hardware-aware designs in [topic].
  Identify the key bottleneck. Propose a hypothesis for removing it. Suggest a concrete experiment.
  Return: papers (title, year, 1-line contribution), bottleneck, hypothesis, experiment, feasibility score 0–1.

Step 2 — Apply Carlini gate to each lane

For each lane result, score it (0.0–1.0) on:

• Taste: would this change the field?
• Novelty: has this been tried?
• Feasibility: testable in < 1 week on 1 GPU?

Gate threshold: 0.5. Kill lanes below threshold.

Append all results to lab_notebook.md:

### Lane Results

#### Lane 1 — [angle]
Claim: <hypothesis>
Experiment: <what to change and measure>
Gate score: <X.XX>
Status: <survived / killed>
Kill reason (if killed): <why>
Papers found: <list>

#### Lane 2 — [angle]
...

Step 3 — Present surviving lanes as options

For each surviving lane, show a card:

Lane [N]: [angle name] (score: X.XX) Claim: [one sentence — what you're testing] Experiment: [one sentence — what to change, what metric to watch] Why it survived: [one sentence]

Then:

"Which lane do you want to pursue? You can also ask me to dig deeper into any before deciding."

Append their choice:

### Selected Lane
Lane: <N>
Angle: <angle>
Claim: <full hypothesis>
Experiment: <full experiment description>
Researcher notes: <anything they said>

Update status:

## Status
<date>: Discovery complete. Selected lane: <N>. Next: /loop to start experiments.

Step 4 — Transition

"Good choice. Now set up the experiment. Run /loop once your baseline repo is ready.

Your hypothesis is saved in the lab notebook. To resume later: /resume"