Mining Session — Master Orchestrator

This is the entry point for every alpha mining session. Read this skill first; it chains all other skills in the correct order.

Session Start Protocol

1. Read THIS skill completely before doing anything else.
2. Determine whether this is a cloud automation or local manual session.
3. Sync to latest main BEFORE branching (mandatory for cloud, strongly recommended for local):

   git fetch origin && git checkout main && git pull --ff-only

   This guarantees the experiment counter, the submitted book (data/book/), and the self-corr baseline are current. Skipping this is what caused duplicate rediscovery and a frozen experiment counter (every run re-branched from stale state and re-found the same families). If pull --ff-only fails, stop and surface the conflict rather than proceeding from stale state.
4. Assign a session ID (see format below) — derive NNN from the freshly-synced data/sessions/ so the counter actually advances.
5. Create the session directory now (Phase 0.5, do not defer to recording): create data/sessions/YYYYMMDD-NNN/meta.md with at least id, date, strategy (best guess, update later), and trigger. This guarantees an audit trail even if the session is interrupted. Append to it as you go.
6. Chain through the workflow in order. The workflow starts with context-gather (Phase 0), then applies the adaptive strategy selection (next section) to choose a mode, then continues through signal generation and beyond.

Adaptive Strategy Selection

After reading context (via context-gather skill), select a strategy using this decision tree evaluated top-to-bottom — first match wins:

READ data/knowledge/opportunities/  (ignore type: submit-candidate files —
                                     those are the submission queue, not ideas)
  └─ High-priority hypothesis/idea items exist?
       YES → HYPOTHESIS mode (test the opportunity)

CHECK: Has the agent run EXPLORE mode in the last 3 sessions?
  └─ NO → EXPLORE mode (novel templates and cross-family interactions)
       (Book is near saturation; novel structures are the highest-value use
        of budget. See data/knowledge/rules/novelty-required.md.)

CHECK multiple gate-passers from different families
  └─ 2+ gate-passers from UNEXPLOITED distinct mechanism families?
       YES → RECOMBINE mode (crossover blending of novel combos)

QUERY HF server for new gate-passers (via hf-server skill)
  └─ New gate-passer from a GENUINELY NEW mechanism family
     (not IV-spread, not analyst revision, not fundamental blend)?
       YES → EXPLOIT mode (directed mutation of the gate-passer)

CHECK existing gate-passers with BRAIN check failures
  └─ Gate-passer exists but fails one or more BRAIN checks?
       YES → REFINE mode (targeted fix for the specific failure)

DEFAULT → EXPLORE mode

Key change from prior behavior: EXPLORE and RECOMBINE now take priority over EXPLOIT. The book is saturated with known pattern families. EXPLOIT should only trigger for genuinely novel gate-passers from unexplored mechanism families — NOT for more IV-spread or fundamental-blend variants.

Workflow Chain

For every session, regardless of mode, execute these phases in order:

1. Context Gather — Read and execute .cursor/skills/context-gather/SKILL.md. This produces a state assessment and confirms the strategy.
2. Signal Generation — Read and execute .cursor/skills/signal-generation/SKILL.md with the chosen strategy mode. This produces candidate expressions.
3. Submit Candidates — Submit expressions to the HF server:

   uv run python3 scripts/hf_submit.py --expressions "expr1" "expr2" \
     --priority 5 --tags <session_tag> <batch_tag>

   Then poll the batch to completion with the canonical poller (do NOT write ad-hoc heredoc poll loops):

   uv run python3 scripts/hf_poll.py --tag <session_tag>

4. Result Analysis — Read and execute .cursor/skills/result-analysis/SKILL.md. This produces verdicts on each candidate AND, for any SAFE/RISKY candidate, labels it on the BRAIN platform via scripts/brain_metadata.py (the HF queue path leaves alphas unlabeled otherwise) and records a status: PENDING data/book/<id>.md entry.
5. Iterate or Stop — If a SAFE/RISKY candidate already meets the session's grade target, STOP and proceed to recording (satisficing — see AGENTS.md). Otherwise, if budget remains AND the improvement trend continues (at least one gate-passer in this round OR metrics improved over last round), return to step 2 with updated context (winners inform the next generation).
6. Record — Read and execute .cursor/skills/experiment-reporting/SKILL.md. This creates session files, updates knowledge, and opens a PR.
7. Cleanup — After the PR is created, the experiment-reporting skill's Step 6 handles switching back to main and restoring any stashed changes. Verify the repo is on main with a clean working tree before ending the session.

Cloud vs Local Differences

Session IDs

Format: YYYYMMDD-NNN where NNN is a zero-padded 3-digit counter.

To assign the next ID:

1. List existing directories in data/sessions/.
2. Filter for entries matching today's date prefix YYYYMMDD-.
3. Take the highest NNN found for today, increment by 1.
4. If none exist for today, start at 001.

Example: if data/sessions/20260603-001/ exists, the next is 20260603-002.

Stop Conditions

Terminate the session (proceed to recording phase) when ANY of these is true:

• Viable candidate found (DEFAULT for autonomous sessions): A SAFE or RISKY candidate exists that meets the user's grade threshold (e.g., EXCELLENT+). The session goal is to find ONE viable candidate, not the best possible one. Once found, proceed immediately to recording/PR. Only continue iterating if the user explicitly requested multiple candidates or specified a budget to exhaust.
• Budget exhausted: simulation count >= budget cap
• Diminishing returns: 3 consecutive rounds with zero gate-passers
• Family blocked: 3+ variants of the same expression family hit the SAME BRAIN check failure (e.g. CONCENTRATED_WEIGHT). Stop mutating that family and pivot to a structurally different template — do not keep burning sims on a structural block. (Lesson from 20260604-001: group_neutralize(IV spread) was EXCELLENT but always failed CONCENTRATED_WEIGHT + SUB_UNIVERSE across all variants; the zscore(ts_mean(...)) template broke through.)
• Exploration complete: all planned directions have been tested
• Explicit human stop: user says to stop (local mode only)
• Clock limit: cloud automation has been running > 45 minutes