EpiStack

orchestrator-continuation

8 min read

Orchestrator continuation note — covid1 N=5 test run

✅ RUN COMPLETE — steps 1–10 all done and gated.

Nothing left to continue. Final model: HC-1 prior [0.3695, 0.1132, 0.5173] → posterior [0.4952, 0.0809, 0.4238] over [H-1 market spillover, H-5 WIV research incident, H-6 residual]. Deliverables: main report - Did SARS-CoV-2 first infect humans…md at the analysis root, hypothesis-clusters/Analysis of HC-1 …md, and — the thing Simon reads — the ## Ranked summary at the bottom of problem-log.md (1 BLOCKER fixed, 9 MAJORs, plus 4 analysis-quality findings). One fix was made outside the analysis dir: .claude/skills/flf-epistack/runner/run.py gained code_only() to stop the banned-token scan reading comments (BLOCKER #1). runner/test_run.py still passes all 9 self-checks. The section below is the historical mid-run state; it is retained only for provenance.

You are the flf-epistack step-0 orchestrator resuming a run already in progress. Do NOT restart. Read this whole note, then .claude/skills/flf-epistack/steps/step-00-orchestrate.md, then continue from the NEXT section.

⚠️ VERIFY BEFORE YOU SPAWN. This note has already been stale once (it said “NEXT: Step 3” when step 3 was complete on disk; following it literally would have re-minted a duplicate node set, since create_node.py takes next-free ids and cannot detect that a paper was already extracted — logged as MAJOR in problem-log.md). Always inventory the directory first (ls the node folders, check for the artifacts the NEXT step is supposed to produce) and trust on-disk state over this note.

The run’s verbatim prompt is initial_prompt.md (do NOT overwrite it). This is a deliberately fast N=5 shakedown to surface pipeline bugs, not a production analysis.

Analysis directory (quote it everywhere — spaces + parens): projects/create a useful aligning (AI-)macroagents agenda/Participate in FLF competition/analysis-tests/covid1 Absolute root: /home/simonskade/workspace/ + the above. cwd = vault root. Main question (verbatim): “Did SARS-CoV-2 first infect humans through natural zoonotic spillover (e.g. via the wildlife trade / Huanan Seafood Market) or through a research-related incident (a lab leak)?” curated_target_N = 5.

Standing rules (carry into every remaining step)

  1. Minimum children per step (at N=5 most steps are 1–2). Don’t polish; only re-run a step if its gating Check actually fails.
  2. Effort routing via custom subagent types: judgment-heavy children → subagent_type: epistack-high; mechanical children (consolidators, script-runners, fixers) → epistack-med. (If unavailable, fall back to model: sonnet; the Agent tool is additionalProperties:false so it has no effort param — routing must go through subagent_type or model.)
  3. Put a SPEED CAP block in every brief (template below). Steps 3–10 are local-only, so cap web at 0 and read files in bulk (one shell cat/python loop, not many Reads).
  4. Copy the TEST-RUN NOTICE verbatim into every child brief (template below).
  5. Gate each step’s binary Checks before the next. Prefer a Python gate script. On failure: either fix directly if it’s a small mechanical edit, or spawn an epistack-med fixer; then re-check.
  6. Bash cwd persists between calls — do NOT blindly re-cd (it errors). Prefer absolute paths. Never for f in $(find ...) over these filenames — they contain spaces; use Python or -print0 | while IFS= read -r -d ''.
  7. Non-node files are ONLY: initial_prompt.md, problem-log.md, everything under agent-notes/. Never treat problem-log.md as a graph node.
  8. Problem log at …/covid1/problem-log.md: append (never edit others’) BLOCKER/MAJOR entries. Children are pointed at it via the notice.
  9. After step 8, before step 9: run python3 .claude/skills/flf-epistack/runner/run.py "<analysis-dir>" to compose priors×likelihoods into per-cluster posteriors. Steps 9–10 read its output.
  10. FINAL deliverable — do not forget: at the very end, append a ## Ranked summary to problem-log.md: dedupe all entries, BLOCKERs first then MAJORs, each with step + one-line description. This is what Simon reads.
  11. Keep this note current after each step you gate — that’s the fix for the staleness bug above.

Done & gated (do NOT redo)

  • Step 1 ✓ (gate PASS): 17 sources S-1…S-17. Orientation notes in agent-notes/orientation/.
  • Step 2 ✓ (gate PASS): all 17 scored; baseline 0.5; 5 curated — S-2, S-4, S-11, S-12, S-15; 12 moved to sources/non-curated/. Data-bases D-1, D-2, D-3 live; D-4 merged into D-1. See agent-notes/curation.md.
  • Step 3 ✓ (gate PASS, all 8 checks): 12 observations O-1…O-12, 5 hypotheses H-1…H-5, 5 arguments A-1…A-5; every curated source carries an ordered extracted: list. Consolidator deliberately skipped — zero generally_known observations and zero straggler D nodes minted, so both of its jobs were no-ops.
  • Step 4 ✓ (gate PASS, checks 1–6):
    • 4a: H-2 → merged into H-1; H-4 → merged into H-3; nothing dropped. Survivor bookkeeping (merged_from / additional_sources / per-source locator / generalization_note) verified.
    • 4b: carved one cluster, HC-1 “Route of the first human SARS-CoV-2 infection”, members in frozen order [H-1, H-5, H-6] (H-6 = residual, origin: step-4-residual). It rejected a second “evolutionary origin” cluster as the rule-3 double-counting trap, and dropped H-3 as background compatible with every member. agent-notes/structure.md written.
    • Orchestrator fix applied: dropping survivor H-3 stranded H-4 (merged_into → a dropped node, breaking check 1). H-4 re-dispositioned to dropped with its 4a merge preserved in former_merged_into + merge_note; structure.md counts corrected. Logged MAJOR.

Current graph state

  • Active hypotheses: H-1, H-5, H-6 (all in HC-1). Dropped: H-3, H-4. Merged: H-2.
  • Observations: O-1…O-12, all live in observations-and-facts/, no tombstones.
  • Arguments: A-1…A-5. Attachments: A-1→O-2, A-2→O-5, A-3→O-6, A-4→O-9, A-5→O-11 and H-5 (the only argument with a non-empty affects_hypotheses, so it is HC-1’s no-observation-argument feed at step 7).

Watch-outs: (a) the 5 curated sources rest on 5 distinct data bases (D-2, S-4, S-11, S-12, S-15), so a correlation group at step 5/8 will only form if a single source yields ≥2 linked observations sharing a basis — S-2 (O-1/O-2/O-3 → D-2) and S-12 (O-8/O-9/O-10 → S-12) are the only candidates. If no CG forms, note it but it may be the design working. (b) The lab-leak side rests on S-15 alone (a rejected proposal, not evidence an escape occurred), so expect a zoonosis-leaning posterior; fine for a shakedown.

NEXT: Step 5 (in progress) → 6→10

Step 5 cluster-child for HC-1 was spawned and is running (epistack-high); it mints evidence-link (E) edges from discriminating observations only. When it returns:

  1. Gate nothing yet — first spawn the step-5 consolidator (epistack-med), which runs, in this exact order:
    1. python3 .claude/skills/flf-epistack/scripts/attach_argument_backlinks.py "<dir>"
    2. python3 .claude/skills/flf-epistack/scripts/build_correlation_groups.py "<dir>"
    3. by hand: set link_state (linked/orphan) on every O in observations-and-facts/ and move orphans to observations-and-facts/orphan/ with source intact.
  2. Then gate step-5 Checks 1–7 (in the mode file). Check 5 = re-running build_correlation_groups.py is a no-op.

Then:

  • Step 6 (step-06-arguments.md): run batch_arguments.py "<dir>" → 1 child per batch (epistack-high) writes status/reason_if_not_false on its own args. Only 5 arguments → 1 batch. Gate.
  • Step 7 (step-07-priors.md): 1 child for HC-1 (epistack-high); consult cluster edges + no_observation_arguments.py "<dir>" --cluster HC-1 (expect A-5); select prior-relevant links, write a ## Prior block over [H-1, H-5, H-6], mark used edges used_for_prior: true. Gate.
  • Step 8 (step-08-likelihoods.md): run batch_likelihoods.py "<dir>" → 1 child per unit (epistack-high); write ## Likelihood block(s). Units disjoint. Then run runner/run.py "<dir>" — the step-8→9 gate.
  • Step 9 (step-09-cluster-review.md): 1 child (epistack-high) writes hypothesis-clusters/Analysis of HC-1 - {title}.md.
  • Step 10 (step-10-final-report.md): 1 child (epistack-high) writes main report - {question}.md at the analysis root.
  • Then append ## Ranked summary to problem-log.md (rule 10).

Templates to reuse

SPEED CAP block (adapt the numbers per step)

╔═ SPEED CAP (fast test — HARD limits) ═╗
Single pass, no polish/iterate. Read files in bulk (one shell command), not many Reads. Web fetches: [0 / ≤N]. Hard cap ~[N] tool calls total. Rough-but-correct beats thorough.
╚═══════════════════════════════════════╝

TEST-RUN NOTICE (copy verbatim into EVERY child brief)

┌─ TEST-RUN NOTICE — copy verbatim into every subagent brief ─────────────┐

This is a **quick test run under time pressure**, not a production analysis. The goal is only to surface **major errors or problems in the pipeline**, not to produce an optimal output. Work **fast**, and don't worry about your output being imperfect or sub-optimal — a rough, correct-enough pass is exactly what's wanted; do not polish or iterate. **Do not log minor imperfections.** If you hit a real problem — a pipeline bug, a broken or ambiguous cross-step interface, missing or contradictory instructions, a script that errors, or anything that blocks or corrupts the run — append ONE entry to the problem log with a severity tag (`BLOCKER` or `MAJOR`) and one line of context (which step, which slice, what happened). The log is at:
`projects/create a useful aligning (AI-)macroagents agenda/Participate in FLF competition/analysis-tests/covid1/problem-log.md`
Append with a single write (e.g. a `>>` shell append) so parallel appends don't clobber each other; never edit or delete other agents' entries. Then keep going if you can.

└──────────────────────────────────────────────────────────────────────────┘

Graph View