EpiStack

CONTINUE-HERE

10 min read

CONTINUE-HERE — orchestrator resume note (eggs1 test run)

RUN COMPLETE (2026-07-20). All 10 steps finished and gated; the ranked summary is written at the bottom of problem-log.md (2 BLOCKERs, 14 MAJORs). Final artifacts: main report - Is habitual egg consumption...md at the analysis root and three Analysis of HC-N - ...md cluster reviews in hypothesis-clusters/. Nothing below needs doing — the rest of this file is retained as the record of what was planned and why.

You are the flf-epistack step-0 orchestrator, RESUMING a run already in progress. Steps 0–4b are done and gated (4b verified against Checks 2–4,6 on 2026-07-20: 13 actives, each in exactly one of HC-1/HC-2/HC-3, backlinks match, one residual last per cluster, no numbers; strain already logged as MAJOR). Read this whole file, then continue from step 5 through step 10 and the final ranked summary. Do not redo earlier steps or overwrite existing nodes.

First actions on resume:

  1. Read .claude/skills/flf-epistack/steps/step-00-orchestrate.md (your role) and this run’s source of truth initial_prompt.md (verbatim — do NOT overwrite it).
  2. Skim this note’s “Current state” and “Remaining plan”. Verify state on disk with the check commands before trusting it.

Run parameters (from initial_prompt.md)

  • Main question (verbatim): “Is habitual egg consumption net beneficial, harmful, or neutral for human health — overall, for whom, and at what level of intake?”
  • curated_target_N = 10.
  • Analysis dir: projects/create a useful aligning (AI-)macroagents agenda/Participate in FLF competition/analysis-tests/eggs1
  • This is a FAST TEST RUN (shakedown), not a production analysis. Optimize for finishing quickly and surfacing pipeline problems, not output quality. Use the minimum children each step needs. Only re-run a step if its gating Check actually fails.

Operating rules for the rest of the run (IMPORTANT — carry these into every child)

  1. Agent types: spawn judgment-heavy children as epistack-high, mechanical children as epistack-med (these exist as custom agent types with baked-in reasoning effort + speed discipline). Do NOT set a model override — let the agent type govern. All fan-out is yours; children never spawn children.
  2. Strong speed cap in every brief. Prepend this to each child brief: ”⚡ HARD SPEED CAP (from Simon): Work FAST, few tool calls, NO web unless truly unavoidable — work from the node files already on disk. Don’t polish or re-verify. Target the minimum tool calls.” The web-heavy re-reads (steps 1–2) were the ~850s/child cost; steps 5–10 are local reasoning and should be fast.
  3. Copy the TEST-RUN NOTICE (bottom of this file) verbatim into every child brief, appended at the end. It applies to you too.
  4. create_node.py gotchas (state these in every brief that mints nodes): (a) it needs the full node markdown, with a {{ID}} placeholder, piped on stdin (the mode-file one-line example omits this); (b) run it from vault root /home/simonskade/workspace OR pass the analysis-dir as an absolute path — a bare relative path fails from other cwds. Never hand-pick an id.
  5. Gate every step with its mode-file Checks before starting the next; on failure spawn a fixer child, then re-check. Move-never-delete: retired nodes go to sub-folders, not deletion.
  6. Problem log at <dir>/problem-log.md: children append BLOCKER/MAJOR only, single >> writes. It, initial_prompt.md, and everything under agent-notes/ (incl. this file) are the only non-node files — no check may move/delete them.

Current state (done + gated: steps 0, 1, 2, 3, 4a)

Verify with:

cd "<dir>"
ls sources/*.md | wc -l                 # 10 curated
ls sources/non-curated/*.md | wc -l      # 10 non-curated
ls data-bases/*.md | wc -l               # 10 live D (D-9 merged→D-4 in data-bases/merged/)
ls observations-and-facts/*.md | wc -l   # 20 live O (O-15 merged→O-3 in .../merged/)
ls hypotheses/*.md | wc -l               # 10 active H
ls arguments/*.md | wc -l                # 6 A
  • Curated sources (10): S-1, S-5, S-8, S-9, S-11, S-12, S-13, S-14, S-16, S-19. (Non-curated S-2,S-3,S-4,S-6,S-7,S-10,S-15,S-17,S-18,S-20 in sources/non-curated/.)
  • Data-bases (10 live): D-1 NHS, D-2 NHSII, D-3 HPFS, D-4 PHS, D-5 WHS, D-6 pooled-US (ARIC/CARDIA/Framingham/Jackson/MESA), D-7 PURE, D-8 China Kadoorie, D-10 EPIC, D-11 NIPPON DATA80. D-9 (dup PHS) merged→D-4.
  • Observations (20 live): O-1..O-21 minus O-15 (merged→O-3). Generally-known survivor: O-3 (egg cholesterol content, data_basis []). D-backed obs carry data_basis overrides: O-4,O-5→D-8; O-6,O-7,O-8→D-4; O-9,O-10→D-10; O-19→D-1/D-2/D-3; O-20,O-21→D-7. Self-basis obs (inherit source): O-1,O-2 (S-12); O-13,O-14 (S-1); O-16,O-17 (S-5); O-18 (S-8); O-11,O-12 (S-19).
  • Hypotheses (10 active): H-1 egg raises atherogenic lipids→CVD (S-12, surrogate); H-2 moderate intake protective for CVD (S-13, hard); H-3 high intake raises mortality esp. diabetics (S-14, hard); H-4 essentially neutral for IHD / inverse=reverse-causation (S-16, hard); H-5 poor bioavailability→weak effect (S-19, surrogate); H-6 2 eggs/day safe in T2D (S-1, heterogeneity); H-7 stable hypo-responder subset (S-5, heterogeneity); H-8 real but saturating lipid effect (S-8, surrogate); H-9 moderate intake not associated with CVD (S-9, hard-null); H-10 globally neutral (S-11, broad null). 4a made no merges/drops (all discriminable) — so the H-merge bookkeeping path is untested (fine; step-3 generally-known merge exercised the same convention).
  • Arguments (6): A-1 blinded crossover isolates egg as cause → affects O-1 (S-12); A-2 mortality-without-events→residual confounding → O-6,O-7 (S-14); A-3 attenuation on excluding first 4y→reverse causation → O-9,O-10,H-4 (S-16); A-4 dose-response strengthens causal reading → O-4,H-2 (S-13); A-5 population-average misleading for subgroups → O-16,O-17,H-7 (S-5); A-6 concavity→linear extrapolation overstates harm → O-18,H-8 (S-8). (Verify each A’s exact affects_* in the files before relying on it.)

Remaining plan (steps 4b → 10)

Step 4b — cluster (NEXT). One child, epistack-high, single (needs whole active set in view).

Brief it to read steps/step-04-structure-hypotheses.md (4b part only; take 4a’s active set as given, do NO merging). Carve the 10 actives into ~3 mutually-exclusive HC clusters. Case steering: keep surrogate-endpoint H (H-1, H-5, H-8) separate from hard-endpoint H (H-2, H-3, H-4, H-9, H-10); heterogeneity (H-6, H-7) its own cluster. Each HC: hypotheses (≥2, ascending id, residual last), exclusivity, exhaustive_by_construction (key present), independence, depends_on (key present, ideally empty). Add exactly one residual per non-exhaustive cluster (mint via create_node.py, origin: step-4-residual, statement names the escape). Set each active H’s cluster backlink. Write agent-notes/structure.md (inventory + counts + generated date). Known tension to watch (ask child to flag in its return): H-4/H-9/H-10 all assert ~null direction, so as separate members of one hard-endpoint cluster their mutual exclusivity is strained. This is useful pipeline signal — if it can’t be carved cleanly without merging, that points to a gap in the 4a-merge/4b-cluster division of labour (log MAJOR if real). Gate (step-4 Checks 2–4, 6): every HC in hypothesis-clusters/, unique HC-N matching filename, ≥2 active members ascending-id/residual-last, non-empty exclusivity+independence, depends_on+exhaustive_by_construction keys present; exactly one residual: true last per non-exhaustive cluster; every active H in exactly one HC with matching cluster backlink, no H in two clusters; no step-4 field holds a number.

Step 5 — link evidence. One child per HC (epistack-high, parallel), then one consolidator (epistack-med).

Per-cluster children read steps/step-05-link-evidence.md; each owns one HC and mints evidence-link (E) edges from every observation that DISCRIMINATES its cluster’s members (obs→HC, per-member breakdown indexed by HC.hypotheses order). They do NOT touch arguments or correlation. Slices disjoint by construction. Then ONE consolidator (after all cluster-children) runs, in order:

python3 .claude/skills/flf-epistack/scripts/attach_argument_backlinks.py "<dir>"
python3 .claude/skills/flf-epistack/scripts/build_correlation_groups.py "<dir>"

then sets link_state on every O (linked iff it’s the from of some edge, else orphan→observations-and-facts/orphan/). Expected correlation groups (obs sharing a data basis): {O-6,O-7,O-8}=D-4; {O-4,O-5}=D-8; {O-9,O-10}=D-10; {O-20,O-21}=D-7; plus each source’s self-basis obs pair. build_correlation_groups mints CG nodes over these — sanity-check they appear. Gate: python3 .claude/skills/flf-epistack/scripts/check_missing_evidence_links.py "<dir>" --cluster all (reports only) — spawn a fixer for any real (observation,argument) gap before proceeding.

Step 6 — assess arguments. One child per batch of ≤20 args (epistack-high, parallel).

python3 .claude/skills/flf-epistack/scripts/batch_arguments.py "<dir>" (reports only) → batches. Only 6 args, so likely ONE batch = one child. It reads steps/step-06-arguments.md, assesses each arg’s validity, writes status/reason_if_not_false (and for a corrected arg an edited statement + ## Original) on its own args only. Gate step-6 Checks before step 7.

Step 7 — priors. One child per HC (epistack-high, parallel; small).

Each reads steps/step-07-priors.md, slice = one HC. Consult the cluster’s evidence-links + its no-observation args:

python3 .claude/skills/flf-epistack/scripts/no_observation_arguments.py "<dir>" --cluster HC-N

Select prior-relevant links (base rates), write one ## Prior block pricing that cluster’s members, mark each used edge used_for_prior: true (so step 8 won’t double-count).

Step 8 — likelihoods. One child per unit (epistack-high, parallel).

python3 .claude/skills/flf-epistack/scripts/batch_likelihoods.py "<dir>" (reports only) → units, each either one CG-N (joint) or a batch of ≤3 lone edges; used_for_prior edges are excluded and listed as “already accounted for”. Pass each unit’s edge ids (assess) + write target (block_on: the CG node, or each edge note) into the brief. CG unit → one joint ## Likelihood block on the CG node; independent batch → one ## Likelihood block per edge note. No consolidator.

After step 8, before step 9 — run the model (gate):

python3 .claude/skills/flf-epistack/runner/run.py "<dir>"

Prints HC-N prior [...] posterior [...] per cluster. Catches wrong-length vectors, a cluster missing a prior/likelihood, or banned tokens. Fix before step 9.

Step 9 — cluster reviews. One child per HC (epistack-high, parallel).

Each reads steps/step-09-cluster-review.md, writes the human-readable review beside its HC: hypothesis-clusters/Analysis of HC-N - {rest of HC-N's title}.md (no id).

Step 10 — final report. One child (epistack-high).

Reads steps/step-10-final-report.md, reads the cluster reviews + posteriors, writes the single main report - {core research question}.md at the analysis root.

FINAL — ranked summary (your job, orchestrator):

Append ## Ranked summary to problem-log.md: dedupe all entries, list BLOCKERs first then MAJORs, each with step + one-line description. This ranked list is the main deliverable of the test run — Simon reads it to decide what to fix before the real N=25/50 runs.

Problem log state so far (all MAJOR, no BLOCKERs yet)

Already logged in problem-log.md:

  1. MAJOR — create_node.py snippet uses relative paths with no stated cwd; fails when run from the analysis dir. (step 1)
  2. MAJOR — briefs asked to tag each source surrogate-vs-hard endpoint, but no endpoint_type field exists in the source schema (added ad-hoc, inconsistent across pool). (step 1)
  3. MAJOR — data-basis granularity undefined for pooled multi-cohort studies (scope-note says “each cohort its own node” but S-7/S-9 treated as single pooled basis); choice changes step-5 independence counting. (step 2a)
  4. MAJOR — my orchestrator brief mislabeled S-14’s basis as “PHS/WHS”; S-14 is PHS-only (WHS is S-4). Scorer caught it and minted PHS-only, else a false step-5 correlation edge. (step 2a)
  5. MAJOR — create_node.py stdin requirement (piped markdown + {{ID}}) absent from mode-file invocation examples; multiple agents’ first mint call failed. (step 3 + recurring)

Watch for more at 4b (exclusivity strain) and step 5 (correlation-group correctness).

TEST-RUN NOTICE — copy verbatim into every subagent 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/eggs1/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