FLF EpiStack pipeline

A 10-step pipeline that turns one contested empirical question (COVID origins, LHC black-hole risk, egg health, …) into a navigable Obsidian knowledge base whose Bayesian conclusion recomputes from the notes. Built for the FLF “Epistemic Stack” competition. The graph is the artifact; the answer is a report over it.

You are invoked with a step number (0-10) or a short description. Match it to the table, then read that mode file — it is your full instruction set. Steps 1-10 each do exactly one step, in place, on a given analysis directory; step 0 is the orchestrator that runs the whole pipeline by spawning the others. Steps run 1→10; each consumes the earlier steps’ output and adds its own, never rewriting theirs.

step	mode file	does
0	steps/step-00-orchestrate.md	run the whole pipeline: brief and spawn each step’s children
1	steps/step-01-find-sources.md	seed the graph: candidate primary sources (a wide pool for step 2)
2	steps/step-02-curate-sources.md	score trust + usefulness per source and map its data-basis (2a, ~10 each); one selector reconciles Ds + flags curated ~N (2b)
3	steps/step-03-extract.md	extract each curated paper’s observation/hypothesis/argument nodes (~5 papers/child)
4	steps/step-04-structure-hypotheses.md	4a merge near-duplicates + drop off-topic; 4b cluster survivors into hypothesis-clusters
5	steps/step-05-link-evidence.md	one child per cluster: obs→cluster evidence-links where they discriminate; consolidator wires args + correlation groups
6	steps/step-06-arguments.md	assess each argument’s validity (batched)
7	steps/step-07-priors.md	select prior-relevant evidence-links (mark used_for_prior) + no-observation args, then write a prior block per cluster
8	steps/step-08-likelihoods.md	one child per correlation-group (joint) or per ≤3 lone edges: write P(obs|H) python blocks
9	steps/step-09-cluster-review.md	one human-readable review per cluster, beside its HC
10	steps/step-10-final-report.md	one main report answering the main question

Filepaths are relative to the skill folder (.claude/skills/flf-epistack/).

Node types

The analysis directory is a graph: each node is one markdown file (typed YAML frontmatter + prose body) in a per-type folder; edges are frontmatter wikilinks. An evidence-link note specifies an obs→cluster edge rather than a node in its own right. Two things are not nodes: initial_prompt.md at the analysis root, and everything under agent-notes/ — agents’ working notes, starting with agent-notes/orientation/, to which later steps may add subfolders of their own.

prefix	folder	type
S	sources/	source
D	data-bases/	data-basis — shared underlying data; observations resting on the same one are correlated
O	observations-and-facts/	observation (a.k.a. observations and facts: empirical findings and general facts about our world; henceforth just ‘observation’)
H	hypotheses/	hypothesis
HC	hypothesis-clusters/	hypothesis-cluster (mutually exclusive members)
A	arguments/	argument
E	evidence-links/	evidence-link
CG	correlation-groups/	correlation-group — joint-likelihood home for edges whose observations share a data basis (built by step 5’s consolidator)

Two output files carry no id/prefix: the per-cluster review lives beside its cluster in hypothesis-clusters/, named Analysis of HC-N - {rest of HC-N's title}.md; the single final report sits at the analysis root, named main report - {core research question}.md.

Shared conventions

1. The whole run turns on the main question, verbatim. It lives (with the FLF case framing) in the run’s initial_prompt.md — the orchestrator reads the pre-staged file, or writes it if handed the case inline; every child gets the question quoted in its brief. curated_target_N and the analysis directory are passed at invocation, not stored in that file.
2. Ids and filenames: every node has an id = PREFIX-N (integer counter, no zero-padding), fixed at creation, never reused or renumbered. Filename = PREFIX-N - Descriptive title.md; there is no title field. (The two report files above are the exception — no id, fixed-form names.) Filenames are immutable — every reference is a full-filename [[wikilink]], so renaming breaks inbound links; if a later edit makes a filename read wrong, fix the story in the body, keep the name.
3. Additive edits, clear ownership: steps mostly add fields to existing nodes; each mode file’s Interface says which fields it owns. Don’t overwrite another step’s fields.
4. Move, never delete: retired nodes move to a sub-folder (non-curated/, merged/, dropped/, orphan/), keeping links and content, so the record stays auditable.
5. Merges (steps 3–4): survivor = lowest id in the merge set; absorbed nodes become tombstones (merged_into → survivor, resolving in one hop, never a chain) and move to .../merged/; the survivor lists the absorbed nodes’ sources in additional_sources with a per-source locator list.