The pipeline

This is the specification the analyses were produced by — the flf-epistack Claude Code skill, published here so the method is auditable without cloning anything. It is the same text the agents actually execute, not a write-up of it.

The pipeline runs as one orchestrator plus per-step child agents. Each step below states what it consumes, what it must write, and the checks that gate it before the next step starts.

Read it

1. The skill definition — the entry point: node ontology, folder layout, the rules every step obeys, and the invocation contract.

The ten steps

1. Step 0 — Orchestrate the run
2. Step 1 — Find primary sources
3. Step 2 — Curate sources
4. Step 3 — Extract observations, hypotheses, and arguments from a paper
5. Step 4 — Structure hypotheses
6. Step 5 — Link observations to hypothesis-clusters
7. Step 6 — Assess argument validity
8. Step 7 — Estimate priors
9. Step 8 — Estimate likelihoods per (cluster, observation)
10. Step 9 — Per-cluster review of the analysis
11. Step 10 — Final report answering the main question

The code

Rendered on GitHub, which does code better than this site does:

1. runner/run.py
2. runner/test_run.py
3. scripts/attach_argument_backlinks.py
4. scripts/batch_arguments.py
5. scripts/batch_likelihoods.py
6. scripts/build_correlation_groups.py
7. scripts/check_missing_evidence_links.py
8. scripts/create_node.py
9. scripts/curation_select.py
10. scripts/no_observation_arguments.py

The runner (runner/run.py) is the piece that makes the answer recompute: it extracts the ## Prior and ## Likelihood python blocks from the notes, enforces determinism (arithmetic over named variables only — no imports, randomness, clock or I/O), and composes each cluster’s posterior. runner/test_run.py checks it against the numbers the step 7 and 8 specs publish.

Generated copy

These pages are generated from pipeline/ by scripts/sync-pipeline-to-content.mjs. Edit the source, not these. Wikilinks appearing in the step specs are syntax examples, shown verbatim rather than linked.