Step 3 — Extract observations, hypotheses, and arguments from a paper

Step 3 fills in the paper-derived body of the graph. You are one child in a parallel fan-out; your slice is ~5 curated sources. Read each of your papers end-to-end, one at a time, and extract that paper’s observation / hypothesis / argument nodes (each resting on exactly that one source), then write an ordered extracted link list back onto the source node. No scoring, no cross-paper merging, no clustering, no observation↔hypothesis edges — those are later steps (4-8).

Relevant context

You extract from source (S) nodes — created in step 1, scored and curated in step 2. The fields you read (you never rewrite a step 1/2 field):

field	set by	meaning
type / id	step 1	source / S-N
url	step 1	open this to read the paper
source_type	step 1	enum (journal-article, preprint, dataset, …)
data_basis	step 2	[[D-N]]/[[S-N]] the source’s findings rest on; you read it and attribute each observation to the relevant subset
motivatedness	step 1	known agenda / COI — travels to your nodes via source
trust_score	step 2	[0,1] data-reliability; your O/H/A inherit it through source, never copied onto them

You ignore usefulness / combined_score (they served only the curation cut).

Interface

Consumes (from step 2): the curated: true source nodes (equivalently, those left directly in sources/), opened via their url; you process the ~5 in your slice. Each node inherits its source’s trust_score as data-reliability through the source link (never copied); motivatedness (a source’s known agenda / COI) travels the same link — its data-trust effect is already priced into trust_score, its argument-selection effect is handled at steps 5-6. Ignore usefulness / combined_score.

Produces for step 4 (cluster hypotheses): the single-source H nodes. Step 4 reads statement (+ question) to drop irrelevant hypotheses, merge near-duplicates across papers, and group them into mutually-exclusive clusters.

Produces for step 5 (link evidence): the O nodes, each with its data provenance attributed (the exact data it rests on — own experiment, external basis, or both — drawn from the source’s step-2 data_basis). Step 5 builds observation→cluster edges by reading the observations directly (step 3 leaves no bears_on hint), and derives each observation’s basis set — its data_basis override, else its source’s — to find correlated observations by shared node identity. An argument about a specific observation reaches its edge on its own: step 5’s script reads affects_observations as a backlink, so nothing here needs a placement hint.

Produces for step 6 (arguments): each A’s statement, reasoning body, affects_observations, and affects_hypotheses. Step 6 assesses each argument’s validity from the reasoning itself — so write the reasoning to stand on its own; an argument carrying only affects_hypotheses (no observation) feeds step 7’s prior, found via no_observation_arguments.py.

Spec

Step 3 creates O / H / A nodes and edits each processed curated source node (adds the extracted list and body ![[embeds]]; the source’s data_basis was set at step 2). It mints no judgment numbers — data-reliability is inherited from the source, and validity / likelihood / prior are later steps. Mint every id’d node (O / H / A / D) by piping its markdown on stdin — cat <<'EOF' | python3 /home/simonskade/workspace/.claude/skills/flf-epistack/scripts/create_node.py <analysis-dir> --type <type> --title "..." … EOF — with {{ID}} in the piped body only and never in --title, and the absolute script path so the call works from any working directory; never a hand-picked id — the script takes the next free PREFIX-N under a lock, so parallel children never collide (one counter per prefix, gaps are fine).

Every node you create rests on exactly one source. You hold ~5 papers, but you extract each in isolation and never merge across them — near-duplicate hypotheses from different papers (even two of your own) are left un-merged here; cross-paper merging is a later job — hypotheses at step 4, and generally_known observations at consolidation (below). source is singular and permanent; a merge survivor’s extra provenance goes into additional_sources only when nodes are merged, never while you extract a single paper.

Shared fields (observation / hypothesis / argument)

field	req	meaning
type	MUST	observation / hypothesis / argument
id	MUST	O-N / H-N / A-N
statement	MUST	~1-2 sentences; the filename title already names the thing, so statement adds the detail. Per-type meaning below
source	MUST	exactly one [[S-N]], pointing into a curated: true source
locator	MUST	section / figure / page / ¶ into that source
quote	OPTIONAL	verbatim snippet (worth adding for load-bearing or surprising items)

Longer prose — methodology, data detail, the argument’s reasoning — lives in the body (which is embeddable). Data-reliability is the source’s trust_score (a [0,1] reliability score set in step 2); step-3 nodes inherit it by reference through source and never copy it onto the node — nothing derived is cached in frontmatter.

Attributing each observation’s data

Step 2 set each source’s data_basis — the [[D-N]]/[[S-N]] bases its findings rest on — and minted those D nodes (a data-basis (D-N) node is the shared link target that lets step 5 detect correlated evidence by node identity, not by matching free text). Read that field, and per observation record the exact data it rests on: the source’s own experiment (self-link [[S-N]]), an external [[D-N]], or both. Reuse an existing base before minting; a needless duplicate hides real correlation, and duplicates parallel batches mint are reconciled by the step-3 consolidator. (Mint any external base step 2 missed.)

Per type

1. observation (node type “observations and facts”, folder observations-and-facts/) — records something taken to be true or actually observed about our world: an empirical finding or a general, uncontested fact about the world (a lot of general facts about the world count as observations here). It must say something true (or presumed observed) about our world specifically, not something that holds across all possible realities — so a purely mathematical fact is not an observation; encode that as an argument (a universally-valid inference you can apply) or, if it is a derived conclusion of interest, a hypothesis. A finding here follows ~with certainty; anything genuinely uncertain or contestable is a hypothesis, not an observation. (Henceforth this node type is called simply “observation”, id O.) Two kinds:

   • paper-derived (the default; leave generally_known absent/false) — a finding / measurement that follows from this source’s data. The body MAY add data detail (e.g. a small table of the dataset’s key properties — not raw data) and MAY ![[embed]] the source’s methodology section so method is written once. data_basis — the exact data this finding rests on, drawn from the source’s step-2 data_basis: the source’s own experiment (self-link [[S-N]]), an external [[D-N]], or both. Omit it to inherit the source’s full data_basis; set an explicit subset when the source is mixed and this observation rests on only some of its bases, so step 5 doesn’t over-count correlation.
   • generally_known (generally_known: true, data_basis: []) — a firmly-established, uncontested general fact or event about the world that the paper references but that rests on no dataset; evidence-like because its probability differs across the rival hypotheses, so knowing it discriminates between them (e.g. that humans were established on the landmass before the extinction pulse). It carries the paper it appears in as its source like any node, but resting on no data it is never correlated (step 5) and same-fact duplicates across papers are merged into one node at consolidation (below). Be sparing.

2. hypothesis — statement is an uncertain candidate answer that predicts observations. question (SHOULD) — the sub-question it answers; a useful hint for step-4 clustering, which groups mutually-exclusive answers to one question into a cluster.

3. argument — a universally-valid logical or mathematical validity (a calculation counts as one) that carries no information of its own: being valid across all realities, it bears on our world only through what it attaches to. statement = the overall implication / conclusion it establishes. reasoning (MUST, in the body) — the inferential content or calculation, free-form but detailed (small-stepped inferences welcome), with no required logical structure or premise/conclusion split. It attaches via two OPTIONAL [[wikilink]] lists, and at least one must be non-empty:

   • affects_observations — the [[O-N]] observations this argument bears on; step 5’s script attaches it to the matching evidence-link edge and step 8 reads it while pricing a likelihood.
   • affects_hypotheses — the [[H-N]] hypotheses it bears on directly, independently of any observation; step 7 reads these while setting the cluster’s prior.

   Most arguments attach to an observation (affects_observations non-empty). A paper that merely states a fact and connects it to a claim is usually giving you an observation (extract the fact as an O), not an argument — create an argument only when the connection itself is nontrivial: non-obvious, or carrying important subtleties. An argument attaching only to a hypothesis (no observation) is rare — mainly math / theoretical-physics reasoning, where a theorem applies to a hypothesis directly. Extract a paper’s observations and hypotheses before its arguments so the ids exist; every id you list must be an O/H extracted from this paper — don’t invent a cross-paper link.

Extract only observations, hypotheses, and logical / mathematical considerations. Exclude opinions, related-work summaries, and authorial recommendations.

Source note: extracted list + embeds

On each processed source node, in order of appearance in the paper — this ordered list is itself the paper’s structured summary:

1. extracted (MUST, frontmatter) — the ordered list of this paper’s O / H / A [[wikilinks]]; Bases-queryable.
2. body (SHOULD) — ![[embed]] each extracted node under section headings so the source note renders as the paper’s summary. Methodology MAY be written once here as a ## Methodology section and embedded into the relevant observation bodies.

Micro-example

Main question “What drove the extinction of Sahul’s megafauna around 45-40 ka?”, paper S-1:

1. O-14 — observations-and-facts/O-14 - Youngest reliable ages for 8 of 12 genera fall within ~2 kyr of first human occupation.md; source [[S-1 - Re-dated last-appearance ages for Sahul's extinct megafauna]]; locator “§Results, Fig 2”; data_basis [[D-4 - Continental late-Quaternary age compilation, group R]] (its ages come from that external compilation, not S-1’s own fieldwork); body: small table of the 12 genera + youngest ages, plus ![[S-1 - Re-dated last-appearance ages for Sahul's extinct megafauna#methodology|Methodology]].
2. H-3 — hypotheses/H-3 - Direct predation by arriving humans drove the extinction.md; source [[S-1 …]]; question “What was the dominant driver of the extinction pulse?“.
3. A-101 — arguments/A-101 - Continental synchrony of the pulse with first arrival favours a human driver.md; source [[S-1 …]]; locator “§Discussion”; affects_observations: ["[[O-14 …]]"] — it turns on the synchrony finding, so step 5’s script will attach it to O-14’s edge; detailed reasoning in the body. (An argument bearing directly on a hypothesis instead — e.g. a theorem constraining one member — would carry affects_hypotheses: ["[[H-N …]]"] and no observation; rare, mainly in math/theory papers.)
4. On S-1: extracted: ["[[O-14 …]]", "[[H-3 …]]", "[[A-101 …]]", …] (order of appearance); body embeds each node under ## Results / ## Discussion. (S-1’s data_basis was set at step 2; abbreviated … links stand for the full immutable filename.)

Checks (binary)

1. Every node sits in its type folder, has type ∈ {observation, hypothesis, argument} with a unique matching-prefix id, and a filename id - descriptive title.md; no id collides with a step-1 id.
2. Every node has exactly one source (into a curated: true source) and a non-empty locator.
3. Every observation statement is a ~certain finding (spot-check: no uncertain / contested claim sits among observations); every generally_known: true observation carries data_basis: [], and a paper-derived one leaves generally_known absent/false.
4. Every argument has a non-empty reasoning body, is nontrivial, and carries ≥1 non-empty attachment (affects_observations and/or affects_hypotheses); no node is an opinion, related-work summary, or recommendation. Every id in affects_observations resolves to an O node, and every id in affects_hypotheses to an H node, each extracted from the same paper.
5. No node carries method, kind, bears_on, premises, conclusion, topic, or section fields; no clusters, edges, priors, likelihoods, or validity scores exist (later steps own those).
6. Every processed curated: true source has a non-empty extracted list in which every id resolves to an O/H/A node whose source is that source; no curated: false source was processed.
7. Every observation data_basis override is a subset of its source’s data_basis; any straggler D node step 3 minted sits in data-bases/ with a unique D-N, basis_type in enum, and ≥1 inbound link.
8. generally_known merge (consolidation): every survivor carries merged_from + additional_sources; every absorbed duplicate is state: merged in observations-and-facts/merged/ with merged_into; no two live generally_known observations state the same fact.

Process

You extract from the ~5 curated papers in your slice, one paper at a time, applying the conventions above to each.

1. Read the paper properly — this is the deep read. Steps 1-2 only skimmed; now go end-to-end. Skim the abstract + figures for the thesis, then read Results and Methods closely (observations live there), then Discussion (hypotheses and arguments live there). Extract in order of appearance so the extracted list reads as a faithful summary.
2. The observation/hypothesis line is the crux. An observation is what the data essentially force — a measurement, a rate, a correlation actually present in the dataset. A hypothesis is any interpretation, causal claim, or extrapolation that could be wrong even if the data are perfect. Test each candidate: would a rival team analysing the same data still have to grant this? Yes → observation; if it turns on a contestable inferential step → hypothesis (and that step is usually an argument). The standard failure is promoting the authors’ preferred interpretation to an observation because the paper asserts it confidently.
3. Track which data each finding rests on. Papers routinely mix their own fieldwork with external compilations. Note, per finding, which dataset / instrument produced it — that is what lets you attribute each observation’s data_basis correctly against the source’s step-2 data_basis, so step 5 neither misses nor invents correlation.
4. Capture arguments as reasoning, not slogans. Write the reasoning body densely enough that step 6 can assess validity without reopening the paper — include the actual calculation for a mathematical argument — but don’t formalise into premises/conclusion. Skip trivial inferences (a bare fact-plus-connection is an observation, not an argument), opinions, and recommendations. Record what the argument bears on: the observations it turns on in affects_observations, any hypothesis it bears on directly in affects_hypotheses — at least one must be set, and most arguments carry an observation.
5. Right granularity, one paper at a time. One node per distinct claim: don’t shatter one finding into near-duplicate observations, and don’t bundle several findings into one. Extract each of your ~5 papers in isolation — do not merge or cross-reference across them, not even two papers in your own batch; cross-paper merging happens later, never here.
6. Write statements to stand alone. A reader who hasn’t seen the paper should be able to parse each statement and filename title: specific (numbers, taxa, intervals), self-contained, neutral. Give every node a precise locator; add a quote for anything load-bearing or surprising, for provenance.
7. Attribute each observation’s data. The source’s data_basis is set at step 2; per observation, record the exact data it rests on — the source’s own experiment (self-link [[S-N]]), an external [[D-N]], or both. Reuse an existing base before minting; a needless duplicate silently hides real correlation. (Mint any external base step 2 missed.)
8. Stay in your lane. No scores, clusters, cross-paper edges, observation↔hypothesis links, or validity / likelihood / prior numbers — later steps own all of those. Everything you write must be grounded in the single paper it came from.

Consolidate — merge generally_known duplicates, after every extractor has returned. A separate consolidation child does this (the same one that reconciles duplicate D nodes); an extractor does none of it. Two papers referencing the same generally-known fact yield one fact, not independent evidence, so merge near-duplicate generally_known observations into a single node with the full step-4 merge bookkeeping: lowest id survives and gains merged_from + additional_sources + a per-source locator list, each absorbed node gets state: merged + merged_into and moves to observations-and-facts/merged/, and any argument affects_observations pointing at an absorbed node is repointed to the survivor so its evidence isn’t stranded. Paper-derived observations are never merged: independent data is separate evidence, and shared data is handled as correlation at step 5, not by merging.