Step 6 — Assess argument validity

An argument carries no information of its own: it is a universally-valid logical/mathematical validity (theorem or inference schema). Being universally valid, it bears on our world only through what it attaches to — it MUST attach to ≥1 affects_observations and/or affects_hypotheses, and most arguments attach to an observation. An argument attaching only to a hypothesis (no observation) is rare — mainly math / theoretical-physics reasoning.

Step 6 judges whether each argument’s inference actually holds and tags it approved / corrected / rejected. It assesses every argument — both those attaching to an observation (affects_observations) and those attaching only to a hypothesis (affects_hypotheses). It mints no nodes and renames nothing; it only adds validity fields, editing a corrected argument’s statement in place.

Validity is judged on the argument’s own merits, conditional on its premises: assume the premises, ask whether the conclusion follows — independent of whether the premises or the conclusion are actually true. Their truth is priced later, as a likelihood (step 8) or prior (step 7); mixing it in here double-counts.

Step 6 runs as children over batches of ~20 arguments; the orchestrator batches every argument with scripts/batch_arguments.py <analysis-dir> and briefs each child with the rest of this file (see steps/step-00-orchestrate.md).

Relevant context

You read argument (A, step 3) nodes and write the two validity fields:

field	set by	meaning
statement	step 3	the one-line inferential claim; you edit it only on a corrected argument
reasoning (body)	step 3	the inferential content / calculation — the whole input to your judgment; you expand it to record your verdict’s reasoning
affects_observations / affects_hypotheses	step 3	what the argument attaches to (≥1 non-empty); you don’t change these, but they tell you what it feeds — step 8 for a with-observation argument, step 7 for a no-observation one

Interface

Consumes — from step 3: statement, the reasoning body, and the attachments affects_observations / affects_hypotheses. Source features (via source) inform only a trusted validity verdict.

Owns — status / reason_if_not_false on A nodes; edits statement on corrected arguments (never the filename); records each verdict’s reasoning in the body. Touches nothing else.

Feeds step 8 (likelihoods) — each with-observation argument’s status (include approved/corrected, exclude rejected) and its clarified reasoning body, read while pricing the observation it rides on. The clearer the body, the better step 8 can judge what probability to assign.

Feeds step 7 (priors) — each no-observation argument (empty affects_observations, attaching only via affects_hypotheses) contributes its status and clarified body, which step 7 consults when Fermi-estimating that hypothesis-cluster’s prior.

Feeds step 9 — reason_if_not_false flags which accepted arguments rest on unchecked trusted reasoning; it is an audit signal for the cluster review, not an input to any number.

Spec

Write onto each argument in your batch:

field	req	meaning
status	MUST	approved = the inference holds as stated. corrected = it holds only in a weaker/adjusted form. rejected = the inference does not hold
reason_if_not_false	MUST unless rejected	checked = you traced the load-bearing step yourself (skimming ≠ checked). trusted = tracing it is infeasible at reasonable cost (long derivation, specialist formalism), so validity is trusted from the source

A corrected argument is fixed in place: edit statement to the version that does hold, and preserve the pre-edit text under a ## Original heading in the body — so downstream reads the corrected claim directly. Never edit the filename, even when the correction makes it read wrong: it is the link target every other node reaches this argument by. A filename that no longer matches the corrected statement is the expected end state; ## Original plus the edited statement carries the true story.

Record your verdict’s reasoning in the reasoning body as you go — the traced steps behind a checked verdict, the specific failure behind a rejected one, or why a direct check wasn’t feasible and what makes the source credible for a trusted one. This body is what step 9 audits and what makes a trusted call defensible; a bare status with no trace is the main thing to avoid.

Rules

1. status is about whether the inferential step is real, not how strong the argument is. A valid-but-weak argument is still approved; how much it moves a number is steps 7-8’s job (keeping these apart prevents double-counting strength).
2. checked vs trusted is judged from the reasoning body, and length alone isn’t the trigger — a long but elementary chain is checked. Take trusted only where tracing the load-bearing step is genuinely infeasible.
3. A checked verdict is author-blind — it rests on the reasoning, never on author, venue, or citation count; source identity may inform only a trusted verdict.
4. rejected arguments are kept, not deleted — their links stay intact. “This step doesn’t go through” is a result; downstream simply excludes them.

Micro-example

Main question “What drove the extinction of Sahul’s megafauna around 45-40 ka?“.

1. corrected. A-101 “continental synchrony of the pulse with first arrival favours a human driver” → status: corrected, reason_if_not_false: checked. statement edited to “…synchrony is evidence for a human driver over a climate driver”; the filename still reads “…favours a human driver” and is left as-is; ## Original keeps the as-stated claim; body: “as stated the implication fails — a climate excursion that also tracks the arrival interval reproduces the synchrony; the evidential version goes through.”
2. approved + trusted. A-118 “the OSL age model’s systematic uncertainty keeps first occupation within ±2 kyr” → status: approved, reason_if_not_false: trusted. Body: “multi-page quantitative derivation, not traceable quickly; independently redone by a second lab, no published refutation.”
3. rejected. A-14 “no butchery site is known for 11 of the 12 genera, so humans did not drive the extinction” → status: rejected (no reason_if_not_false). Body: “needs an unstated premise that butchery sites would preserve and be discovered at a detectable rate; without it the conclusion doesn’t follow — a validity failure, independent of whether that premise is true.”

Checks (binary)

1. Every argument in your batch has status ∈ {approved, corrected, rejected}, and every non-rejected one also has reason_if_not_false ∈ {checked, trusted}.
2. Every corrected argument’s statement reflects the corrected claim and preserves the pre-edit text under ## Original; no argument’s filename differs from its step-3 filename.
3. Every argument carries a body trace for its verdict; spot-check 3 checked arguments — none rests on author, venue, or citation count.
4. No argument was moved, deleted, or renamed; no argument outside your batch was touched; affects_observations / affects_hypotheses are unchanged.

Process

The reliable method is reconstruct, then evaluate, kept apart — most bad verdicts come from judging an argument you never actually pinned down.

1. Reconstruct. State the premises and the conclusion explicitly, then name the single inferential step that carries the argument. Real arguments are usually enthymemes — a load-bearing premise is left implicit — so the core move is surfacing that hidden premise. Reconstruct it in its strongest reasonable form (the principle of charity), but stop short of completion by force: if the only premise that rescues the argument is one no one would grant, or one so strong it does all the work itself, that is a fault in the argument, not a licence to invent.
2. Evaluate the step, conditional on the premises. Assume the premises true and ask whether the conclusion follows. The productive probe is the undercutting defeater — an alternative that breaks the link between reason and conclusion without denying the premises (e.g. “the synchrony is equally explained by a climate excursion that also tracks the arrival interval”). No such defeater survives → approved. One survives but a weaker conclusion is immune to it → corrected to that weaker form; in practice this is the most common outcome — a “proves / refutes” claim that only really supports an “is evidence for”. The step needs a premise that fails completion-by-force → rejected.
3. Keep validity separate from truth. You are not asking whether the conclusion is true, or whether a premise is factually right — a premise’s truth is priced as a likelihood or prior in steps 7-8. Reject only when the inference fails; an argument with a false premise but a sound step is still approved.
4. checked vs trusted. Trace the load-bearing step yourself and mark checked whenever that is feasible — this is the default, and it must not lean on who wrote the paper. Reserve trusted for steps whose verification is genuinely out of reach at reasonable cost (specialist formalism, a multi-page derivation); there, say in the body why a direct check wasn’t feasible and what makes the source credible (independent replication, no published refutation). trusted is a cost concession, not deference — keep it rare.