Ars Inquirendi

AI-generated conjecture · below the evidence/publication boundary

← All conjectures · Philosophy, theology & the schools

Hard sayings are a small stock

Status: Anticipated · untested

Status is derived only from the shepherd-authored triage/prediction data above -- community submissions and claims are a separate overlay and can never change it (see the participation panel below).

This is a proposed connection between two domains, generated by a language model. It is not an article and not evidence: it sits below the evidence/publication boundary. A quantitative prediction and a named kill-dataset are attached (when registered) so the claim stays falsifiable rather than merely evocative.

Claim (verbatim)

Hard sayings are a small stock: in the scholastic classroom the objection-side authorities functioned as a memorized bank of classic difficulties — an objection had to be recognizable to master and audience to carry disputational force, so the same hard sayings of Augustine, Aristotle, and Jerome were recycled from disputation to disputation — while build-side authorities were drawn from whatever exemplars the master actually had open on the desk. Testing and building therefore ran on different inventories: the test inventory small, canonical, and repetitive; the build inventory broad and refreshed. This is a claim about concentration, orthogonal to whether the quotations are direct or mediated.

Prediction clause (verbatim)

In the 1,424 judged citation contacts, compute the share of contacts falling on the ten most-cited authorities separately for test-assigned and for build-assigned contacts. Primary clause: top-ten concentration among test contacts exceeds top-ten concentration among build contacts by at least 10 percentage points; the verdict follows this clause. Secondary clause: among authorities appearing in both roles, mean contacts per distinct cited work run at least 1.5 times higher in test than in build, i.e. testing re-hits the same loci.

Kill-dataset (verbatim)

Kill: the in-house Aquinas citation corpus (1,424 judged citation contacts with cited-authority identification and build-vs-test reading assignment); two grouped tallies decide it.

Nobody has run this test. The kill-data is named above. If you can run it — or you know the paper that already settles it — claim the kill or submit the prior. Kills and priors are credited here, by name, as they come in.

In the atlas

This conjecture is bridged, as an L1 lead, onto these Inferpedia subject pages.

Provenance

Run: Fresh agent generation · model: claude-fable-5

Generated by a fresh Fable-tier instance (claude-fable-5) at maximum effort with generation-first blindness (no repo reads, no web searches, no DB queries; single packet Write was the only tool use); 188-title list supplied at launch, titles only, no verdicts/dossiers seen; prompt pre-committed in docs/generated/conjectures_1001_wave_ledger.md and docs/generated/conjecture_fresh_fable_w02_prompt_20260708.txt. Wave W02 of docs/GOAL_CONJECTURES_1001_CAMPAIGN_20260708.md. Novelty unverified by construction. DEVIATION DISCLOSURE: after an output-limit resume, one accidental no-op placeholder Write to the session scratchpad preceded this packet Write; nothing was read and no information entered the generation; blindness intact, but the run used two Writes, not one.

Novelty / leakage triage

anticipated in the literature — this exact test has never been run

The disputational machinery — recognizable stock objections drawn from canonical authorities — is described qualitatively in the quaestio disputata literature, but no study measures differential authority-concentration between objection-side and solution-side citations, and the prediction runs on an in-house judged contact dataset. The test/build asymmetry as a measured quantity is un-run.

Predictions

Inconclusive registered 2026-07-11 calibration prediction (parent triage: leaked/adjacent)

Resolution: Inconclusive

Caveats: The registered identification guard fired in both groups — 31.3% of test contacts and 45.9% of build contacts lack an identified cited authority against a 20% ceiling — so the verdict is INCONCLUSIVE by my own rule. The point estimate is stated plainly: DIFF = +1.28 pp against a +10 pp threshold, and the secondary ratio 1.096 against 1.5x; had the guard held, this row would have been a clean kill on both clauses. Three honest observations about the number the guard suppressed. First, the missingness is asymmetric and almost certainly non-random: unresolved citations skew toward obscure authorities, so excluding ~821 of them inflates CR10 in both groups and inflates the noisier build group more, biasing DIFF in an unknown direction. Second, the crude authority normalization the operator flagged ('pseudo', 'john', 'school', 'marcus' as authority tokens, plus the 'vulgate' subset that escaped the scripture mapping) merges and splits authorities arbitrarily, and CR10 is exactly the statistic such token damage distorts. Third, and against the conjecture's framing rather than its measurement: both groups sit at ~94% concentration, so a +10 pp difference was close to arithmetically unreachable at this ceiling — the registered threshold was poorly matched to a corpus this top-heavy, which is an authoring flaw of mine, not a data flaw. A better build with resolved authorities and the vulgate mapping fixed would likely still show a small DIFF, and this conjecture would then need to survive an honest re-registration at a threshold that can actually discriminate.

Public prediction (kept in substance): 'compute the share of contacts falling on the ten most-cited authorities separately for test-assigned and for build-assigned contacts. Primary clause: top-ten concentration among test contacts exceeds top-ten concentration among build contacts by at least 10 percentage points; the verdict follows this clause. Secondary clause: among authorities appearing in both roles, mean contacts per distinct cited work run at least 1.5 times higher in test than in build, i.e. testing re-hits the same loci.' Restated for THIS corpus as it exists: the public text says 'In the 1,424 judged citation contacts'; the corpus now holds 2,245 judged contacts across the three host works, and this registration binds to the full 2,245 — the count is the only clause the corpus's growth touches, and it is registered as superseded. The build-vs-test reading assignment is taken per contact from the host packet's build_or_test field, as recorded, with no reassignment. TOP-TEN OPERATIONALIZATION, fixed before computation (the public text is ambiguous between a global and a per-group top ten): the ten most-cited authorities are computed WITHIN each group separately — i.e. each group's own CR10 concentration: for group g in {test, build}, rank authorities by contact count within g, take the top ten, and let C_g = (contacts in g on those ten) / (all contacts in g with an identified cited authority). Ties at rank 10 are resolved by including, among the tied authorities, the alphabetically first by recorded authority name until exactly ten are held. Contacts with unidentified cited authority are excluded from both numerator and denominator; contacts with missing build_or_test are excluded. Secondary clause operationalization: restrict to authorities with at least one contact in each role; within each role compute (contacts on that authority) / (distinct cited works of that authority contacted in that role), average across the qualifying authorities unweighted, and compare test to build. The corpus spans works to c. 1268 rather than the full career; the disputational build/test structure is present throughout, so the prediction applies to the available window as-is, with no truncation carve-out.

Resolution criteria: Let C_test and C_build be the within-group top-ten (CR10) concentrations defined in the statement, and DIFF = C_test - C_build in percentage points. SUPPORTED: DIFF >= +10. KILLED: DIFF < +10. The verdict follows this primary clause at the public threshold, unsoftened. DATA GUARDS (any triggers INCONCLUSIVE instead): build_or_test missing or unassignable on more than 20% of the 2,245 contacts; or either group containing fewer than 200 contacts with an identified cited authority (below which a ten-authority concentration statistic is dominated by a handful of loci); or cited-authority identification missing on more than 20% of contacts within either group. SECONDARY, reported alongside but never flipping the verdict: PASS if the unweighted mean contacts-per-distinct-cited-work among both-role authorities is >= 1.5 times higher in test than in build; declared 'not evaluable' (without affecting the primary verdict) if distinct-work identification is unavailable for more than 30% of the both-role authorities' contacts, or if fewer than 5 authorities appear in both roles.

Known priors disclosure: Shepherd-adjacent. The qualitative prior is real and was known to the shepherd at authoring time: the quaestio disputata literature describes the disputational machinery as drawing recognizable stock objections from canonical authorities. But no study measures differential authority-concentration between objection-side and solution-side citations, and the build/test asymmetry as a computed quantity is un-run anywhere; the 10-point and 1.5x thresholds are the shepherd's authored magnitudes. Additional adjacency, disclosed: the build_or_test assignment itself comes from host packets produced by the in-house pipeline the shepherd oversees, so the grouping variable and the conjecture share a house of origin; mitigation is procedural — this registration, including the per-group CR10 operationalization, tie rule, exclusions, and guards, is fixed before any tally has been computed or seen.

Within each group separately, authorities were ranked by contact count and the within-group top-ten concentration CR10 computed over contacts with an identified cited authority; DIFF = C_test - C_build, threshold +10 pp. Secondary: unweighted mean contacts per distinct cited work among both-role authorities, test vs. build, threshold 1.5x. Registered guard: cited-authority identification missing on more than 20% of contacts in either group triggers INCONCLUSIVE.

Dataset: The same 2,245-contact judged corpus, split per contact by the host packets' build_or_test field as recorded: 1,432 test contacts and 813 build contacts.

computed 2026-07-11

Weigh in

No community feedback yet.

Add your take

Posted immediately (spam is removed). Community feedback is never an adjudicated verdict and never changes this conjecture's triage label or status above.

Working on this?

Sign in to claim this conjecture and let others know you're working on it.