Right Direction, Wrong Magnitude

Status: Open

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.

A conjecture about this corpus itself

This is one of five meta-conjectures the generating model posed, blind, in the corpus’s final wave — predictions not about the pre-print world but about this collection of 1001 conjectures: how its verdicts would distribute, where its novelty would concentrate, whether its own governing architecture would survive its own data. Its kill-dataset is the corpus’s adjudication records, which is why it sits outside the ordinary triage ladder (“out of scope” below): it could only be judged once everything else had been.

Those records now exist, and on 11 July 2026 the model formally resolved this conjecture against them. Verdict: unresolvable as posed

Only 7 conjectures (10 prediction records; cj-079's three variants counted once, a flagged choice) have been resolved against kill-datasets — 0.7% of the corpus. Directional accuracy 4/7 = 57.1% (right: entropy-floor, precision-recession, zipf, yule; wrong: dealers-razor, universality-class, timurs-fingerprint), binomial P(X>=4 | n=7, p=0.5) = 0.50; numeric-threshold coverage is also 4/7 = 57.1%, so the predicted >=20-point direction-over-magnitude gap measures 0 points. At n=7 the stated binomial test can only reach p<0.05 at 7/7 (0.0078; even 6/7 gives 0.0625), so the primary clause was structurally undecidable on this sample.

The resolved sample is too small for the stated test in the strict sense: no achievable outcome short of perfection could have satisfied the p<0.05 binomial clause, and the 95% CI on 57.1% spans roughly 18-90%, consistent with both the conjecture and its negation. A second as-posed defect: this corpus's kill-clauses bundle direction and magnitude into single supported/killed verdicts (all 4 supported items satisfied their stated thresholds by construction; all 3 killed items failed direction itself), so the predicted dissociation — right arrow, wrong number — had almost no way to register in the records; it occurred in 0 of 7 cases. Honesty note: this verdict is NOT a flattering dodge — the point pattern runs against the conjecture (57.1% direction misses the 60% bar, 57.1% coverage exceeds the 40% ceiling, gap 0 not 20+), and had these proportions held at n>=40 the verdict would be killed. What would resolve it: 40+ resolutions with sub-clause scoring that separates directional sign from numeric-interval capture at adjudication time — a scoring change future waves should adopt. Counting all 10 records instead of 7 conjectures gives 7/10 direction (p=0.17) and 7/10 coverage; same conclusion.

All five self-judgments, with the full exhibit →

Claim (verbatim)

Every quantitative conjecture in this corpus makes two separable bets: a directional bet (which way the effect points) and a magnitude bet (the specific numbers and intervals named in the prediction). I conjecture that the finished corpus will display a characteristic calibration asymmetry: among conjectures that are actually resolved against data, the directional bets will be right well above chance, but the stated numeric thresholds and intervals will capture the measured values at a rate far below what naive confidence in the predictions would imply. The mechanism is that a model's directional intuitions are distilled from genuine structural regularities in its training corpus, while its specific numbers are largely stylistic confabulation — plausible-sounding precision generated to satisfy the prompt's demand for quantitative form. If this holds, it prescribes exactly how machine-generated conjectures should be consumed: trust the arrow, re-derive the number — and it makes the corpus itself a measured datapoint on the calibration of large models as hypothesis engines.

Prediction clause (verbatim)

Among all corpus conjectures adjudicated as resolved (falsified or confirmed against their kill-datasets), at least 60% will be correct in predicted direction, but at most 40% of the specific numeric thresholds or intervals stated in their prediction_text will contain or be satisfied by the measured value. Primary clause (the verdict follows it): the directional accuracy exceeds 50% by a binomial test at p < 0.05. Secondary clause: the numeric-interval coverage is at most 40%, and lower than the directional accuracy by at least 20 percentage points.

Kill-dataset (verbatim)

The One Thousand and One Conjectures corpus and its triage/adjudication records; kill is a statistical test (binomial test on directional accuracy and coverage-rate comparison).

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

Composed blind by claude-fable-5 with zero tool use, emitted as a single JSON text message per the fresh-lane blindness protocol.

Novelty / leakage triage

out of scope (a meta-conjecture about the corpus itself)

Meta-conjecture about the calibration asymmetry of this corpus's own predictions, resolvable only against the project's adjudication records; out of shepherd scope.

Predictions

No prediction registered yet.

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.