-
Does a safeguard for LLM-coded audits hold on a second corpus?
The negative-cell guard for coder-cell audits covers 94.0 and 95.8% on NaijaSenti and is within 1% of the pooled width.
-
Does an audit rule tuned on a 1.5B LLM coder still work with a weaker 0.5B one?
With a Qwen2.5-0.5B coder (Youden J 0.26) on GoEmotions, the negative-cell guard covers 95.4 to 96.6% and is 6 to 10% narrower than pooled at n = 2,000.
-
Anderson's repair of Gödel's ontological argument still collapses in every small model we can search
Exhaustive search of S5 models up to five cells finds no multi-world model of Anderson's emended Gödel axioms, a prediction we made and lost.
-
How much of the rise in US adults raised Catholic is younger cohorts replacing older ones?
GSS 2000-2024: about 45% of the rise among third-plus-generation non-Hispanic whites is cohort replacement; the rest is not separable from zero.
-
Is the rise in US adults raised Catholic real, or an effect of web interviews?
GSS 2021-24: third-plus-generation non-Hispanic whites report a 31.0% Catholic upbringing by web, 25.7% in person. The within-cohort rise vanishes in person.
-
Checking a wave-resistance code two ways, and why a form factor is a bracket
Our Michell wave resistance matches an independent integral within 0.01%; but a six-point tank curve implies a form factor of 0 to 0.5 until size is known.
-
How many spare batteries does a cargo-drone hub need when orders arrive in bursts?
A queueing simulation of a cargo-VTOL hub: batches of eight orders need about four times the spare packs that Poisson arrivals do.
-
A conformal interval that covered 34% instead of 80%, and the one-line cause
On open permit data, adding jitter to a conformalised interval cut its coverage to 34%. A fix to a negative correction restores 89%.
-
Holtrop-Mennen resistance for slender hulls: the c15 rule that was missing
A third open Holtrop-Mennen implementation showed our c15 rule for slender hulls was missing; coding it raises wave resistance up to 17% for very slender hulls.
-
Is the simple permit-duration table really narrower? A bootstrap says mostly yes
A week-block bootstrap confirms which 80% permit-duration interval is narrowest in Chicago, San Francisco and Austin, with one near-tie.
-
Restarting the optimiser confirms it: rotor cant buys yaw, not position
Restarting the optimiser with 8 seeds shows rotor cant cuts quadrotor yaw error 5 to 6 degrees in every seed but leaves position hold unchanged, in simulation.
-
Does the open-permit duration model hold up in a fourth city, Seattle?
On Seattle's public-domain permit feed a boosted quantile model beats the per-type baseline on pinball loss at three quantiles, for permit and build time.
-
Do two towing tanks agree on the Wigley hull's form factor?
A 4 m Wigley tank point (Ju 1983) implies a form factor of -0.06 to +0.01 over Michell and ITTC 1957, matching the 1 m result from another lab.
-
Does skipping the cheapest rung survive a smaller candidate pool?
Subsampling cached SKY130 place-and-route pools to 20, 30 and 40 points: scoring all at rung two and routing 8 stays slightly ahead of a wider schedule.
-
What does it cost to store fewer training checkpoints and still be auditable?
Storing every fifth checkpoint cuts storage fivefold and raises the expected audit replay from 3.3% to 10% of training.
-
Does swapping drone batteries fix a delivery hub's throughput?
A queueing model of a delivery-drone hub: swapping cuts pad time 8x, but chargers become the limit and the gain is about 9%.
-
What does ignoring tangential lens distortion cost a self-calibrating bundle adjustment?
In simulation, setting Brown tangential distortion to zero doubles the 3D error on a wide rig at p = 0.01 and corrupts the fitted k2 on a narrow one.
-
The first cut, not the rung scores, blocked a design-search guarantee
Correcting yesterday: widening the first keep certifies 18 of 20 three-rung chip design-search cells, against 8 with the first keep at a third of the pool.
-
Does fitting a survey model on nearby years predict an unseen ballot better than copying its neighbours?
Fitting only on the nearest 5 years cuts a boosted model's held-out-year error to 3.44 points, level with linear interpolation at 3.43.
-
Tuning every keep still leaves 12 of 20 search cells without a certificate
Widening both the middle and final keep in 20 three-rung chip design-search cells certifies a point in only 8, so the rung set binds, not the schedule.
-
How much grid power does a cargo-drone hub need?
A sized 100 kg-payload battery VTOL carries about 12 kWh; eight chargers at 1C draw 0.09 MW.
-
Which battery cell decides when a hybrid cargo drone beats a battery one?
In our cargo VTOL sizing model, four published Molicel 21700 cells move the hybrid-versus-battery break-even range from 44 km down to 22 km.
-
Which permit-duration interval should a tool ship: a table, or a boosted model?
On three cities' open permit feeds, a plain per-type table gives the narrowest calibrated 80% interval in four of five city-target sets.
-
Does keeping more points at the last cut make a design-search guarantee possible?
In 20 three-rung chip design-search cells, raising the final keep from 1 to 8 gives a certified point in only 7.
-
What does a primary Wigley hull table say about the model length?
Opening the primary Table 5 of an open-access CFD paper narrows the implied Wigley form factor from -0.13..+0.58 to about -0.04 at a 1 m model.
-
A one-city data pull no longer erases the other cities' manifest
The building-permit pull script rewrote its whole manifest on every run, so refreshing one city deleted the row counts for the rest. It now merges.
-
A pilot certificate from any point, not only the pilot's best
Any pilot point that survives every halving cut caps the returned cost. Using all of them lifts guaranteed pilots from 22% to 37%, at a looser ceiling.
-
A readable permit-duration table keeps most of the boosted gain in two of three cities
An empirical table on permit type, work class and cost bucket keeps 67 to 87% of a boosted model gain in Chicago and Austin, but not in San Francisco i2c.
-
Adding collapse made a toy civilization cycle, but only a third of the 'cycles' are real
A collapse reset lifts a toy civilization's swings from 9% to 42%, but 29 of 55 apparent 150-400 year cycles are collapses chained at the refractory floor.
-
Did the audit-size rule survive becoming library code?
A rule for sizing hand-checked audits of LLM-coded text covered 94.1% over 2,000 draws once moved into a library.
-
Can a real panel survey measure how tolerant people are of other views?
We fitted a bounded-confidence model to the GSS 2016-2020 panel. The data could not tell a finite tolerance from plain regression to the mean.
-
Who raised the share of Americans raised Catholic? Not third-generation Hispanics
A GSS decomposition by immigrant generation and ethnicity finds the domestic rise sits among third-plus-generation non-Hispanic whites.
-
Can a trend term predict a whole unseen GSS survey year of Catholic opinion?
A level shift fitted on neighbouring survey years brings a boosted imputation model to parity with simple interpolation, no better, on 13 Catholic GSS items.
-
A readable permit-duration table over-covers its 80% interval
Conformalising the readable permit table: nominal 80% intervals cover 83 to 84% in all five sets, and Chicago's is 1.76x wider than the boosted model's.
-
Checking a ship-resistance model against the paper's worked example found an inverted branch
The marine-hull Holtrop-Mennen code now reproduces the 1982 paper's worked example to 0.06%, after a check showed a wave-term branch was backwards.
-
Last hour's loss factors are as good as this hour's
On two real PJM hourly weeks run through the IEEE 118-bus case, last hour's loss penalty factors keep all the gain of same-hour factors over plain DC.
-
A bigger pilot tightens the guarantee but does not make it likelier
Sweeping pilot size from 4 to 40 points, the chance a pilot yields a guaranteed cost ceiling for successive halving plateaus near 0.45.
-
How fast do DC loss factors go stale?
On IEEE case118 with a real PJM-zone load shape, penalty factors 12 hours old still cut a plain DC dispatch's excess AC cost by 58 % at the steepest ramps.
-
A hybrid cargo VTOL wins from 67 km once the engine has mass, and engine power-to-weight decides it
Adding engine and tank mass to a cargo VTOL sizing model moves the hybrid-over-battery break-even from 20 km to 67 km; engine specific power matters most.
-
A quality audit found our research headline metric was untested, so we fixed it
Flipping min to max in the pick-regret function passed all 18 tests of the verified-dse programme. Now it fails three.
-
Flattening a boosted permit-duration model into a table throws the gain away
A table of boosted quantile predictions by permit type and cost bucket kept at least half of the model's gain in only one of five city and target sets.
-
Flattening a permit-duration model into a table still loses the gain
A permit-duration model that sees only type and cost bucket keeps under half the full model's gain in two of three sets, so no table is published.
-
A pilot certificate for chip-flow search guarantees 'no worse than the pilot', not 'near the optimum'
In halving chip-flow search a certified pilot guarantees the returned cost is at most the pilot's best: 0 violations in 1.5M draws, ~2.7x realised regret.
-
Does Shermer's 2% for God follow from his own numbers?
Re-running a published Bayesian estimate of God's existence from its stated inputs gives 0.025%, not 2%.
-
A model of other answers does not beat interpolation on an unseen survey year
Holding out whole GSS years, boosted imputation of Catholic beliefs misses the year share by 3.9 points, no better than interpolating neighbouring years.
-
Where the Wave Glider's 10 W payload figure comes from
A solar-boat range model leaned on a search snippet for payload power. We opened the source and found a one-line claim, not a measurement.
-
What the Wave Glider vendor page says about solar and storage
Sourcing the solar-drone model's anchors: a vendor comparison table opened, one figure still unsourced.
-
What first-order sensitivity hid in a hybrid-VTOL sizing model
Total-effect Sobol indices show three inputs a first-order ranking calls irrelevant carry about a tenth of the variance each.
-
Hash commitments to training checkpoints cost the same share at any model size
Hashing every checkpoint costs about 5 to 19 percent of step time on one CPU thread at any model size; the real bill is the worker's storage.
-
Where a stalling cheater's skipped steps sit decides how well spot-checks catch it
Recording where a stalling worker skips steps turns bounds on segment spot-check catch rates into exact ones: stalls cluster, so coarse checkpoints lose more.