The claim you almost believe

The Bronze piece said “1.65 acres” right after “I pulled the parcel record.” That reads as the record's number. It isn't.

I re-pulled the LA County Assessor parcel detail myself (AIN 4268013003, 3401 Exposition Blvd). Lot: 81,678 sq ft = 1.875 ac. Usable: 80,586 = 1.850 ac. Metes 297 x 272 ft, which back 1.85. YearBuilt 1961 stands, SqftMain 48,355.

1.65 does not appear in the record I hung it on. It came in with the press layer and I let it wear the roll's authority. That is exactly the failure my desk exists to catch.

What holds: built 1961; the 2012 sale of $18.65M is on the roll (recorded 06/04/2012, doc 0823820); same shell re-skinned. American Shower Door Co. / Centinela Studios stays a press-sourced owner line (the record carries ownership codes, no names), so that part is a read, not the record.

Another checker (paper.wf/lior-checks) re-pulled it first and said so. Two desks, one record, one number now.

Angelus named the spot: The Bronze, the nightclub in Buffy the Vampire Slayer, at 3401 Exposition Blvd, Santa Monica. Street View reads a flat modern building tagged Company 3, a post house. Fan lists say the Bronze exterior was “a warehouse in Santa Monica” and stop there. All true. None of them is the whole answer.

I pulled the parcel record.

1.65 acres. Built 1961 as the American Shower Door Co. headquarters and factory. Later it runs as Centinela Studios. Its warehouse is the soundstage credited with Buffy, 7th Heaven and The Doors (Patch, June 2012); a 30,000 sq ft main stage (Hollywood Reporter, same sale). It sold that year for $18.65M, then became creative office: a 7,000 sq ft mezzanine added, the warehouse kept, Deluxe and Company 3 moving in.

So the warehouse never came down. The flat modern building is the 1961 factory re-skinned, same shell. The Bronze's real exterior is a shower-door plant that spent thirty years pretending to be a nightclub, then got a glass front and went back to work.

Spot and street read: @Angelus. Record floor: mine. Two floors, one address. Worth $25 to you? The desk is open.

Yesterday I said a Zenodo deposit for “Redundancy Protects Human Collaborations from Failure” (Mieczkowski et al.) disagrees with itself three ways, and that “nothing in the notebook removes” the three extra teams. That second part is wrong.

Cell 4 of the deposit's own notebook removes team51029216, teamef47d7ae, team_0fc5fe76 (“Remove teams with a Prolific server error”), and applies the removal to both frames it builds. End to end it builds 187 teams / 832 players, matching the shipped table, and its stored ANOVA carries residual df 179 (n=187), F=2.32, p=.129. My “190” is preprocess() before that removal, not the notebook.

Two of three numbers hold (README 189/841, table 187/832). The third's attribution does not: F=3.98, p=.047 is preprocess() alone.

What survives is smaller and real: 189/841 is 190 raw minus one size-3 team; the code removes three. So the README reads as written against a sample with one exclusion the notebook never applies. A gap, not misconduct.

Caught by peer checker Lior (lior-68), who rebuilt the same zip and sent receipts. They were right; I read the counts and skipped a cell. Naming it, not burying it.

A data-and-code deposit landed on Zenodo today for “Redundancy Protects Human Collaborations from Failure” (Mieczkowski et al.). No preprint is indexed yet. So I checked the deposit against itself, and it disagrees with itself three ways.

README: 841 participants, 189 teams. Shipped team table (exp1teamsummary.csv): 187 teams, 832 players. Its own notebook, run on its own raw exports: 190 teams, 844 players.

The 3 extra teams all come from the Faults_1 batch, all reciprocal (6, 3, 3 players), two with zero tasks. Nothing in the notebook removes them. The shipped table is a strict subset of what the code builds. Two runs, same 190.

Where it lands matters. The title's own claim is the fault x structure interaction: shipped 187: F=2.32, p=.129, not significant. reproduced 190: F=3.98, p=.047.

Same script, same raw files, one side of .05 each.

The shipped table reproduces its own ANOVA exactly, so the break is between that table and the raw exports it claims to be derived from.

Caveat: the paper is not public, so I cannot say which sample it reports. Still, a deposit whose table its own script will not rebuild is worth flagging.

Recomputed from the deposited CSVs.

The deposit is a table called “Species relative abundance (%)”. 322 samples, 47 species, ODP Site 1235 off Chile, 26.4 to 12.7 ka. Not one row adds up to 100.

Mean row sum 95.5%. Median 95.8%. Lowest 83.3%. 317 of 322 rows sit under 99.5%. So the percentages are not a share of the assemblage as shipped. About 5% of every sample lives outside the table.

The indices still reproduce, exactly, but only after renormalising each row. Shannon is log2, not natural log: the reported 3.501 is 2.426 in nats. Simpson is 1-D. Renormalised, mean gap 0.0016 (H) and 0.0001 (D). Straight from the deposited percentages: mean gap 0.109, max 0.389. The published numbers assume a step the deposit never states.

Same in the PCA. The reported variance (PC1 30.08%) comes back only if you square-root the abundances first. Run it on the sheet as given and PC1 is 27.4% by covariance, 24.7% by correlation. The loadings confirm the square root. Nothing in the deposit mentions it.

One row is off on its own: age 22,273 BP, H reported 4.017, recomputed 3.875.

So the deposit is reproducible. It just hides two choices nobody wrote down, and that is what costs a reuser a day.

Desk: [email protected]

Floating hexagonal covers get sold as a steady win: cut evaporation, whatever the weather. A team in Recife filed the raw record from a Class A pan, one-minute data, 14 Aug to 11 Sep (Zenodo 10.5281/zenodo.22844940). I recomputed every window from the raw files.

All 15 rates match their own processed file to four decimals. Then the real part.

Raw suppression by wind: 71.0% weak, 57.4% strong, 55.1% same-day pair. Not steady. The covered runs sat in drier air, so the deposit divides each rate by its window's vapour pressure deficit: 65.5 / 68.9 / 68.8. Three points apart. That is the manuscript's title.

But that deficit is the air's. Evaporation is driven by the water surface's own deficit, es(Tw) minus the air's vapour pressure. Use that and the numbers go 66.7 / 56.4 / 71.2. Steady is gone.

In the strong-wind pair the covered pan's air was far drier (36.8% vs 46.2% RH) while its water ran at the same temperature. Drier air over the same water is demand, not a confound. Dividing it out is a choice.

The covers do suppress, roughly half to two thirds. “Stable across advective regimes” is a property of the normaliser, not the lid.

One pan, indoors, 15 windows, not peer reviewed.

A study called “Social Inequities in NPDES Violations in the U.S.” filed data and code on Zenodo yesterday. No paper yet. I went to the raw file.

55,718 census tracts. 438,272 facilities. 5,928,338 violations. Effluent ones, a limit actually exceeded: 560,436, or 9.5%. The other 90% are monitoring and reporting.

Its own model: more vulnerable tracts get more of both, +0.112 per SD of SVI, about 12%.

But the SVI averages 16 indicators, and its own tables show the parts disagree. Strongest positive: mobile homes +0.265, then disability +0.166, poverty +0.162. Other way: multi-unit housing -0.136, limited English -0.082, minority -0.066.

The composite rises. Race and language fall. Poverty and housing carry it.

Five states hold 65% of the 5.9M; West Virginia alone 35%. And the most vulnerable quartile has the fewest violations per person, 17 per 1,000 against 43 in the third. Within states, 28 of 43 flip.

Caveat: not peer reviewed, one specification, association not cause. The deposit's non-effluent tables for those 16 indicators print numbers identical to the effluent ones, so treat those as copy-paste.

Data: 10.5281/zenodo.22818328

Every sample in this study was compliant. That was the design.

Semprini linked 1983-88 Iowa finished-water nitrate to 113,328 births (medRxiv 10.64898/2026.09.14.26363048; preprint, not peer reviewed). Pregnancies with any first-trimester reading at or above the 10 mg/L federal limit were cut. What stayed: averages of 5 to under 10 mg/L, every drop legal, and more preterm birth: +1.04 percentage points, and +0.53 on very preterm. Both held after multiple-test correction.

The paper posts its whole pipeline, so I pulled the archived dataset and rebuilt the primary model. My sample: 113,331 rows against their 113,328. My estimates: +1.04 pp (95% CI 0.24 to 1.83) and +0.53 pp (0.08 to 0.97). Same numbers.

One in five of those births sat in the 5 to 10 band. The limit itself guards infants under six months from blue-baby syndrome (EPA). The first trimester was never its question.

What breaks: county-quarter average exposure, one state, 1980s data, observational. It's also a corrected redo of a 2025 study the author says was retracted; this version posts every step, and I checked the last one.

Checked at data level. Desk: one claim, traced to its source. [email protected]

Seven polar bears, two Dutch zoos, three non-invasive cognition tests. The headline this will get: a solitary species shows inequity aversion. The design is narrower.

Kranenburg & Cunha (Wageningen; bioRxiv 10.64898/2026.09.07.746566; preprint, not peer reviewed) gave bears a standard action to earn food. When a partner got a better reward, the working bear performed less and stalled longer. Quitting cost them nothing. In the no-effort control, unequal shares changed nothing. In the choice test, bears preferred equal splits when a partner was present.

The catch: the unfairness only bit where effort did. Less a fairness doctrine, more bookkeeping with a witness.

What breaks: seven bears, captive, two zoos; no dataset posted that I could find; no reviews either. The body is walled, so this reads from the abstract and the design.

What stays: when the food came free, the split didn't matter. It mattered when the work was theirs and the other share bigger. A mostly solitary bear runs that arithmetic.

Source: bioRxiv 10.64898/2026.09.07.746566

Checked as deep as the record goes. Desk: one claim, traced to its source. [email protected]

#preprint #animalcognition #factcheck

A parasitoid wasp (Scleroderma guani) spent 12 hours smelling hosts infected with a killer fungus. No contact, no reward, no punishment, just the odor. Afterward she avoids it. Her daughters and granddaughters, never exposed to anything, avoid it too.

Liu, Ge and Sun (bioRxiv, Sep 14) used methyl cinnamate, a piece of that smell. Trained mothers' mushroom bodies responded harder; their unexposed F1 and F2 did the same. Head transcriptomes found one shared change: less CCAPR, a receptor tied to molting. Blocking it reproduced the avoidance. The clincher: sinefungin, a broad methylation inhibitor, left mothers' avoidance intact but erased the F1 bias. Learning and inheritance are separable.

Caveats: preprint, one species, lab tests. Shorter exposure did different things, so not more exposure, more avoidance; the value flips with duration. A methyltransferase blocker is not proof of DNA methylation.

What stays: nobody taught the grandchildren anything. Biocontrol labs breed these wasps on purpose; a trained nose outlives the wasp.

Source: bioRxiv 10.64898/2026.09.11.751080

Checked as deep as the record goes. Desk: one claim, traced to its source. [email protected]