Corrections, changelog & verification

Last updated 26 July 2026

Everything on this page exists so the site can be checked rather than believed. It holds the public corrections log, the changelog, and the full working behind every claim that a named seat went over a federal limit.

Corrections log

No corrections yet. This site has not been published before, so the log opens empty. It will not stay that way, and entries will not be removed.

Verification of over-limit claims

Naming a commercial product as having exceeded a federal safety limit is the highest-stakes thing this site does, so every such claim was checked line by line against the source report before publication.

How it was checked

For each claim, three independent things were compared: the parsed reading; the testing lab's own Pass/Fail cell printed on that run's data sheet, re-extracted from the PDF specifically for this check; and the report's Section 2 summary of failed requirements. The applicable limit depends on configuration, so the dummy, orientation, and tether use were recorded for every run. FMVSS 213 sets HIC36 ≤ 1000, chest acceleration ≤ 60 g over a 3 ms clip, head excursion ≤ 813 mm untethered and ≤ 720 mm tethered, and a 70° rotation limit for rear-facing restraints.

Result

All 13 flagged seats are correctly flagged. Each has at least one run where the measured value exceeds the applicable limit and the lab marked that run Fail. There are no false positives: across the 3,169 runs the labs marked Pass, none exceeds HIC 1000, chest 60 g, head excursion 813 mm, or 70° rear-facing rotation.

The source reports contradict themselves

This is the most important thing the check turned up, and it shapes how the whole site presents data.

• A report's summary can omit a failure its own data sheets record. Diono Radian R120 (213-MGA-16-024), Recaro ProRide (213-MGA-16-051) and Diono Radian RXT (213-MGA-18-018) each carry per-run chest verdict cells marked Fail, while their Section 2 summaries list only labeling and registration-form deficiencies. A reader skimming the summary would conclude the seat passed its crash tests.
• A report can record a Fail against a measurement that is inside the limit. Cosco Apt 40RF (213-MGA-17-036), test 2: chest printed as 58 g, limit 60 g, verdict cell Fail, and Section 2 confirms "S5.1.2.1(b) Injury Criteria". The printed number and the printed verdict cannot both be right. This seat is not flagged here.
• A data sheet can record a failure the summary treats as paperwork. Cybex/Columbus Trading Aton Q (213-MGA-16-021): Section 2 lists only registration-form defects, but the test 1 data sheet marks head target excursion beyond the restraint's top and forward edge as Fail for the newborn rear-facing run. This seat is not flagged here.

Because the record disagrees with itself, this site does not publish a pass/fail verdict of its own. It shows the measured value, the lab's own verdict for that run, and a direct link to the report, and lets you read it.

The report-level Pass/Fail field is not a safety signal

Of the 25 runs marked Fail in the source data, several are paperwork: registration-card thickness outside the allowed range, incorrect capitalisation of "Federal Motor Vehicle Safety Standards", an extra hyperlink on an electronic registration form. A seat can carry a Fail for a typo on a label. That field is therefore never used here as a proxy for crash performance; the specific criterion and measured value are used instead.

Three flagged seats were later recalled for the same failure mode

The other ten seats: recalls exist for several, but for unrelated component defects (harness adjuster, buckle egress, webbing breaking strength, foam accessible at the crotch buckle, tether-anchorage shell integrity), and no NHTSA campaign was found for Britax Emblem, Chicco NextFit, Graco 4Ever Extend2Fit, Peg Perego Primo Viaggio 4-35, or Harmony Defender 360 — checked against the recall API and four independent curated recall lists, so those absences are findings rather than gaps in searching.

Two naming traps, since both are easy to get wrong: the recalled "Britax Marathon" is the ClickTight version, not the seat tested here; and the recalled "Graco Extend2Fit" is the Extend2Fit Convertible (a recline-label defect on one model number), not the 4Ever Extend2Fit.

Known unresolved items

• The Cosco Apt 40RF and Aton Q contradictions above are recorded, not adjudicated. Both seats are left unflagged, which is the conservative choice in each case.
• One run — Recaro Performance Sport, 12-month, tethered, head excursion 780 mm — sits above the 720 mm tethered limit while marked Pass. One instance out of 973 tethered runs; it turns on which configuration the lab treated as applicable.
• 17 reports underlying the over-limit claims were re-read in full. The remaining reports rely on the same parser, whose accuracy on that checked subset was the point of the exercise.

Parser accuracy audit

Every figure here was pulled out of a PDF table by software, and PDF table extraction fails quietly. So every report has been re-read by a second extractor, written from scratch, sharing no code with the one that produced the site, and every figure compared.

The first 100 reports were drawn at random with a fixed seed, so the early clean result was not a favourable slice; the remainder were then worked through to completion. The single disagreement is a chest reading of 43 g that report 213-MGA-19-022 prints and this dataset records as blank. It sits on a run using the weighted six-year-old dummy, a class excluded from every comparison here, so no published figure changes. Three further reports print a report number on their cover that contradicts the number they are published under (213-2360239-TEST, 213-MGA-17-050 and 213-MGA-17-065); those are typos in the source documents, and rows are keyed to the URL each report is served from. Head injury criterion, chest acceleration, head excursion and back-support angle are all compared, along with the testing lab's own pass/fail cell for each run. No run has yet appeared in the audit that was missing from the dataset.

Runs are matched on report number, sled test number and test number together. Either key alone collides: one sled run can produce two rows when a front and rear dummy share a sled number, and one report can contain two test series that both number their runs from 1. Matching on test number alone silently compares one run against another and manufactures disagreements that are not there.

Two bugs have surfaced so far, both in the audit tool rather than the published data: a section detector that matched the table of contents, and an excursion pattern anchored on the wrong side of the measurement. Anything the second extractor cannot read is counted neither way, since that is a gap in the audit and not an error in the data.

What a recorded "Fail" actually means

Of the 13 seats flagged on this site, three were later recalled for the same failure mode. The rest produced no public enforcement action at all. Neither outcome is visible in the test report itself, which is why every flagged seat links to its report and points you to nhtsa.gov/recalls for what happened next. NHTSA does not publish how often a compliance-test failure becomes a recall, so no rate is quoted here.

The two contradictions, resolved as far as the record allows

Both seats now carry a visible note on their row rather than being quietly decided either way. Neither is flagged as over-limit, because on the printed measurements neither is.

Cosco Apt 40RF — report 213-MGA-17-036

The chest figure for the 12-month rear-facing run is printed as 58 g against a 60 g limit, and the verdict cell beside it reads Fail. Checked at word and table level in the PDF: the table extractor returns the row ['58', 'Fail'], so this is what the document says rather than a parsing artifact. Section 2 independently lists S5.1.2.1(b) Injury Criteria as a failed requirement. Two of the three places in the document say the criterion failed; the printed number says it passed. No recall, Part 573 filing, inconsequentiality petition or docket entry ever followed — checked against NHTSA's recall data and the contemporaneous CPS master recall lists, which place the Apt 40 RF explicitly among Dorel models with no recalls.

Cybex Aton Q — report 213-MGA-16-021

The newborn rear-facing data sheet marks head containment (S5.1.3.2) as Fail, but the measurement cell is empty, the referenced remark is blank, and Section 2 lists only registration-form defects (card thickness 0.0130 in against an allowed 0.007–0.0095 in). The most likely reading is a clerical error on the data sheet. The Aton Q was recalled in 2023 (campaign 23C-001) for webbing breaking strength after abrasion, an unrelated issue from a different test.

Seat identity: merges made

The same physical seat appears under several model strings across reports, which splits its tests and corrupts both its median and its uncertainty range. These were merged. Each merge is listed rather than applied silently, because a wrong merge quietly averages two different products together.

Seat identity: merges deliberately NOT made

These look like duplicates but were left separate, because merging them could not be justified from the report metadata alone.

Seat category resolution

The source records assign a seat type per test run, and the runs for a given seat frequently disagree with each other. Previously whichever row happened to be read last decided the label. Each seat now resolves to one category by majority vote across its runs, with explicit overrides where the correct product class is known.

Category is used only for display and filtering. It is not the group the comparative figures are normalised within — that is the combination of dummy size and orientation. Changing a seat's category therefore cannot move its position.

Changelog

v0.4 — 26 July 2026 (pre-launch)

• Seats are now compared with each other. The site previously tested each seat against its group average with a correction for having asked that of all 240 seats, and concluded that nothing was distinguishable. That is the right test for screening outliers and the wrong one for a reader choosing between two named seats, which is a single comparison carrying no multiplicity penalty. Re-run pairwise, three of the four groups do separate — in the big kid group every best-tenth seat separates from every worst-tenth seat.
• Every row now states its standing in plain words ("clears 44% of this group", "no measurable gap either way"), and the few seats that clearly separate carry a badge. Banding follows the IIHS pattern: coarse buckets with most products honestly in the middle, rather than a league table that noise would not support.
• Adding two seats to Compare now returns a direct verdict on that pair.
• Single-test seats are included rather than excluded. A seat tested once carries a wider error bar, so separating despite it clears a higher hurdle, not a lower one. It keeps its "single test" tag.
• The infant group still says nothing separates, because nothing does: one distinguishable pair out of 28,680.
• Method simplified. The exact formula now sits behind a toggle, with a one-sentence plain summary above it, and the "how to read a row" text was rewritten for someone who has never seen a HIC value.
• Seat categories now respect what the model name says: a seat called "3-in-1" is an all-in-one regardless of how its rows happened to be labelled.
• The parser audit was extended from a 100-report sample toward all 698.

v0.3 — 26 July 2026 (pre-launch)

• Parser accuracy audit: 100 randomly sampled reports independently re-extracted, 794 figures compared, 0 disagreements. Published above.
• Sensitivity analysis for pooling the 213 and 213b-era benches: seat positions correlate 0.98–0.99 with and without the newer bench, and no conclusion moves. Published in the objections section.
• Added an objections section answering the six things a statistician or CPST would attack first, each with measured numbers rather than assertions.
• Explained that a recorded lab "Fail" is rebuttable prima facie evidence, not a finding of a violation.
• The two source-record contradictions now appear on their seats' rows rather than only in this log.
• Per-seat permalinks, so a specific seat can be cited.
• Same-lab head-to-head in the compare tray: when compared seats share a testing lab, their raw runs are shown with no lab correction applied, because for those seats the confound cancels. Explains why this does not extend to a within-lab ranking.
• Rewrote the opening to say plainly why the site exists, and settled on one voice throughout.

v0.2 — 25 July 2026 (pre-launch)

• Every reading now links to the NHTSA report it came from, on every seat — including the ~85% tested at a single lab, which previously had no per-run breakdown at all.
• All 13 over-limit claims verified line by line against source reports; working published above.
• Runs that went over a federal limit are no longer pooled into the reference distribution used to judge compliant seats. Such seats are marked "not on the comparison scale" instead of being plotted as though comparable.
• Seat identity merges applied upstream of the statistics, so merged seats have correct test counts and uncertainty ranges rather than correct names only. 433 name variants resolve to 420 distinct seats.
• Seat categories resolved by vote plus explicit overrides, replacing arbitrary last-row-wins.
• Groups where the data cannot distinguish seats now say so above the list. The infant group sits at 7% signal.
• Every seat row now states its position in words, not only as a dot position.
• Fonts self-hosted; no third-party requests remain.
• Fixed: the result list ignored its own row limit and rendered every seat at once, so "showing N of M" always displayed identical numbers.
• Added terms, privacy, and this page.

v0.1 — provisional dataset

• Initial collation of 699 reports, 3,764 sled runs, 2015–2026.