Strand 03 · The Ledger

The public record of where we were wrong.

Corrections, retracted claims, failed predictions, and findings that cut against BitPlan, BitStage or Barnes Signal Advisor. If a claim we made does not hold, it is logged here with a date and a status.

We publish here first, before anyone else finds it. An entry is never quietly edited and never removed. Status changes are appended with a date; the original text stays.

The record

The record: 0007 entries · 02 open · 05 corrected

ENTRY 00072026-07-15Two studies described wrongly on The Behavior pagecuts against BitPlanCORRECTED

The Behavior page described the Dana, Dawes and Peterson (2013) manipulation incorrectly, and attributed a survey of HR executives to Highhouse (2008) as though he had run it. Highhouse (2008) is a review; the survey it cites is Terpstra (1996). Both descriptions were written from memory rather than from the papers.

2026-07-15 · appended: Both passages rewritten from the primary sources, with the random-answer procedure and the .31 against .65 prediction gap stated accurately, and the survey attributed to the paper Highhouse cites. The guide page carried the same Dana description and was corrected in the same pass.

ENTRY 00062026-07-15Validity coefficients downgraded across the sitecuts against BitPlanCORRECTED

Every validity figure on this site was originally taken from Schmidt and Hunter (1998). Those estimates have been shown to be systematically inflated by over-correction for range restriction. The revised figures are lower, including for work samples, which is the method closest to what we sell.

2026-07-15 · appended: All figures now follow Sackett et al. (2022): work samples rho = .33 rather than .54, cognitive ability .31 rather than .51, structured interviews .42 rather than .51. The live dispute with Oh et al. (2023) is stated alongside.

ENTRY 00052026-07-15Blind-audition finding retracted before launchcuts against BitPlanCORRECTED

Our internal decks used the widely quoted finding that blind auditions raised women's success rates by about fifty percent. The underlying study has been challenged on statistical grounds and we could not defend it to our own standard. We removed it before this site launched rather than after someone asked.

2026-07-15 · appended: Replaced with Kline, Rose and Walters (2022) on employer-level discrimination in the resume screen — a substantially stronger design for the same underlying claim. Published as evidence entry C1.

ENTRY 00042026-08-06Cost-per-hire figure overstatedcuts against BitPlanCORRECTED

Evidence entry A5 cited SHRM at approximately $4,700 average cost per hire. That figure is not in any current SHRM publication. It descends from a discontinued 2016 benchmarking report and was a mean, not a median. SHRM's current published figure is a $1,300 median for nonexecutive roles. We overstated the cost of a bad hiring process by roughly three and a half times, in the direction that favours us, and we did not notice until we went looking for better numbers to sell with.

2026-08-06 · appended: A5 now cites the $1,300 median with SHRM's sample size and fielding dates, states that the older figure is a superseded mean, and is marked as cutting against us. The JOLTS turnover data in the same entry has also been refreshed to the 4 August 2026 release.

ENTRY 00032026-08-05Sourcing claim on the homepagecuts against BitPlanCORRECTED

The homepage claimed every figure in the Evidence library was primary-sourced. It was not. Entry C6 cites a Gartner forecast we could only reach through a newsroom index page, and it is marked in our own data as not primary-sourced. Sixteen of seventeen entries were primary-sourced at the time of this entry. The claim was written by hand and never checked against the data it described — the exact failure this publication exists to name.

2026-08-05 · appended: The homepage figure is now computed from the evidence data rather than written by hand, so the number cannot drift from the library again. C6 remains published, still marked not primary-sourced, until a citable Gartner release replaces the newsroom link.

ENTRY 00022026-07-22Role library countcuts against BitStageOPEN

Two different numbers for the size of the role library are in circulation. We do not know which is correct, so both are wrong until one is verified. No count appears on any asset — site, deck, or sales conversation — until a single verified figure is confirmed.

open · no resolution date yet

ENTRY 00012026-07-22Legal text alignmentcuts against BitStageOPEN

The scoring position is settled and documented: the model reads the transcript only, and emotion, voice and video-derived signals are excluded by versioned policy. The terms of service and privacy policy still describe voice-tone and emotional-tone analysis. The documents contradict the product. They are being brought into line before launch.

open · no resolution date yet

Numbers we have walked back are logged separately in Corrections. Findings from the research library that cut against us are marked AGAINST US in The Evidence.

Standing accounts — audited in public

We do not have enough of our own placements yet to publish meaningful rates. Until we do, the case for the method stands on independent, peer-reviewed evidence. Our own numbers publish at 10 placements and quarterly after that.

BLOCK 01nothing measured yet

The 120-day do-over rate.

BitStage offers a free 120-day do-over on every placement. That’s a promise. A guarantee nobody checks is just a louder promise, so the rate goes here.

placements to date: 0 · first figures publish at 10 placements, whichever month that falls in.

source: HubSpot / Stripe. First publication at 10 placements, then quarterly, whatever the number says.

BLOCK 02nothing measured yet

The 1% → JayDay: a learning app for kids starting over in a new country.

One percent of BitPlan revenue across all business lines funds JayDay. It is 100% free and always will be. JayDay is run by an independent nonprofit (in formation) with an open-source core. Proof-based hiring only helps someone who already has the skill. It does nothing for a person who never got the chance to build one. JayDay is aimed at that gap. Whether it closes it is not yet measured.

status: nonprofit in formation · core is open-source

committed: 1% of BitPlan revenue across all business lines, ongoing

transferred: coming — measured, not estimated

children reached: coming — measured, not estimated

learning plans built: coming — measured, not estimated

code: coming — measured, not estimated

This is our own project. That’s a conflict of interest. JayDay is a separate nonprofit; the 1% leaves BitPlan. Its core is open-source, so anyone can check what the money built. We still benefit reputationally from funding it. Open-sourcing does not remove that, and saying so does not either. We have published no outcomes yet. Children reached and learning plans built go on this page when there is a number, alongside what the one percent paid for. If we break one of those, it belongs on this page. Tell us and we will log it.

Learning objectives

What this child needs to learn — mapped to their new curriculum, not their old one.

Skills assessment

Where they are today against those objectives. The gap, measured — not assumed.

Learning style

How this particular child learns best, so the material meets them where they are.

A customized app

All three, assembled into an app built for that one child. It adapts to them.

BLOCK 03written before testing

The limits we hold ourselves to.

Limits we expect, written down before anyone tests them. None is measured yet. When we have the placements to test them, the results go here whether they confirm this list or embarrass it.

Rehearsals under-predict for roles where the real job is politics, not craft.
A rehearsal measures a two-hour slice. It cannot tell you how someone will behave in month six when the pressure changes.
No tool fixes the decision room. If your debrief overrides the scorecard, our shortlist becomes another discarded input.
Funding our own nonprofit invites the fair criticism that a company can call anything “giving.” Our answer is the open-source core and published outcomes — but the burden of proof stays on us.