The Blind Test: Masking Company Names and Judging Facts Alone

Anyone can look at Quibi in 2026 and say the outcome was obvious. Hindsight is cheap. The harder question is whether an assessment process, given only the facts knowable well before a company's terminal event, and with the company's name stripped out, sorts the doomed from the durable. That is what our blind backtest was built to check.

This article explains how the test worked, what it showed, and, just as importantly, what it does not license us to claim.

The setup

Verdicts land in three buckets: called safe, called uncertain, or not called safe.

The blind backtest took 116 dead companies from the graveyard and asked the production judges to assess each one using only facts knowable 18 months before its terminal event. Identity was masked and context was time-boxed. The survivor arm was run twice, as two independent replicates.

The pairs were real head-to-head matchups: Aeropostale against American Eagle Outfitters at a November 2014 cutoff, Quibi against Netflix at April 2019, Convoy against Uber Freight at April 2022, Yellow Corporation against Old Dominion Freight Line at February 2022, Jawbone against Fitbit at December 2015, InVision against Figma at July 2022, Rite Aid against CVS Health at November 2023, Katerra against Skanska at December 2019. Sixty-six of the 116 dead companies had a matched survivor twin.

What the numbers showed

In the July 2026 blind backtest, 90% of the 116 dead companies received a negative verdict on masked pre-collapse facts, and not one received a called safe call. The remaining 9% landed in the middle bucket, and zero landed positive. On the paired subset of 66 dead companies, the pattern held: 89% negative, zero positive.

The survivor arm looked completely different. In the first replicate of the survivor arm, only 17% of 58 matched surviving peers drew a negative verdict and 24% drew a positive one, assessed July 2026. The second replicate on 54 survivors came out at 20% negative and 29% positive. Across the 53 survivors judged in both replicates, the two runs agreed on the verdict bucket 49 times out of 53, a 92% agreement rate.

Given the same masking, the same time-boxing, and companies from the same markets at the same dates, they treated the two arms very differently.

Could the models just be recognizing the companies?

A fair objection: a model might see "masked meal-delivery startup, San Francisco, 2015" and quietly recognize Munchery, then recall the ending. We ran an identifiability probe that tried to name each masked company, and split the results.

On the clean subset of 17 dead companies that the identifiability probe could not name, 88% still drew negative verdicts and none drew a safe call, assessed July 2026. The recognized dead subset came out at 90% negative, essentially the same. The clean survivor subset is only 5 companies, too small to lean on, so we claim recognition-independence only for the dead arm. That is an honest limit of the design, and we state it rather than papering over it.

The most interesting result was an accident

Ten of the companies we seeded into the survivor arm as healthy peers drew negative verdicts from the blind judges in both replicates: Chegg, Course Hero, Fisker, Fitbit, Getir, Flink, Nikola, Rabbit Inc., Sun Basket, and Unacademy. When we investigated, each of the 10 turned out to be a genuinely failed or gutted company that had been mis-selected as a healthy twin during test construction. The blind assessment flagged all 10 seeding errors, in both replicates, before any human noticed the mistake.

Fisker is the sharpest example: it was placed in the survivor arm as a peer for Canoo at a July 2023 cutoff, the probe did not identify it, and both replicates still returned not called safe. The facts alone carried the verdict.

The flip side of the readout matters too. The dead-arm miss list, blind judges saying positive on a company that was dead within 18 months, is empty.

What we do and do not claim

Our publishing rule is strict: we never say "we predicted these deaths," and we never declare an outcome for any live company. The supported claim is narrower and, we think, more useful: given only pre-collapse facts, the assessment separated the doomed from matched survivors, and it never once said called safe on a company that was dead within 18 months.

For live companies, everything forward-looking on WillItSurvive is published as a dated watch-item with a disclaimer telling the reader to verify. The blind test is not a crystal ball.

Questions this data answers

How was the blind backtest constructed? Dead companies were judged on facts knowable 18 months before their terminal event, with identity masked and context time-boxed, by the same production judges used across the assessment. Each was matched where possible with a surviving peer judged at the same cutoff with the same masking, and the survivor arm was run as two independent replicates.

Did the blind judges ever mark a soon-dead company as safe? No. Across all 116 dead companies in the July 2026 blind backtest, zero received a called safe verdict; 90% were negative and the rest fell in the middle bucket. The dead-arm miss list is empty.

Matched survivors from the same markets, judged at the same dates with the same masking, drew negative verdicts only 17% of the time in replicate one and 20% in replicate two, versus 90% for the dead arm. The two survivor replicates agreed on 49 of 53 shared companies.

Could the models have recognized the masked companies and recalled their fates? An identifiability probe tested this. The 17 dead companies the probe could not name were still 88% negative with zero safe calls, so recognition does not explain the dead-arm result. The unrecognized survivor subset was only 5 companies, so we do not claim recognition-independence for that arm.

What were the seeding errors and why do they matter? Ten companies placed in the survivor arm as healthy peers, including Chegg, Fisker, Nikola, and Getir, drew negative verdicts in both blind replicates and turned out on investigation to be genuinely failed or gutted. The blind process caught all 10 construction mistakes before humans did.

Assessments reflect evidence available July 2026. Companies change; treat watch-items as things to verify, not conclusions.