Before it tests your decisions, the test bench takes its own exam.

How a grade is earned

Five steps, every one of them readable in the code published below.

• Step 1. Real decisions from the past, with a known outcome and reactions documented by the press of the day. Each case is written ex-ante: nothing that followed the announcement appears in it, and the year of the case is a hard cutoff on the pipeline's documentary research. The model's own memory cannot be cut off: that is the stated limit of the exercise.
• Step 2. One answer key per case: every objection actually raised at the time, with a named source and a link that opens, checked by script.
• Step 3. The real production pipeline builds a panel of 30 voices and replays the case in independent passes. The number and the verdict come out of a computation; the language model provides the texture of the voices, never the math.
• Step 4. A judge distinct from the simulator (openai/gpt-4.1-mini, a different house from the voices' model), at temperature zero, classifies each objection as recovered, grazed or missed, under the instruction "never credit out of charity," with a readable justification per objection. The judge's model is written into every published artifact.
• Step 5. A case's grade is the average of its passes, published with its min-max range, never the best run. Both concordance tables are written in the code and published: the first, the verdict against whether the decision was kept, from the start of the exam; the second, against whether it actually succeeded, on September 26, 2026, after the replays. The two readings stay side by side, column by column.

The range is not a flourish. On Facebook's Libra project, the average is 75%, with passes running from 63 to 100: a single run could have said 100 as easily as 63. The average and the range always travel together.

Why the grade covers the process, and the verdict also faces how the story ended

The rule was written before the measurement

This rule governs the American corpus you are about to read. It was written and sealed before the corpus grew from 16 to 43; the 27 cases that completed parity all came in under it, or not at all. Five criteria, all mandatory: a US company decision publicly announced between 2005 and 2024, at least two national outlets from the period, at least three ex-ante objections attributable with a quote and a link that opens, an outcome classifiable without ambiguity twelve months later, zero overlap with the existing corpora. The cases excluded up front are named in the rule. Production runs in waves of 5, timed: 10 to 30 minutes per case, median 15.

If the documentation contradicts the expected outcome, the documentation wins. Applied in the very first wave: Qwikster, expected "withdrawn," reclassified "adjusted" by the sources.

Demand the same rule from any tool that wants to get near your decisions: written before the measurement, applied even when it stings.

Two exams, every scorecard

The American knowledge base behind the panel held 19 frames at the first US exam (July 27) and 33 as of August 5.

And one stratum kept at arm's length: 13 startup decisions never used to calibrate anything, 83% average coverage. That is the answer to "you calibrated on your own cases."

And the verdict itself, against what really happened. Against real success, 40 of the 43 US verdicts are right or within the accepted margin, 18 of them exactly; in France, 37 of the 41 cases that can be graded, 12 of them exactly. Exactly right means the expected verdict: adopt for a goal that was met, adjust for a goal met in part, hold off for a goal that was missed; the accepted margin allows the neighboring step, and both rules are published. We grade every verdict twice because a decision that was kept is not a decision that worked: Ÿnsect, a French insect-farming startup, stuck with its giant plant, and two years later it was still producing nothing. Seven verdicts out of 84 went the wrong way, and we name them. Three green lights (an "adopt" verdict) on missed goals: Google Glass, Ÿnsect, and free period products for French students (1,500 dispensers promised, a target later cut to 963). Four calls to hold off on decisions that reached their goal: Reddit's API pricing and Facebook's news ban in Australia in the US, Netflix and Bridgestone in France. Against whether the decision was kept, the count is 10 exact verdicts and 5 wrong-way ones in the US, 5 and 7 in France; two French cases stay ungraded, for lack of an established result. The success labels and their sources (JSON, with an English note for each case).

Here is every scorecard, the strong and the weak, one line per case, one table per market.

The US exam, replayed August 15, 2026 (decoupled judge)

The French exam, replayed August 15, 2026 (decoupled judge)

Generated straight from the run files you can download below (replays of August 15, 2026, with the decoupled judge openai/gpt-4.1-mini, written into each artifact). No line in it is written by hand.

WeWork and its $47 billion score 96%, range 88-100, in the table above.

The test bench also grades what almost nobody documents: the backing. 127 documented endorsements across the two corpora, about 65% recovered by the same strict judge, judgments published one by one.

The 121 objections the test bench raised beyond the answer key on the French corpus were triaged one by one: 116 hold up, 5 faulty, published with everything else.

We are the only ones on this market publishing a worst scorecard.

Second wave, September 27, 2026

114 more decisions, 57 per market: the exam base grows to 100 cases in the US and 100 in France.

At 43 cases, a single verdict weighs more than two points, and a demanding executive is right to find the base thin. The new cases came in through the same door as the old ones: a list of candidates written before any run on the test bench, deduplicated against the original exam and against the Hub cases; case files written ex ante, every quote checked on its page by a script; real success graded by two readers who could not see the verdict, disagreements settled by rules written before the reading; and a dedicated review hunting for any trace of the outcome in the decision as submitted. Success labels: in the US, 25 goals met, 20 missed, 12 met in part; in France, 25, 19 and 13.

Against real success, 55 of the 57 US verdicts are right or within the accepted margin, 21 of them exactly; in France, 50 of 57, 14 of them exactly. Same rules as the sealed exam, published above. Against whether the decision was kept, the count is 20 exact verdicts and 6 wrong-way ones in the US, 19 and 8 in France.

Of the 114 verdicts in this wave, nine went the wrong way on success, and we name them. In the US, one green light on a missed goal, Union Square Hospitality Group ending tipping, and one call to hold off on a success, Dollar Tree moving to $1.25. In France, one green light on a missed goal, Cajoo's 15-minute grocery delivery, and six calls to hold off on decisions that reached their goal: Fnac buying Darty, Renault's competitiveness agreement, Ford turning down a buyer for its Blanquefort plant, the closing of Goodyear's Amiens-Nord plant, the end of the A380, and Sanofi selling control of Opella. Our reading: all six French hold-off calls are restructurings (plant closings, the end of a program, a takeover, a divestiture, a labor agreement), where the reception was harsh and the goal was met anyway. The test bench measures how an announcement is received; it does not say that a badly received decision will fail.

The total, sealed exam plus second wave

US: 95 of 100 verdicts are right or within the accepted margin against real success, 39 of them exactly; 5 went the wrong way. France: 87 of the 98 cases that can be graded, 26 of them exactly; 11 went the wrong way, two cases stay ungraded. The total adds two measurements of a different nature, one sealed, one not: we give it for the size of the base, and keep each table readable on its own.

The US second wave, September 27, 2026

Generated straight from the run file you can download below (run of September 27, 2026, decoupled judge openai/gpt-4.1-mini). No line in it is written by hand.

The French second wave, September 27, 2026

The tuning log

A precision instrument keeps a log. Here is ours, dated: on each line, the trial, the measurement, the lock.

• 07/01: the French corpus opens, every source link verified that day; it is frozen at 43 decisions and 185 objections in early July.
• 07/25: the exam replayed in 5 passes catches two display drifts, including an old worst scorecard still sitting on five satellite pages. The lock: the shared-facts manifest is born, and a deployment gate has locked every figure on every page since.
• 07/27: one startup-stratum case leaves for lack of a primary source; 14 becomes 13, the 83% average does not move.
• 08/03: the measurement shows the homepage was displaying Dunkirk's best run, 92, instead of the average, 89. The lock: "never the best of the runs" applies everywhere, including when it lowers a good grade.
• 08/08: the American selection rule is written and sealed before the corpus grows from 16 to 43; the Qwikster outcome gap is logged the same day, the documentation wins.
• 08/09: parity is reached: 43 American cases, full exam, zero failures.
• 08/15: the coverage judge changes houses: openai/gpt-4.1-mini, decoupled from the voices' model, validated on a known-answer control before being hired. Both exams are replayed in full and the figures published as they come: the averages move to 74% and 84%, the worst scorecards to 13% and 33%. The lock: the judge's model is written into every published artifact.

An instrument that catches its own drift before its readers do is worth more than an instrument that never drifts on paper.

Request access

Five percentages, zero published cases

On this market, the reliability scores on display run from 82.7 to 96%. All self-declared, none tied to a protocol a reader can replay. Here are their exact words, and what each one measures.

The challenge stands: publish a corpus, a rule written before the measurement and a worst scorecard, and this page will say so.

The science behind the grading · How to judge any tool on this market

The limit in writing, the sealed answer

See who leans in, who pushes back, and on which blind spot, before you announce. This exam is retrospective, the model may have seen how the story ends: that is written in the record, and it is exactly what the 15 sealed verdicts below are built to close. It is not a poll, not a prediction, and it is never a measure of public opinion: a test bench, grades published.

On August 8, 2026, 15 verdicts on company decisions still in progress were sealed: Netflix, Apple, Ubisoft, OpenAI, among others. Their outcomes do not exist yet; the verdicts are frozen anyway. The file is public: kapari.eu/registre-prospectif-2026-08-08.json. Its fingerprint, a2b02a494359f96b3cf04885bdb38da5b3541a648d3fc735b17e6ee6ae2080f0, is anchored through OpenTimestamps into the Bitcoin blockchain. The seals open on February 8, 2027, then on August 8, 2027, both counts set in the seal published, hits and misses. Alongside them, added after sealing and labeled as such, a second reading against real success, for the decisions whose goal can already be judged; the others will be declared ungraded. The seal is free. The six months are not. And you do not have to trust us: the exact fingerprint recipe, with a downloadable verification script, is published.

Find another vendor whose page says that.

Replay the exam

Everything here is downloadable: both corpora (JSON), the real-success labels for all 86 decisions with their sources, the reference runs (replays of August 15, 2026, with the decoupled judge), the earlier runs kept as history, the endorsement judgments, the judge prompts, the triage of the 121 objections from the French corpus. Voices model: mistralai/mistral-large-2512; judge: openai/gpt-4.1-mini at temperature zero, written into each run's synthesis along with the pass count. The script survives bad nights: patient retries, a save after every case, a resume mode, a no-anchor control. Of the 141 source links behind the French exam, 130 open today. Click them.

• The US corpus and the French corpus
• The real success of all 86 decisions, with sources (English notes included)
• The reference runs: US, August 15, 2026 (decoupled judge) and French, August 15, 2026 (decoupled judge)
• Second wave: the US corpus and the French corpus (success labels and sources included), and the runs of September 27, 2026: US and French
• Earlier runs kept as history: US, August 9, 2026, French, July 25, 2026 and the French replay, August 8, 2026
• The endorsement judgments
• The judge prompts and the triage of the 121 objections (on the French replay of August 8, 2026)

Hear the objections before the room does

Your next difficult announcement already has its objections. The only open question is whether you hear them before or after.

Request access

A 30-day trial at sixty decisions a month, then five a month, until payment opens at the latest. No card, no sales call.

We do not promise you will make the best decision. But we assure you that at the moment you make it, you will have the most complete set of elements to help you.

Want the proofs first? Open the scorecards on the Kapari Hub, download the sealed registry, and read how the test bench works.

The worst objection is the one that finds you first.