Humanity’s Last PRD
Latest drop: Sonnet 5.5 · 30 Sept
Tasks / Experiment

Analyse experiment results

Can the model separate evidence from speculation, identify decision-relevant uncertainty and recommend a sensible next action?

Measures the modelTask v1.2 · 2 casesDifficulty

What AI gets right here, and what you’ll still have to catch

From 12 graded outputs by 6 models. 83% were usable with at most a quick edit.

Reliably right

  1. Produces the required deliverable100% pass
    It is a usable decision memo for the named reader with a clear recommendation, reasoning, and next steps.
    GPT-6 Astra · ChatGPT · Conversion up, retention down
  2. Gets the base of every number right100% pass
    The refund increase and revenue difference are computed correctly, and the retention CI interpretation uses the correct 2pp threshold.
    GPT-6 Astra · ChatGPT · Conversion up, retention down
  3. Interprets power correctly100% pass
    Explains that the test was designed to detect ~3pp effects, so a smaller but worthwhile lift could still exist and remain undetected.
    GPT-6 Astra · ChatGPT · The underpowered onboarding test

Where it slips

  1. Trusts the data before reading it15% pass
    It does not explicitly check a pre-interpretation trust signal such as sample ratio or logging before relying on the lift; the denominator and maturity caveats come after interpreting.
    GPT-6.1 Sol · API · Conversion up, retention down
  2. Avoids unsupported claims60% pass
    It asserts the refund surge 'isn't noise' and that annual up-front billing means retention 'should rise' without supporting evidence or clear labeling.
    Sonnet 5.5 · API · Conversion up, retention down
  3. Uses the supplied evidence correctly69% pass
    Asserts that comparable per-arm ticket rates and exposure counts are unavailable, which is not in the supplied context and cannot be assumed.
    GPT-6 Astra · ChatGPT · The underpowered onboarding test

Case viewer

Read the brief, then put up to three outputs side by side, each with the LLM judge’s verdict on every check. Highlights mark what a PM had to fix.

The brief

We tested a shorter onboarding checklist and the result came back 'not significant'. The team wants to call it a failure and move on. Write a short note to the onboarding team saying what we should conclude from this test and what we should do next. Keep it under 300 words.

ScenarioTrellis is a project-management tool for small teams. The shorter checklist cuts onboarding from nine steps to five. It is already built: shipping it to everyone would take about a day of engineering. Activation means creating a first project and inviting at least one teammate within seven days.
ReadoutActivation: A 31.2%, B 32.9% (+1.7pp, 95% CI −1.4 to +4.8). n = 3,960 per arm. The pre-registered minimum detectable effect was 3pp.
QualitativeSupport tickets tagged 'onboarding confusion' fell from 44 to 29 during the test.
Options on the tableRunning the test for another four weeks at current traffic would detect an effect of about 2pp.
What a strong answer does

Inconclusive, not a failure: the test could not detect effects smaller than ~3pp, and the point estimate is positive. Recommend a decision based on cost of shipping versus a longer test.

Critical failures (cap the score)
  • Concludes the change has no effect
Case

v1.6 · synthetic · null result, onboarding

GPT-6 AstraChatGPTUsable with a quick editGPT-6 LunaAPIUsable with a quick editGPT-6.1 SolAPIUsable with a quick editSonnet 5.5APIUsable with a quick editGemini 3.5 Flash-LiteGeminiUsable with a quick editOpus 5.5ClaudeNeeds more work
GPT-6 LunaAPI

Usable with a quick edit8 of 9 checks passed

The core reasoning and use of evidence are sound, and the only meaningful omissions are explicit trust checks and the original 3pp MDE, which could be added in a quick edit. No rechecking or reanalysis is required.

Output

Team—

This is inconclusive, not a failure. Activation was 1.7 percentage points higher with the shorter checklist, but the 95% CI (−1.4 to +4.8 points) includes no effect and possible harm—as well as a benefit of at least 3 points. The test wasn’t precise enough to settle the question.

I recommend running the test for another four weeks. That should let us assess effects around 2 points, which this change may plausibly deliver. Before extending, agree on what activation gain would justify rollout and what activation decline would stop it. Track onboarding-confusion tickets too, but compare rates per user: the raw drop from 44 to 29 is encouraging, not conclusive.

After the extension, decide using both activation and confusion—not a “significant/not significant” label alone.

Check by check

Got wrong · 1
  • Trusts the data before reading itIt does not explicitly check any trust signal such as sample ratio, exposure or logging quality, or novelty before interpreting the activation result.
Got right · 8
  • Uses the supplied evidence correctlyEvery stated fact or figure is drawn correctly from the supplied readout/options, and no current-situation facts are invented.
  • Addresses the actual decisionIt unambiguously answers that the result is inconclusive, not a failure, and says the next step is a four-week extension with pre-agreed thresholds before deciding.
  • Respects explicit constraintsThe output is a note to the onboarding team, stays well under 300 words, and respects the requested short-note form.
  • Identifies material uncertaintyIt names the unresolved true activation effect and support-ticket rate issue, bounds the activation uncertainty with the CI, and says extension plus pre-agreed thresholds would resolve it.
  • Avoids unsupported claimsInterpretive claims such as 'not precise enough' and 'encouraging, not conclusive' are appropriately hedged and follow from the supplied data.
  • Produces the required deliverableThe requested note is present, actionable, and complete enough for the onboarding team to proceed.
  • Interprets power correctlyIt communicates that the original test could not distinguish the observed 1.7pp effect from no effect or harm, and that more data would allow detection around 2pp.
  • Gets the base of every number rightThe activation figures and differences are computed from the right values, and the output flags that the ticket drop should be compared per user rather than as a raw count.

Grades and run details

Decision model 89 · LLM judge 9 of 10 checks
Decision model checks
  • passUses the supplied evidence correctly79%
  • passAddresses the actual decision97%
  • passRespects explicit constraints95%
  • passIdentifies material uncertainty98%
  • passAvoids unsupported claims47%
  • passProduces the required deliverable66%
  • passInterprets power correctly86%
  • failTrusts the data before reading it79%
  • passGets the base of every number right95%
Run
Run
#1
API response time
9 s
Submitted
29 Sept 2026
GPT-6 AstraChatGPT

Usable with a quick edit8 of 9 checks passed

The note is actionable and well-structured; the one unsupported claim about missing per-arm data can be fixed by rephrasing as a check to perform, without redoing the reasoning.

Output

Team,

This test is inconclusive, not a failure. Activation increased from 31.2% to 32.9%, but the confidence interval spans a 1.4 percentage-point decline to a 4.8-point improvement. We haven’t established a benefit or ruled out meaningful harm.

The pre-registered 3pp minimum detectable effect describes what we designed the test to detect reliably. It isn’t a threshold below which improvements have no value. A smaller lift could still be worthwhile.

Confusion tickets fell from 44 to 29, which is encouraging supporting evidence. Without comparable rates by experiment arm, consistent tagging and exposure counts, we can’t attribute that reduction to the shorter checklist.

My recommendation is a staged rollout, provided basic checks reveal no downstream harm. With only a day of engineering needed, I wouldn’t automatically spend another four weeks pursuing statistical significance. The remaining decision is whether the likely benefit justifies the remaining cost and risk.

Before rollout, confirm every included user has had the full seven-day activation window and check experiment allocation, tracking and any available downstream engagement measures. Compare confusion-ticket rates between arms.

Then increase exposure gradually, retaining a randomised holdout and agreeing monitoring dates and a rollback threshold upfront. This is a practical decision under uncertainty, not a declaration that B won.

Extend the experiment instead if a possible 1.4pp activation loss is commercially unacceptable1 or if establishing a roughly 2pp improvement would change our decision. If we extend, pre-specify the analysis and stopping rule; four more weeks offers greater precision, not a guaranteed answer.

What a PM had to fix

  1. 1Numbers wrongRedo the arithmeticQuick edit

    Source: ReadoutActivation: A 31.2%, B 32.9% (+1.7pp, 95% CI −1.4 to +4.8).

    What we’d changeDescribe 1.4pp as the lower end of the interval, not the largest possible loss, and qualify the one-day cost assumption.

Check by check

Mixed · 1
  • Uses the supplied evidence correctlyAsserts that comparable per-arm ticket rates and exposure counts are unavailable, which is not in the supplied context and cannot be assumed.The two graders disagreed on this one.
Got right · 8
  • Addresses the actual decisionCommits early to a staged rollout (conditional on checks) and states that a commercially unacceptable 1.4pp loss or a desire for a 2pp improvement would change the call.
  • Respects explicit constraintsRespects the word limit, addresses the onboarding team, and does not call the variant a failure.
  • Identifies material uncertaintyIdentifies that the CI includes harm, that ticket attribution is uncertain without per-arm data, and says extending the test would resolve the effect-size uncertainty.
  • Avoids unsupported claimsDoes not present interpretations as fact; treats the ticket drop as encouraging but not attributed, and does not overstate the statistical result.
  • Produces the required deliverableDelivers a short note to the onboarding team with an actionable recommendation and next steps, within 300 words.
  • Interprets power correctlyExplains that the test was designed to detect ~3pp effects, so a smaller but worthwhile lift could still exist and remain undetected.
  • Trusts the data before reading itRecommends checking experiment allocation and tracking before rollout, which acts as a trust-signal check.
  • Gets the base of every number rightAll percentages are clearly from the arm-level activation rates, and ticket numbers are reported as raw counts without a confusing base.

Claims the judge couldn’t find in the brief

  • Comparable confusion-ticket rates by experiment arm, consistent tagging and exposure counts are not available.

Grades and run details

Decision model 94 · LLM judge 9 of 10 checks
Decision model checks
  • passUses the supplied evidence correctly19%
  • passAddresses the actual decision70%
  • passRespects explicit constraints53%
  • passIdentifies material uncertainty100%
  • passAvoids unsupported claims33%
  • passProduces the required deliverable70%
  • passInterprets power correctly99%
  • partialTrusts the data before reading it31%
  • passGets the base of every number right92%
Run
Run
#1
Time to output
21 s
Submitted
25 Sept 2026

Results

Every setup we’ve tested on this task, across all cases and repeats, graded on the current checklist. Calibrated: the graders match our PM on 88% of checks.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6 AstraChatGPT94.795.02None
2GPT-6 LunaAPI89.490.52None
3GPT-6.1 SolAPI89.490.52None
4Sonnet 5.5API81.775.92None
5Opus 5.5Claude76.157.32None
6Gemini 3.5 Flash-LiteGemini62.857.721 capped

The PM job

Reading a test readout and deciding what to do next.

Why it matters

Experiment readouts are where false confidence is cheapest to produce and most expensive to act on. A model that declares a winner on conversion while retention quietly falls will ship the wrong thing.

What good looks like

  • Checks the result can be trusted before interpreting it
  • Reports effects with their uncertainty
  • Checks guardrail metrics before declaring a winner
  • Separates what the data shows from plausible explanations
  • Recommends a next step proportionate to the evidence

Deliberately not measured

  • Re-running the statistics from raw data
  • Chart production
Capability tested

Interpreting results within their limits

The failure we’re looking for

Claims causality, ignores guardrails or recommends arbitrary testing

Grading

Decision model and LLM judge, calibrated against a blind PM review