Tasks / Discover

Customer research call guide

Can the model write a guide that uncovers behaviour rather than opinions?

Measures the modelTask v1.0 · 2 casesDifficulty

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

From 9 graded outputs by 5 models. 50% were usable with at most a quick edit.

Reliably right

  1. Identifies material uncertainty100% pass
    Names specific unknowns (cost of lateness, ease of chasing, automation fit, willingness to pay) and says mixed findings mean targeted follow-up.
    GPT-6.1 Sol · API · Would freelancers pay to stop chasing invoices?
  2. Designed to falsify100% pass
    Every learning goal has disconfirming questions, e.g., 'What did you do next?', 'When do you avoid sending another reminder?'
    GPT-6.1 Sol · API · Would freelancers pay to stop chasing invoices?
  3. Asks what people did, not what they think100% pass
    Core questions ask for the most recent overdue invoice and specific actions taken, not opinions or predictions.
    GPT-6.1 Sol · API · Would freelancers pay to stop chasing invoices?

Where it slips

  1. Fits the call33% pass
    Timings add up to 30 minutes but no must-ask questions are marked and no guidance on what to cut if time runs short.
    GPT-6.1 Sol · API · Would freelancers pay to stop chasing invoices?
  2. Respects explicit constraints81% pass
    The guide is clearly over 900 words, violating the explicit length constraint.
    Opus 5.5 · Claude · Would freelancers pay to stop chasing invoices?
  3. Different questions for user and signer81% pass
    The guide does not provide distinct question tracks for the daily warehouse operations manager and the executive who signed; it only differentiates churned vs at-risk customers, with the same questions for both roles within each track.
    GPT-6 Luna · API · Why are warehouses leaving? Two execs, two theories

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

You're a Staff PM at Crate. In four weeks, the exec team decides which of two bets gets our two squads next half, and both bets rest on a theory of why customers are leaving. We have three weeks for 12 customer calls of 45 minutes. Write the guide for them. It should include: the learning goals; the questions, with timings, for the two kinds of people we'll talk to; guidance for whoever runs the calls; any changes you'd make to the plan below; and what we'd need to hear to back each bet, or neither. Keep it under 1,500 words. The exec team will read it before the calls start.

About CrateWarehouse management software for mid-size third-party logistics companies (3PLs). 340 customers, $22M ARR. Gross logo churn rose from 11% to 18% over the last year.
The decisionBet A: integrations with warehouse robotics (autonomous carts, pick-assist robots). Bet B: rebuild implementation so customers go live faster. Two squads for the next half go to one of them.
CEO, Ana Ferreira“We're losing warehouses because they're automating and we don't talk to their robots. Every lost customer I've spoken to mentions it.”
CPO, Marcus Hill“We lose them in the first year because implementation takes forever. They're gone before they ever see the value.”
Exit survey (41 churned customers, last 12 months)A single-choice question, 'Why are you leaving?'. Missing features or integrations: 46%. Too hard to implement: 22%. Price: 17%. Other: 15% (free-text box, left blank by all but two). 'Missing features or integrations' was the first option in the list.
Implementation and CRM dataMedian time from signing to go-live: 94 days; the contract promises 45. Customers who went live after 90 days churned at 2.4× the rate of the rest in their first year. 3 churned customers moved to a competitor with robotics integrations; robotics was cited in 9 lost new-logo deals. We don't know when in the year churned customers decided to leave.
The call plan12 calls: 6 with customers who churned in the last 9 months, 6 with at-risk customers (health score red). On each account we want to talk to the warehouse operations manager who used Crate every day and, separately, the VP of Operations or COO who signed the contract. Ana wants to run 4 of the calls with former customers she knows personally.
Notes from Customer Success and Legal5 of the 6 at-risk accounts are in renewal talks, and Customer Success is offering them discounts. Interviewers must not discuss pricing, discounts or what's on our roadmap. Every call must start by asking permission to record.
What a strong answer does

A guide that tests both theories fairly and leaves room for a third: the goals ask what actually drove each decision to leave or doubt, when it happened, and who made it, with questions that could disprove robotics (did they automate, did robots come up before they decided, what did they switch to and why) and implementation (how long go-live took, what happened in those months, whether value arrived) alike. It treats the exit survey as weak evidence (single choice, integrations listed first, 'Other' left blank) and the 2.4× figure as a correlation to probe, not a cause. The core is a timeline of the account: the story from signing to the moment they decided to leave or started doubting, with sequential probes. It has separate tracks for the daily user (what went wrong day to day) and the signer (how the decision was made, what alternatives they weighed), fits 45 minutes with timings, and opens with consent to record. It guards the calls: no pitching, pricing, discounts or roadmap promises, with a script for when an at-risk customer asks, and it addresses Ana's bias: she has a public view and personal relationships, so she should observe or take notes rather than lead, or at minimum use the same script and have her calls read separately. It ends with decision signals: what pattern across the 12 would back Bet A, what would back Bet B, and what would back neither (price, a business change, a missing capability that isn't robotics).

Critical failures (cap the score)
  • Includes questions that promise roadmap items or discuss pricing or discounts with customers in renewal talks
  • Asks customers to choose between robotics and implementation as the reason they left
Case

v1.0 · synthetic · B2B SaaS, logistics, churn, exec bias, staff-level

Usable with a quick edit11 of 13 checks passed

The guide is thorough and evidence-based; a PM could add a note on which questions to prioritize if time is tight and use it immediately.

Output

Crate churn research: executive pre-read and interview guide

Decision: Allocate two squads next half to robotics integrations (A), implementation rebuild (B), or neither if the evidence does not justify either.

Research stance: Test both explanations, not confirm either executive’s account. These interviews can reveal mechanisms and decision timelines; they cannot estimate their prevalence across 340 customers or prove causality.

1. Learning goals

  1. Reconstruct the departure process: When did concerns emerge, when did leaving become likely, and when was the decision effectively made?
  2. Identify consequential problems: Separate reasons volunteered after leaving from problems that materially changed warehouse outcomes and the buying decision.
  3. Test Bet A: Was a specific robotics incompatibility a binding obstacle? Would addressing it have plausibly retained the account?
  4. Test Bet B: Did implementation delay prevent value, damage trust, or trigger departure? What caused the delay, and could Crate control it?
  5. Find alternatives and interactions: Price, service, reliability, business changes, or other missing capabilities may dominate. Robotics work may itself complicate implementation.
  6. Connect learning to an investable scope: What could two squads deliver next half that would address the demonstrated retention mechanism?

What the existing evidence does—and does not—say

  • The exit survey’s broad integration category, single-choice format, first-position placement, and sparse free text do not establish robotics as the cause.
  • The 2.4× churn association makes implementation worth investigating, but complexity, customer readiness, or other factors could drive both delay and churn.
  • Three competitor moves and nine lost deals indicate a robotics opportunity; acquisition losses are not evidence of a retention mechanism.
  • Unknown decision dates make chronology essential.

2. Changes to the call plan

Resolve the arithmetic: Twelve separate interviews cannot cover twelve accounts with two roles each. Use six accounts × two separate role interviews = twelve 45-minute calls.

Recommended mix: Four churned accounts and two at-risk accounts. This reduces reliance on renewal negotiations while retaining a prospective view. Findings will be depth-oriented, not representative.

Select accounts from the full CRM population, not executive contacts or survey answers alone:

  • Across churned accounts, include slow and faster implementations, first-year and later departures, and robotics-exposed and non-exposed customers.
  • Include at least one confirmed move to a robotics-enabled competitor and one slow implementation without an apparent robotics need.
  • For at-risk accounts, prefer one with active automation plans and one with implementation/value-realization problems. Include the non-renewal account if suitable.
  • Seek contrasting retained accounts through existing operational data, even though the call budget does not cover them.

Recruit both the daily warehouse operations manager and the contract-signing VP Operations/COO. If the original signer has left, recruit someone directly involved in the departure or renewal decision and document that substitution.

Ana should not lead calls with her personal contacts. Use a neutral interviewer without renewal responsibility. Ana can help recruit; observing requires explicit customer consent. Her existing conversations are useful leads, not independently verified findings.

CS should introduce research as separate from renewal discussions, then step out. Participation must not affect commercial treatment.

3. Interview guides: 45 minutes each

Use the relevant role column. In every section, distinguish direct experience from hearsay. Ask broad questions before naming either bet.

TimeDaily warehouse operations managerVP Operations / COO
0–3 min: permission and framingFirst ask: “May we record this conversation for internal research?” If declined, continue with notes. Then explain purpose and confidentiality limits.Same opening.
3–7 min: context and expected value“What did your warehouse handle, and what was your role with Crate? What was supposed to improve?”“What led you to buy Crate? What outcomes and deadlines mattered? How would you judge success?”
7–17 min: chronological reconstruction“Walk me from signing through onboarding, first operational use, and the point when problems became serious.” “Describe a particular shift or incident.” Establish dates, workarounds, operational impact, and who knew.“Walk me from purchase to the first concern, evaluating alternatives, and the decision.” “When did staying stop being the default?” Establish dates, decision-makers, triggers, and alternatives.
17–27 min: implementation and value“What had to happen before you could use Crate successfully? Where did work stall? Who owned each step?” “When did you first achieve useful results?” Probe migration, configuration, training, integrations, staffing, and rework after an open answer.“What go-live date did you expect, and what happened? What consequences did that have?” “What other factors accompanied the delay?” “If the same product had gone live in 45 days, what would still have put the relationship at risk?” Ask why.
27–36 min: workflow, automation, and unmet needs“Which workflows did Crate support poorly? Show or describe a recent example.” Then: “Were robots in use or planned? Which systems, what workflow, and what connection was required?” “What happened without it?”“What capabilities influenced staying or switching?” Then probe automation: vendor, deployment date, committed budget, required integration, and alternatives considered. “If that connection had existed, what else would have needed to change for you to stay?”
36–42 min: decision and competing explanations“Which problem mattered most in daily operations? What did you escalate, to whom, and when?” “What worked well?” “What have we missed?”“Which issues were necessary to the decision, and which were secondary?” “What would have had to be different for you to stay?” Ask about business changes, service, reliability, and other alternatives without forcing a category.
42–45 min: verify and closeSummarize the timeline and mechanism: “What have I misunderstood?” Request relevant artifacts and permission for a brief clarification follow-up.Same; verify whether operational problems actually influenced the commercial decision.

Status-specific wording

  • Churned: Ask what the replacement actually delivered, whether it went live, and whether the cited problem improved—not merely what the competitor promised.
  • At-risk: Ask what has already happened, current unresolved consequences, and what would cause escalation or departure. Do not imply a departure decision exists. Distinguish funded plans from aspirations.

4. Guidance for interviewers

  • Begin with recording permission; record only after consent. Explain that this is research, not a support, roadmap, or renewal discussion.
  • Do not discuss prices, discounts, or roadmap commitments. If raised, acknowledge without probing commercial terms: “I can’t discuss that here; your account team handles commercial conversations.” Record volunteered price concerns as an alternative explanation.
  • Avoid “Did missing robotics make you leave?” and “Would faster onboarding have saved you?” Start with events; use counterfactuals only afterward.
  • Ask for concrete examples, dates, artifacts, and consequences. “It was painful” is not yet a mechanism.
  • Do not defend Crate, disclose executive hypotheses, or offer fixes.
  • Interview roles separately before reconciling accounts. Treat disagreement as evidence, not an error to erase.
  • Use a note-taker and common template: timeline, problem, consequence, decision link, alternative explanation, evidence source, and confidence. Count accounts, not two interviewees as independent votes.

5. Evidence required for the decision

Back A: robotics integrations

Look for a repeated chain across independent accounts:

Committed robotics deployment → specific missing connection → material operational/business constraint → documented role in departure or credible near-term risk.

Stronger evidence includes decision-maker confirmation, pre-decision correspondence, and a replacement that solved the problem. Match the implicated vendors/workflows to current customer exposure. Before funding, establish that a feasible integration scope could address meaningful retention risk next half.

Mentions of “automation,” hypothetical plans, or competitor marketing are insufficient.

Back B: implementation rebuild

Look for:

Crate-controllable implementation bottleneck → delayed usable value or broken operational commitments → loss of confidence/avoidable cost → departure decision before value realization.

Seek repeated bottlenecks, role-level corroboration, and contrasts with faster implementations. Validate the mechanism against the full cohort, accounting where possible for complexity and customer readiness. Identify changes two squads could deliver; do not assume a wholesale rebuild is necessary.

Back neither—or declare the result inconclusive

Neither is justified if dominant causes lie elsewhere, robotics demand is narrow or speculative, delays are largely outside Crate’s control, or either intervention arrives too late to affect retention. If interviews remain contradictory or unsupported, report insufficient evidence, not “neither causes churn.”

Compare both bets on reachable ARR at risk, evidence strength, controllability, delivery feasibility, and time to impact—not mention counts.

6. Three-week execution

  • Week 1: Audit cohorts and timelines, recruit, pilot one role pair.
  • Week 2: Complete remaining interviews; synthesize after each pair without changing standards to favor an emerging winner.
  • Week 3: Verify discrepancies, triangulate with records, assess engineering scope, and prepare the decision memo with supporting evidence, counterexamples, uncertainties, and a recommendation.

Do not commit two squads merely because one story sounds more persuasive.

Check by check

Got wrong · 1
  • Fits the callThe guide includes timed sections that add up to 45 minutes but does not say what to cut if time runs short, nor does it mark must-ask questions.
Mixed · 1
  • Addresses the actual decisionThe output recommends more research before deciding, states the supplied evidence cannot yet support a call, and specifies exactly what evidence would back each bet or neither.The two graders disagreed on this one.
Got right · 11
  • Uses the supplied evidence correctlyAll factual statements about the current situation are taken directly from the supplied context without invention.
  • Respects explicit constraintsThe guide is under 1,500 words, includes all requested sections (learning goals, timed questions for two roles, guidance, plan changes, decision signals), and is addressed to the exec team.
  • Identifies material uncertaintyIt names key unknowns (decision dates, whether robotics was a binding obstacle, whether delays were Crate-controllable) and says how the interview evidence would resolve them.
  • Avoids unsupported claimsInterpretations are clearly labelled as such (e.g., the exit survey does not establish robotics as the cause), and no cause is presented as established fact.
  • Produces the required deliverableThe output is a complete guide with learning goals, timed questions for both roles, interviewer guidance, plan changes, and decision signals; an exec team could act on it.
  • Tests both theories fairlyQuestions could disprove robotics (e.g., 'Were robots in use? What happened without it?') and implementation (e.g., 'Where did work stall? If go-live had been 45 days, what would still have put the relationship at risk?'), and open questions leave room for a third cause.
  • Different questions for user and signerThe table provides distinct tracks: daily user gets questions about specific shifts, workarounds, and escalations; the signer gets questions about the purchase decision, alternatives, and go-live expectations.
  • Protects the calls and the accountsIt forbids discussing pricing/discounts/roadmap, provides a script to deflect, starts with recording consent, and has a concrete plan for Ana (she does not lead, observes with consent, her prior conversations are treated as leads).
  • Designed to falsifyEach learning goal has disconfirming questions (e.g., 'What happened without the robot connection?' for Bet A, 'If the same product had gone live in 45 days, what would still have put the relationship at risk?' for Bet B).
  • Asks what people did, not what they thinkCore questions ask for a chronological walk-through, specific incidents, dates, and what they did, not opinions or predictions.
  • Doesn't lead the witnessQuestions are open and neutral; robotics and implementation are not named until after broad problem questions, and no answer is hinted at.

Grades and run details

Decision model 88 · LLM judge 13 of 14 checks
Decision model checks
  • passUses the supplied evidence correctly62%
  • failAddresses the actual decision55%
  • passRespects explicit constraints20%
  • passIdentifies material uncertainty99%
  • passAvoids unsupported claims95%
  • passProduces the required deliverable86%
  • passTests both theories fairly100%
  • passDifferent questions for user and signer100%
  • passProtects the calls and the accounts95%
  • passDesigned to falsify64%
  • passAsks what people did, not what they think71%
  • passDoesn't lead the witness58%
  • partialFits the call68%
Run
Run
#1
API response time
80 s
Submitted
30 Sept 2026

Needs more than a quick edit10 of 13 checks passed

The guide lacks separate question tracks for the daily user and the decision-maker, and it doesn't prioritize questions for time management. A PM would need to add role-specific questions and mark must-asks, which is more than a quick edit.

Output

Crate customer research guide: why are warehouses leaving?

Purpose: In three weeks, learn whether churn is more plausibly driven by missing robotics integrations (Bet A), slow implementation (Bet B), or something else. These calls are for understanding customers’ experiences—not selling, negotiating renewals, or validating either executive’s theory.

Learning goals

For each account, establish:

  • What happened, and when? Separate implementation, go-live, first value, automation needs, and the point when leaving became likely or decided.
  • What mattered most? Distinguish a decisive problem from a feature request, contributing frustration, or reason offered after the fact.
  • What evidence supports the explanation? Look for actual workflow impact, workarounds, alternatives evaluated, and actions taken—not just opinions or hypotheticals.
  • What might change the outcome? Explore credible counterfactuals without promising anything.
  • What else explains churn? Surface causes beyond A and B.

Recommended plan changes

The proposed 12 calls cannot cover both the daily user and contract signer on all 12 accounts: that would require 24 interviews. Keep the 12-call cap, but interview two roles separately at each of six accounts: three churned accounts and three red-health accounts, with one warehouse operations manager and one VP of Operations/COO per account. This gives paired perspectives and preserves both customer situations, but is a small, directional sample—not a prevalence estimate. Select accounts for varied implementation times and automation situations where possible; don’t choose only accounts with known robotics issues.

Do not have Ana conduct four calls with former customers she knows personally. Her involvement risks courtesy bias and leading the conversation. Use an independent interviewer; ideally don’t include personal contacts in this small core sample. If one is included, disclose the relationship, and have Ana neither attend nor receive attributable notes.

Because five at-risk accounts are in renewal talks and receiving discount offers, use an interviewer outside the account/renewal team. Tell participants their answers won’t affect service or renewal discussions. Don’t share interview content with the account team in a way that could be used in negotiation. Follow Legal’s recording requirement below.

45-minute guide: churned customers

TimeQuestions
0–3Start by asking: “May I record this conversation?” If no, take notes instead. Explain the purpose, that there are no sales or renewal implications, and that we won’t discuss pricing, discounts, or roadmap.
3–7“What was your role in choosing and using Crate?” “What did your warehouse need the system to help you do?”
7–18“Take me through your experience with Crate, from signing through the decision to leave.” “When did you first consider switching? What happened just before that?” “Who was involved, and when was the decision effectively made?”
18–27“How did implementation and go-live go?” “Where, if anywhere, did it take longer or require more effort than expected?” “What was the impact on your operation?” “When did you first get value from Crate?”
27–35“Did automation or robotics play a role in your warehouse during this period?” If yes: “What equipment or workflow?” “How did it work with Crate?” “What did you do when it didn’t?” “When did this become a problem relative to the decision to leave?”
35–41“What alternatives did you consider, and what drove the choice?” “What, if anything, could realistically have changed your decision?” “What was the most important reason—not just one frustration?”
41–45“What haven’t I asked that matters?” “Is there anything we should correct about our understanding?” Thank them.

45-minute guide: red-health customers

Use the same opening and close. Do not imply that Crate can fix an issue or influence renewal terms.

TimeQuestions
0–3Start by asking: “May I record this conversation?” If no, take notes. Explain that this is research, not a sales or renewal conversation; participation won’t affect service or renewal discussions.
3–7“What is your role in using or overseeing Crate?” “What does a successful warehouse operation look like for you?”
7–17“How has your experience with Crate unfolded so far?” “When have things gone better or worse than expected?” “What is working well today?”
17–27“Where are you in implementation and go-live?” “What has taken the most effort or time?” “What value, if any, have you realized so far?” “What is the operational impact of anything that remains difficult?”
27–35“How, if at all, does automation or robotics fit into your operation?” If relevant: “What equipment or workflow?” “How does it work with Crate today?” “What workarounds do you use?” “When did this become important?”
35–41“Have you considered alternatives to Crate? What prompted that, if so?” “What would need to be true for Crate to work well for your operation?” “What are the most important unresolved issues?”
41–45“What haven’t I asked that matters?” “Is there anything we should correct?” Thank them.

Guidance for interviewers

  • Ask the questions as written, but follow the participant’s timeline. Use neutral probes: “What happened next?” “Can you give me an example?” “How do you know?”
  • Don’t offer Ana’s or Marcus’s theories, mention the exit-survey results, or ask leading questions such as “Would robots have saved you?” Ask about events before asking for interpretations.
  • Separate when an issue occurred from when the participant says it mattered. Probe for specific actions and evidence: evaluations, workarounds, escalations, or a switch decision.
  • Don’t discuss pricing, discounts, renewal negotiations, or roadmap. If asked about roadmap, say you can’t speak to future plans. Don’t make promises or relay an interview answer as a renewal signal.
  • Interview roles separately. Record role and account context in notes, but distinguish direct experience from second-hand views. Afterward, write a short timeline, leading cause(s), evidence, and uncertainties for each account.
  • Treat recordings and notes as confidential; use only with consent and Legal-approved handling. Don’t present anecdotes as representative of all customers.

What would support each bet?

Support Bet A (robotics integrations): In multiple independent accounts, a specific robotics workflow or integration gap caused meaningful operational friction, and that issue arose before the decision to leave or seriously consider leaving. Stronger evidence: an actual robot deployment or concrete plan, a costly workaround, and a competitor selected or evaluated for this capability. A feature mention alone—or robotics cited in a new-logo deal—doesn’t establish that it drove existing-customer churn.

Support Bet B (faster implementation): In multiple accounts, implementation delays or effort materially postponed go-live or first value, caused operational consequences, and preceded the decision to leave or consider leaving. Stronger evidence: specific preventable blockers, an unmet expectation, and a credible account that earlier go-live would have changed the outcome. The 94-day median, 45-day promise, and 2.4× association make this worth testing, but do not prove causation.

Support neither / remain uncertain: The call evidence points more strongly to another cause; A or B appears only after the decision; accounts describe issues without meaningful consequences; or evidence is mixed and no cause consistently precedes churn. Don’t force a winner from 12 interviews. Compare patterns across accounts, not individual interview counts, and triangulate with CRM timelines and other churn evidence. The exit survey’s single-choice format, first-listed feature option, and sparse free text limit what it can establish.

Check by check

Got wrong · 2
  • Different questions for user and signerThe guide does not provide distinct question tracks for the daily warehouse operations manager and the executive who signed; it only differentiates churned vs at-risk customers, with the same questions for both roles within each track.
  • Fits the callAlthough timed sections add up to 45 minutes, the guide does not mark must-ask questions or say what to cut if time runs short, so an interviewer could run out of time without covering the most critical probes.
Mixed · 1
  • Addresses the actual decisionThe guide clearly states what evidence from the calls would back Bet A, Bet B, or neither, which is the decision framework the exec team needs.The two graders disagreed on this one.
Got right · 10
  • Uses the supplied evidence correctlyEvery statement about the current situation is taken directly from the supplied context, with no invented facts.
  • Respects explicit constraintsThe guide respects all constraints: it includes learning goals, timed questions for both customer types, interviewer guidance, plan changes, and decision signals; it stays under 1,500 words; it enforces no pricing/roadmap discussion and recording consent.
  • Identifies material uncertaintyIt names key unknowns (exit survey weakness, correlation vs causation, timing of churn decision) and says how the calls will resolve them through patterns and timelines.
  • Avoids unsupported claimsInterpretations like the exit survey's limitations and the 2.4× association are clearly labeled as not proving causation, not presented as established fact.
  • Produces the required deliverableThe output is a complete call guide with all requested sections, written for the exec team, and well under 1,500 words.
  • Tests both theories fairlyBoth the robotics and implementation theories get questions that could disprove them, open timeline questions come first, and the guide explicitly leaves room for a third cause.
  • Protects the calls and the accountsIt gives a clear rule and script for pricing/roadmap questions, starts with recording consent, and addresses Ana's bias by recommending she not lead calls and, if a contact is included, that she not attend or receive attributable notes.
  • Designed to falsifyEvery learning goal has questions whose honest answers could disprove the hypothesis, such as asking when automation became a problem relative to the decision to leave.
  • Asks what people did, not what they thinkCore questions ask for specific past events and actions (e.g., 'When did you first consider switching? What happened just before that?'), with opinions and predictions clearly secondary.
  • Doesn't lead the witnessQuestions are neutral and open; the guide avoids mentioning either executive's theory until after the participant's own timeline, and prohibits leading probes.

Grades and run details

Decision model 85 · LLM judge 11 of 14 checks
Decision model checks
  • passUses the supplied evidence correctly81%
  • failAddresses the actual decision50%
  • passRespects explicit constraints26%
  • passIdentifies material uncertainty95%
  • passAvoids unsupported claims91%
  • passProduces the required deliverable73%
  • passTests both theories fairly99%
  • partialDifferent questions for user and signer74%
  • passProtects the calls and the accounts79%
  • passDesigned to falsify75%
  • passAsks what people did, not what they think70%
  • passDoesn't lead the witness55%
  • partialFits the call29%
Run
Run
#1
API response time
29 s
Submitted
30 Sept 2026

Results

Every setup we’ve tested on this task, across all cases and repeats, graded on the current checklist. Provisional The checklist is still being calibrated against our PM.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6 AstrawithChatGPT95.892.31None
2GPT-6.1 SolwithAPI92.192.62None
3Opus 5.5withClaude94.169.22None
4Sonnet 5.5withAPI86.484.62None
5GPT-6 LunawithAPI90.273.921 capped

About the task

The PM job

Preparing for customer calls.

Why it matters

Leading questions produce the answers you wanted. A good guide is designed to prove you wrong.

What good looks like

  • Past-behaviour questions
  • No leading prompts
  • Clear learning goals
  • Could disprove the team's hypothesis

Deliberately not measured

    Capability tested

    Question design

    The failure we’re looking for

    Hypothetical, leading questions

    Grading

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