Tasks / Leadership

Hire a PM

Can the model design a hiring process that finds the right PM, and make the call on real candidates from the evidence?

Measures the modelTask type v1.1 · 3 tasksLast changed 6 Oct 2026 · ChangelogDifficulty

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

From 19 graded outputs by 7 models. 79% were usable with at most a quick edit.

Reliably right

  1. Produces the required deliverable100% pass
    Complete recommendation memo with evidence, risks, and actionable next steps, usable as is.
    GPT-6 Luna · API · Two finalists, one Group PM role
  2. Judges on evidence, not presence100% pass
    Judges on concrete evidence (metric moved, team credit, learning from failure) and treats impressions like presence as weak signals.
    GPT-6 Luna · API · Two finalists, one Group PM role
  3. Tests what the last hire failed at100% pass
    The loop includes an influence simulation with the Head of Sales and a scorecard dimension on influence without authority, directly testing what the last hire failed at.
    GPT-6.1 Sol · API · A loop for the first growth PM

Where it slips

  1. Spots the interviewer pattern65% pass
    Does not notice that VP Engineering scores big-tech candidates higher and those hires were rated lower; only reports a negative correlation without the pattern.
    GPT-6 Luna · API · What our interviews predict
  2. Avoids unsupported claims70% pass
    The claim 'A growth PM in a senior role is likely to be exactly that kind of person' is presented as fact without evidence and is not labelled as an assumption.
    Opus 5.5 · Claude · A loop for the first growth PM
  3. Identifies material uncertainty72% pass
    The output does not name unknowns that could change the hiring decision or the loop design, nor does it say how they would be resolved.
    Sonnet 5.5 · API · A loop for the first growth PM

How it’s graded

The checks come from what the best product leaders have said about doing this job well on Lenny’s Podcast. Each one names the guest it comes from: follow a name to the idea on the Lenny’s Podcast wiki.

  1. Judges on evidence, not presence

    Does the output judge candidates, or design the process to judge them, on specific evidence of results they drove and how they worked with their team, rather than impressions such as presence, confidence or polish?

    Passes when Asks for or weighs concrete evidence: the metric a candidate moved, the part they played, how they shared credit, what they learned from a failure. Impressions are treated as weak signals.

  2. Defines good for this role first

    Does the output set out what good looks like for this specific role (the competencies that matter most at this level, and what strong and weak look like) before judging candidates or designing interviews?

    Passes when Names the competencies weighted for this role and level, with strong and weak signals, and the judgement or the process follows from them.

  3. Keeps each judgement independent

    Does the output keep each interviewer's judgement independent (written down before the group discusses it), and guard against the loudest voice or the first speaker deciding?

    Passes when Requires or relies on written, independent evaluations before discussion, and discounts views that changed under group pressure.

Plus our standard checks

Uses the supplied evidence correctly · Addresses the actual decision · Respects explicit constraints · Identifies material uncertainty · Avoids unsupported claims · Produces the required deliverable · and 3 written for each task, which you’ll see in the tasks below.

The tasks

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 Director of Product at Medvia, and you're the hiring manager for a Group PM. The debrief was yesterday. Write your recommendation to Rosa Lindqvist, our VP Product, in no more than 1,000 words: who to hire (Dev, Joy, or neither), the evidence for it, the risks, and what you'd check before making the offer. The CEO wants a decision by Friday. The interview notes and everything else are below.

What the model was given6 items: The role and the scorecard (agreed before interviews began), Dev Malhotra: interview notes, Joy Adeyemi: interview notes, Scores (1 to 4, entered after the debrief discussion), References, Compensation
The role and the scorecard (agreed before interviews began)Group PM for patient payments, leading three PMs. Must set the strategy for patient payments, lift the online payment rate (now 41%), and lead PMs while working closely with Finance and Compliance. Weighting: execution and influence 35%, product sense 25%, analytics 20%, people leadership 20%.
Dev Malhotra: interview notesProduct sense (Head of Design): 'Brilliant. Redesigned our statement flow live, lots of ideas, very creative.' Notes show six feature ideas in the first five minutes, before asking who pays and why. Analytics (Data lead): 'Strong.' Notes: asked what metric he moved at his last job, he said 'we grew the payments platform massively'; pressed, 'the team shipped 40+ features'. No number. Leadership (VP Engineering): 'Confident, inspiring.' Notes: every story in the first person; said his last PM team 'struggled to keep up with me'. Presentation: 'the best deck we've ever seen'.
Joy Adeyemi: interview notesProduct sense (Head of Design): 'Solid, slower.' Notes: spent 15 minutes asking who pays and why they don't pay online, then proposed one change: text-to-pay for the 58% of patients who pay by phone, with its trade-offs. Analytics (Data lead): notes say she raised payment completion at her last company from 33% to 47% over 18 months, and walked through an experiment that failed and what it taught her. Leadership (VP Engineering): notes say she coached a struggling PM into a different role, promoted another, and named her engineers when describing wins. Debrief comments: 'Not much presence.' 'Not sure she'd command the room with the CFO.'
Scores (1 to 4, entered after the debrief discussion)Dev: 4, 4, 4, 4. Joy: 3, 3, 3, 3. The Head of Design spoke first in the debrief. The VP Engineering's own notebook, written before the debrief, says: 'Joy: 4, strongest leader we've seen.' They entered a 3 after the discussion.
ReferencesDev: two references, both people who report to him now. Rosa knows someone at his previous company, who says: 'great in front of execs; his team turned over a lot.' Joy: her former manager and a peer, both strongly positive. Nobody has asked either reference about working with a CFO.
CompensationBand $220k to $250k. Dev is asking for $245k, Joy for $230k.
What a strong answer doesThe answer key the graders mark against

Recommends Joy, or at minimum not Dev on this evidence. Weighs both against the agreed scorecard rather than presence or the deck: Joy gives a metric she moved (33% to 47%), learns from failure, credits her team and coaches PMs, which fits execution, analytics and people leadership; Dev couldn't name a metric he moved, frames everything as 'I', jumped to features before asking about users, and the back channel reports high team turnover, all red flags. Points out that the scores were entered after a discussion the Head of Design opened, and the VP Engineering moved Joy from 4 to 3, so they show anchoring, not independent judgement. Treats 'presence with the CFO' as the one real open question and proposes a targeted reference check on how Joy worked with Finance leaders before the offer. Suggests the next loop collect written scores before the debrief.

Critical failures (cap the score at 40)
  • Recommends hiring Dev on the strength of the debrief scores or the presentation, without addressing that he couldn't name a metric he moved
Task

Staff-level · v1.0 · synthetic · healthcare, fintech, hiring, staff-level

The outputs

7 setups answered this brief. Pick up to three to read side by side.

Usable with a quick edit12 of 12 checks passed

The output is well-structured, evidence-based, and actionable. A PM could use it with minor wording edits, without re-checking evidence or redoing reasoning.

Output

To: Rosa Lindqvist From: Director of Product Re: Group PM, Patient Payments: recommendation

Recommendation

Hire Joy Adeyemi, conditional on two checks I can finish before Friday (below). I would not hire Dev on the current evidence. If the checks on Joy go badly, I would hold the search open rather than default to Dev, because his file has gaps too.

Why the recorded scores mislead

The recorded scores are Dev 4/4/4/4 and Joy 3/3/3/3, which would make this an easy call for Dev. But I don't trust them, for three reasons:

  • They were entered after the debrief discussion, and the Head of Design spoke first. That is a classic anchoring setup.
  • The VP Engineering's own pre-debrief notebook said "Joy: 4, strongest leader we've seen." They entered a 3 after the discussion. Uniform 3s for Joy and uniform 4s for Dev look like a group narrative, not independent assessments.
  • Dev's scores aren't supported by the interview notes (below).

Evidence, by scorecard area

Analytics (20%) - Dev was asked what metric he moved and answered "we grew the payments platform massively." When pressed, he said "the team shipped 40+ features." That is output, not outcome, and he never gave a number. The Data lead's "Strong" doesn't match their own notes. - Joy raised payment completion from 33% to 47% over 18 months and walked through an experiment that failed and what it taught her. That is directly relevant to our 41% online payment rate.

Product sense (25%) - Dev generated six feature ideas in five minutes, before asking who pays and why. The redesign was creative, but it started from solutions. - Joy spent 15 minutes on who pays and why they don't pay online. She then proposed one change, text-to-pay for the 58% of patients who pay by phone, and discussed its trade-offs. "Slower" is how it looked, but it is the better strategic instinct for a role whose first job is setting payments strategy.

People leadership (20%) - Dev told every story in the first person and said his last PM team "struggled to keep up with me." Two of his references are current direct reports, so they aren't independent. Your contact at his previous company said his team "turned over a lot." That is one informal source, but it is consistent with the interview. - Joy coached a struggling PM into a better-fit role, promoted another, and credited her engineers by name. The VP Engineering's original instinct was that she was the strongest leader we've seen. This role leads three PMs.

Execution and influence (35%, the heaviest weight) - Dev's strength is real here. His deck was "the best we've ever seen," and your contact says he is "great in front of execs." That is evidence of influence upward, though not of execution. - Joy's evidence is thinner. The only concerns are "not much presence" and "not sure she'd command the room with the CFO." Those are impressions, not observed behavior, and nobody has tested them. Her 14-point completion lift is execution evidence, but I want to confirm her role in it. - Neither file has much direct evidence on delivery, such as shipping through Finance and Compliance constraints. That is a gap for both.

My rescoring. These are my judgment, not data. I gave Dev a 3 or 4 on execution and influence, and Joy a 3.

Exec/influenceProduct senseAnalyticsLeadershipWeighted
Dev4322~3.0
Joy3344~3.4

Even generous to Dev on the heaviest-weighted category, Joy comes out ahead. The result is sensitive to the influence score, which is why the CFO check matters.

Risks

  1. Executive presence with the CFO. This is Joy's real risk, because the role requires working closely with Finance. But the feedback is vague and unverified, and "presence" comments are also where style bias tends to enter. We should test it rather than assume it.
  2. Attribution of the 33%→47% result. I haven't verified what she owned versus what the team or market drove.
  3. Dev's strengths are not nothing. He may be a strong operator with weak interview discipline on metrics. But the absence of a single number, combined with the team-turnover signal and the "struggled to keep up with me" comment, is a pattern. For a role leading three PMs, that is the riskier error.
  4. Process risk. If we hire on the recorded scores, we are acting on data we know was contaminated.

What I'd check before the offer

All of this fits before Friday:

  1. A 30-minute working session with the CFO and a Compliance lead. Joy would take a real case, such as a payment-plan or fee trade-off. Interviewers would score against defined criteria: clarity, handling of pushback, and command of the numbers. This tests the specific concern directly.
  2. Targeted reference calls on Joy. I would ask her former manager and peer how she handled disagreement with Finance or senior executives, with a specific example. I would also ask for one additional reference from a Finance or Compliance partner. Nobody has asked this yet.
  3. Verify the 33%→47% result. I would ask Joy what she personally owned and how it was measured, and ask her references to confirm.
  4. Ask the debrief group to put concrete observations behind "presence." I would also ask the VP Engineering what changed their score from 4 to 3. If the answer is "the discussion" and not new evidence, that tells us something.

If Dev stays in consideration, I would require two things first: a specific metric he moved and how, and two independent references (former peers or reports, not current ones). I am not proposing to hold the process open for him, but it is the bar I'd want met.

Compensation

Joy asks for $230k and Dev for $245k, in a $220–250k band. I would offer Joy at or near her ask. If the CFO session is strong, there is room to move up within the band if needed to close. Her ask leaves headroom for growth, and I don't want to lowball her because she asked for less.

Process note for next time

We should collect written scores independently before the debrief, and have the most senior person speak last. Otherwise the loudest early voice becomes the score.

Bottom line: Joy is the stronger fit on the evidence, with one open question (influence with the CFO) that I can resolve before Friday. I recommend we make the offer if the CFO session and reference checks come back clean.

Grades and run details

Decision model 100 · LLM judge 13 of 13 checks
Decision model checks
  • passUses the supplied evidence correctly64%
  • passAddresses the actual decision100%
  • passRespects explicit constraints50%
  • passIdentifies material uncertainty100%
  • passAvoids unsupported claims44%
  • passProduces the required deliverable96%
  • passWeighs against the agreed scorecard99%
  • passCatches the red flags in Dev's evidence100%
  • passSays what would change the call100%
  • passJudges on evidence, not presence76%
  • passDefines good for this role first68%
  • passKeeps each judgement independent96%
Run
Run
#1
API response time
34 s
Submitted
5 Oct 2026

Usable with a quick edit12 of 12 checks passed

The memo is clear, evidence-based, and actionable. A PM could use it with only minor wording edits, without re-checking facts or redoing reasoning.

Output

To: Rosa Lindqvist, VP Product From: Director of Product Re: Group PM, Patient Payments: hiring recommendation

---

Recommendation

Hire Joy Adeyemi, subject to two reference checks I can finish before Friday. If those checks go badly, I'd reopen the search rather than default to Dev.

The debrief scores (Dev 4/4/4/4, Joy 3/3/3/3) point the other way, so I'll explain why I'm not following them.

The scores don't reflect the evidence

All scores were entered after the group discussion, and the Head of Design spoke first. The clearest sign of how much the discussion moved people: the VP Engineering's notebook, written before the debrief, says "Joy: 4, strongest leader we've seen." They entered a 3 afterwards. We are looking at a consensus, not five independent assessments.

So I went back to the interview notes and scored them against the scorecard we agreed before interviews began.

Evidence against the scorecard

Execution and influence (35%) - Joy: She raised payment completion at her last company from 33% to 47% over 18 months. That is the exact problem this role exists to solve, since our online rate is 41%. - Dev: Asked twice what metric he moved, he answered "we grew the payments platform massively," then "the team shipped 40+ features." Shipping volume is output, not outcome. - Gap: On influence specifically, we have no direct evidence for either candidate on working with a CFO or Compliance. I come back to this under risks.

Product sense (25%) - Dev: He produced six feature ideas in five minutes before asking who pays or why. That is creative, but it is solution-first. - Joy: She spent 15 minutes on who pays and why they don't pay online. She then proposed one change, text-to-pay for the 58% who pay by phone, and laid out its trade-offs. That is the more relevant skill for setting a payments strategy. - I'd call this at least even, and arguably Joy's. It does not support Dev 4, Joy 3.

Analytics (20%) - Joy: She gave a quantified result and walked through a failed experiment and what she learned from it. - Dev: He gave no number when pressed. - Our Data lead wrote "Strong" for Dev, but the notes underneath don't support it. This is Joy's dimension clearly.

People leadership (20%) - Joy: She coached a struggling PM into a better-fitting role, promoted another, and credited her engineers by name. - Dev: Every story was told in the first person. He said his last team "struggled to keep up with me." Rosa's contact at his previous company reports that "his team turned over a lot." - This role leads three PMs. This is the dimension where the gap is widest, and the VP Engineering's independent view agreed.

Summary: On the evidence, Joy is ahead on analytics and leadership, at least even on product sense, and ahead on demonstrated execution. Dev's advantages are presentation and room presence. Those are real, but they aren't what we weighted.

Risks

Hiring Joy - Executive presence. The debrief concern, "not sure she'd command the room with the CFO," is legitimate for a role that depends on Finance and Compliance. It is also an impression, not evidence. Nobody has asked her references about it, and nobody observed her with a finance stakeholder. - Pace. "Slower" could mean considered or could mean slow to decide. Her 18-month track record suggests sustained progress, but I want to confirm it. - Verification. The 33% to 47% figure is self-reported.

Hiring Dev (for completeness) - No evidence he moves outcomes. We have no number, only output. - Team retention. The turnover signal comes from one secondhand source, so it's not conclusive. It does match his own interview language, and three PMs would report to him. - Weak references. Both of his references currently report to him, which limits how candid they can be. - Cost. He is asking $245k, near the top of the $220k–$250k band, against Joy's $230k. This is minor next to the other points.

Hiring neither - The role stays open while the online payment rate sits at 41%. I'd only choose this if Joy's checks fail.

What I'd check before making the offer (by Thursday)

1. Call Joy's former manager and peer again with targeted questions. - How did she handle a CFO or finance leader who disagreed with her? - Can you give an example of her winning an argument with a senior executive? - Can you confirm the 33% to 47% completion figure, and how much of it was her doing? 2. Add a reference who saw her work with Finance or Compliance, ideally someone from a finance function at her last company. 3. Optional, if timing allows: a 30-minute conversation between Joy and our CFO or a senior Finance partner, framed as mutual fit. This tests the presence concern directly instead of debating it.

Decision rule: If the references describe someone who holds her ground with finance leadership and confirm the result, I'd offer at or near her $230k ask. If they confirm the presence concern, I'd recommend we reopen the search. Dev's leadership and outcome gaps are more serious for this role than a presence gap. If you want Dev kept live as a fallback, I'd first need a reference from a former peer or manager, not a current report, and a concrete metric he owned.

Process note for future loops

We should have interviewers submit scores before the debrief, and rotate who speaks first. The VP Engineering's notebook shows we lost an independent signal this time, and it was the one that mattered most.

Happy to walk through this before Friday.

Grades and run details

Decision model 100 · LLM judge 13 of 13 checks
Decision model checks
  • passUses the supplied evidence correctly88%
  • passAddresses the actual decision100%
  • passRespects explicit constraints57%
  • passIdentifies material uncertainty100%
  • passAvoids unsupported claims59%
  • passProduces the required deliverable98%
  • passWeighs against the agreed scorecard100%
  • passCatches the red flags in Dev's evidence100%
  • passSays what would change the call100%
  • passJudges on evidence, not presence96%
  • passDefines good for this role first77%
  • passKeeps each judgement independent98%
Run
Run
#1
Time to output
27 s
Submitted
5 Oct 2026

Usable with a quick edit12 of 12 checks passed

The output is well-structured, evidence-based, and actionable; a PM could use it with minor wording edits without re-checking evidence or redoing reasoning.

Output

To: Rosa Lindqvist, VP Product Subject: Group PM recommendation — Joy Adeyemi

Recommendation: Hire Joy, subject to focused reference and influence checks before the offer. She has the stronger evidence against the role’s core needs: measurable payment outcomes, thoughtful product judgment, and the ability to lead through other people. The CEO needs a decision by Friday; I would not let the debrief scores alone drive it.

Evidence

  • Execution and influence (35%): Joy gave a specific result from her last company—payment completion rose from 33% to 47% over 18 months—and explained a failed experiment and what she learned. Her leadership examples showed investment in others: coaching a struggling PM into a different role and promoting another. Dev’s evidence was less concrete: he described platform growth and 40+ shipped features but supplied no outcome metric when pressed. His stories were consistently framed around his own contribution, and he said his team struggled to keep up with him.
  • Product sense (25%): Joy investigated who pays and why before proposing text-to-pay for the 58% who currently pay by phone, including trade-offs. That is relevant to our 41% online payment rate and reflects problem-first judgment. Dev’s creativity and presentation were strengths, but six ideas in the first five minutes—before clarifying who pays and why—are a concern for this role.
  • Analytics (20%): Joy’s quantified result and discussion of a failed experiment are stronger evidence of analytical ownership. Dev did not provide a measurable result.
  • People leadership (20%): Joy’s examples demonstrate coaching, development, and shared credit. Dev’s confidence and presentation are positives, but they do not offset the limited evidence of developing a team or his comment about the team struggling to keep up.

Risks and scorecard

Joy’s main risk is whether she can influence senior stakeholders, particularly the CFO and Finance. The “not much presence” and “command the room” comments are concerns to test, not proof that she cannot do it; we should assess her ability to make a clear, evidence-based case rather than equate influence with a particular presentation style.

Dev may be compelling with executives, but we need to understand the team-turnover signal and whether his leadership style scales. His two references are current direct reports, which may limit candor. His former-company reference’s report of high turnover is important but unverified.

I would treat the 4/4/4/4 versus 3/3/3/3 ratings as weak evidence. The VP Engineering’s pre-debrief note called Joy the strongest leader seen and rated her a 4, then the score became a 3 after discussion; the Head of Design also spoke first. This does not establish why the ratings changed, but it is reason to rely on the underlying evidence and check for independent calibration.

Before making the offer

  1. Ask Joy’s former manager and peer specifically about her influence with Finance/CFO-level stakeholders, including a difficult disagreement and the outcome.
  2. Ask for one additional Joy reference who can speak to cross-functional execution; ask Dev’s references about team turnover, feedback, and how he develops PMs. Seek candid, comparable evidence where possible.
  3. In a short structured conversation, ask Joy to make the case for a payment-rate strategy to a skeptical CFO. Evaluate clarity, use of evidence, and handling of challenge—not polish alone.
  4. Confirm Joy’s 33%→47% result, her role in it, and the scope of her people leadership.

If those checks reveal a material gap in Joy’s influence or ownership, I would choose neither rather than default to Dev on presentation strength. Otherwise, proceed with Joy at her requested $230k, within band.

Grades and run details

Decision model 100 · LLM judge 13 of 13 checks
Decision model checks
  • passUses the supplied evidence correctly87%
  • passAddresses the actual decision100%
  • passRespects explicit constraints76%
  • passIdentifies material uncertainty100%
  • passAvoids unsupported claims60%
  • passProduces the required deliverable99%
  • passWeighs against the agreed scorecard100%
  • passCatches the red flags in Dev's evidence100%
  • passSays what would change the call100%
  • passJudges on evidence, not presence100%
  • passDefines good for this role first69%
  • passKeeps each judgement independent87%
Run
Run
#1
API response time
12 s
Submitted
5 Oct 2026

Check by check

Where the outputs differ comes first. Open a check to read the judge’s reasons. “Mixed” means the two graders disagreed.

All got right 12

Uses the supplied evidence correctlyRightRightRight
Sonnet 5.5 · API

Every claim about the current situation is directly supported by the supplied context, with no invented facts.

Opus 5.5 · Claude

Every claim about the current situation is directly supported by the brief or supplied context.

GPT-6 Luna · API

All factual claims about the current situation are directly from the brief, with no inventions.

Addresses the actual decisionRightRightRight
Sonnet 5.5 · API

The output commits early to hiring Joy, conditional on checks, says not to hire Dev, and states that if checks go badly it would hold the search open.

Opus 5.5 · Claude

The output unambiguously recommends hiring Joy, subject to reference checks, and states what would change the call.

GPT-6 Luna · API

Recommends hiring Joy early, framed for Rosa, and states that material gaps in influence or ownership would change the call to neither.

Respects explicit constraintsRightRightRight
Sonnet 5.5 · API

The output is a memo to Rosa, under 1,000 words, and respects the form and length requested.

Opus 5.5 · Claude

The output is a recommendation to Rosa Lindqvist, under 1,000 words, and respects all stated constraints.

GPT-6 Luna · API

Memo to Rosa, under 1000 words, respects the Friday deadline by making a decision now with pre-offer checks.

Identifies material uncertaintyRightRightRight
Sonnet 5.5 · API

It names the unknowns (CFO presence, attribution of the metric, Dev's pattern) and specifies how to resolve them with targeted checks before the offer.

Opus 5.5 · Claude

It names specific unknowns (executive presence with CFO, pace, metric verification) and how to resolve them before the offer.

GPT-6 Luna · API

Names unknowns (Joy's CFO influence, Dev's turnover, debrief score reliability) and specifies how to resolve them with targeted checks.

Avoids unsupported claimsRightRightRight
Sonnet 5.5 · API

Interpretations like anchoring and output vs outcome are clearly presented as the author's judgment, not as established fact.

Opus 5.5 · Claude

Interpretations like the scores reflecting consensus are clearly presented as such, not as established fact.

GPT-6 Luna · API

Interpretations are clearly labelled as risks or weak evidence, not presented as established facts.

Produces the required deliverableRightRightRight
Sonnet 5.5 · API

The memo is complete, addressed to Rosa, within the word limit, and actionable with a clear recommendation and next steps.

Opus 5.5 · Claude

The memo is complete, addressed to the VP Product, within the word limit, and actionable as is.

GPT-6 Luna · API

Complete recommendation memo with evidence, risks, and actionable next steps, usable as is.

Weighs against the agreed scorecardRightRightRight
Sonnet 5.5 · API

The recommendation walks through each scorecard competency with evidence for both candidates and notes that presence and the deck are not on it.

Opus 5.5 · Claude

The recommendation explicitly weighs both candidates against the agreed scorecard competencies and weighting, and notes that presence and the deck are not on it.

GPT-6 Luna · API

Weighs both candidates against the agreed competencies (execution, product sense, analytics, people leadership) and notes that presence and deck are not on it.

Catches the red flags in Dev's evidenceRightRightRight
Sonnet 5.5 · API

It names at least four red flags: no metric moved, first-person credit and team struggled to keep up, jumping to features, and back-channel turnover report.

Opus 5.5 · Claude

It names at least three red flags: no metric moved, first-person credit and team struggled to keep up, jumping to features, and back-channel turnover.

GPT-6 Luna · API

Names at least three red flags: no metric moved, first-person credit and team struggled to keep up, jumping to features, and back-channel turnover report.

Says what would change the callRightRightRight
Sonnet 5.5 · API

It identifies Joy's influence with the CFO as the open question and proposes a working session with the CFO and targeted reference checks to resolve it.

Opus 5.5 · Claude

It identifies Joy's influence with senior finance leaders as the open question and proposes targeted reference checks and a CFO conversation, with a clear decision rule.

GPT-6 Luna · API

Identifies Joy's influence with CFO as the open question and proposes specific reference checks and a structured conversation to test it.

Judges on evidence, not presenceRightRightRight
Sonnet 5.5 · API

It weighs concrete evidence (metric moved, team credit, learning from failure) and treats impressions like presence as weak, unverified signals.

Opus 5.5 · Claude

The output judges on concrete evidence (metric moved, team credit, learning from failure) and treats impressions as weak signals.

GPT-6 Luna · API

Judges on concrete evidence (metric moved, team credit, learning from failure) and treats impressions like presence as weak signals.

Defines good for this role firstRightRightRight
Sonnet 5.5 · API

It sets out the scorecard competencies and weightings for the role and uses them to define strong and weak signals before judging candidates.

Opus 5.5 · Claude

It uses the pre-agreed scorecard with weighted competencies to define what good looks like for this role before judging candidates.

GPT-6 Luna · API

Sets out the weighted competencies and what strong evidence looks like before comparing candidates.

Keeps each judgement independentRightRightRight
Sonnet 5.5 · API

It points out that scores were entered after discussion, notes the VP Engineering's pre-debrief score changed, and recommends written independent scores before debrief.

Opus 5.5 · Claude

It points out that scores were entered after discussion, notes the VP Engineering's changed score, and recommends written independent evaluations before debrief.

GPT-6 Luna · API

Relies on the VP Engineering's independent pre-debrief note and discounts group-influenced scores, keeping judgement independent.

Results

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

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6.1 SolwithAPI94.4100.03None
2Sonnet 5.5withAPI95.896.22None
3Gemini 3.8 FlashwithAPI89.692.32None
4GPT-6 AstrawithChatGPT95.887.23None
5Opus 5.5withClaude91.787.23None
6GPT-6 LunawithAPI93.184.63None
7Gemini 3.5 Flash-LitewithGemini75.064.13None

About the task

The PM job

Hiring product managers.

Why it matters

A bad PM hire costs a team a year. Most loops reward polish and presence, and the debrief goes to whoever speaks first.

What good looks like

  • Defines what good looks like for this role before judging anyone
  • Judges on evidence of results and how they work with a team
  • Tests the real job, not a rehearsed framework
  • Keeps each interviewer's judgement independent

Deliberately not measured

  • Employment law
  • Compensation benchmarking
Capability tested

Hiring judgement

The failure we’re looking for

Hiring for presence and polish over evidence

Grading

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