Humanity’s Last PRD
Latest drop: Haiku 5.5 · 8 Oct
Tasks / Leadership

Write a performance review

Can the model write a review that is clear, fair and specific enough to change what the person does next?

Task type v1.1 · 2 tasksLast changed 6 Oct 2026 · ChangelogDifficulty

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

From 16 graded outputs by 8 models. 50% were usable with at most a quick edit.

Reliably right

  1. Judges outcomes, not activity100% pass
    Every rating change and keep is justified by outcomes against goals, with shipping and activity treated as context, not the result.
    GPT-6 Astra · ChatGPT · Fourteen ratings, three managers
  2. Owns the manager's part100% pass
    Ana plainly acknowledges her Q3 mistake and commits to reviewing goal metrics first in every monthly 1:1.
    Opus 5.5 · Claude · Shipped a lot, moved nothing
  3. Clear and direct, with care97% pass
    The note to Ravi is direct, specific about what needs to change, and framed to help him succeed; the whole pre-read is clear and actionable.
    GPT-6 Astra · ChatGPT · Fourteen ratings, three managers

Where it slips

  1. Identifies material uncertainty50% pass
    The output does not name any specific unknowns that could change the rating or how they would be resolved.
    Opus 5.5 · Claude · Shipped a lot, moved nothing
  2. Notices who's underrated69% pass
    It does not propose Exceeds or ask Sara to justify Meets despite Nora beating both goals; it merely keeps Meets, missing the underrating.
    GPT-6 Luna · API · Fourteen ratings, three managers
  3. Catches the inflated team78% pass
    The output proposes lowering Ben from Exceeds to Meets, failing to keep Ben at Exceeds as required; Ben beat both outcome goals and should remain Exceeds.
    Opus 5.5 · Claude · Fourteen ratings, three managers

How it’s graded

Each check is built on how world-class product and growth leaders do this work, as they’ve described it on Lenny’s Podcast, and names the leader it comes from.

  1. Clear and direct, with care

    Does the review say plainly what needs to change, in words the person couldn't misread, while showing it is written to help them succeed?

    Passes when Every main point is stated directly, with the specific change expected and the support on offer.

    Kim ScottLenny’s Podcast

  2. Judges outcomes, not activity

    Does the review judge the person on the outcomes they achieved against their goals, rather than on how much they shipped or how busy they were?

    Passes when Rates against the goals' outcomes, with the evidence, and treats output as context, not the result.

    Product and growth leadersOn Lenny’s Podcast

  3. Names gaps you could see

    Is each area to improve described as specific, observable behaviour with what good would look like, rather than labels like 'be more strategic'?

    Passes when Every gap names what the person did or didn't do, in a specific situation, and what doing it well would look like.

    Product and growth leadersOn Lenny’s Podcast

  4. Weighs the whole period

    Does the review weigh the whole review period, rather than letting one recent event or one strong impression decide the verdict?

    Passes when Covers the whole period in proportion, and explicitly guards against a recent event or one impression dominating.

    Product and growth leadersOn Lenny’s Podcast

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 compare up to three outputs, each with the LLM judge’s verdict on every check. Highlights mark what a PM had to fix.

The brief

You're the Director of Product at Northgate, and you're facilitating Thursday's calibration for 14 PMs across three managers (Ravi, Sara and Tom). Write the pre-read, in no more than 1,100 words: the ratings you'd change and why, each with the evidence, the ones you'd keep, and a short note to Ravi about his ratings. Use the guidance below.\n\nThe proposed ratings and evidence are attached.

What the model was given2 items: calibration_guidance.md, proposed_ratings.csv (one row per PM, as each manager submitted it)
calibration_guidance.md7 lines · Download
# Calibration guidance

- Ratings judge outcomes against goals first, then how the work was done. Shipping is not an outcome.
- Any rating other than Meets needs written evidence of outcomes.
- A Below rating needs documented earlier feedback on the gap.
- Expected spread across the org: most PMs Meets; Exceeds for clear outperformance on outcomes.
- The calibration facilitator proposes changes; managers can respond before ratings are final.
proposed_ratings.csv (one row per PM, as each manager submitted it)pm,manager,proposed_rating,goal_1,goal_1_result,goal_2,goal_2_result,evidence_cited_by_manager Aisha,Ravi,Greatly exceeds,Raise trial conversion 8% to 11%,8.4%,Ship self-serve billing,Shipped,Shipped 9 features; great energy; leadership loves her Ben,Ravi,Exceeds,Cut onboarding time 10 to 5 days,4.5 days,Raise activation 30% to 38%,39%,Hit both goals Chloe,Ravi,Exceeds,Grow API usage 20%,+4%,Launch partner portal,Shipped,Partner portal launched on time; strong specs Dev,Ravi,Exceeds,Reduce support tickets per account 25%,-27%,NPS from 31 to 40,41,Hit both goals; mentored two APMs Emma,Ravi,Exceeds,Expansion revenue +15%,+6%,Ship usage dashboard,Shipped,Very responsive to Sales; shipped 6 Sales requests Femi,Sara,Meets,Raise weekly retention 41% to 46%,46%,Cut churn of new accounts 20%,-22%,"Hit both goals, steady" Gita,Sara,Below,Raise invoices paid online 35% to 45%,47%,Launch reminders v2 by September,"Launched, rolled back after 4 days (support not briefed)",Reminders launch in September was a mess Hugo,Sara,Meets,Ship mobile app v2,Shipped late,Mobile weekly actives +20%,+3%,Delivered a hard project Ines,Tom,Exceeds,Cut checkout drop-off 30% to 22%,21%,Launch two payment methods,Shipped,"Hit both, specific evidence" Jon,Tom,Meets,Search success 60% to 70%,66%,Ship filters,Shipped,"Good progress, more to do" Kai,Tom,Meets,Reduce refund rate 4% to 3%,3.1%,Fraud losses -20%,-24%,Solid year Lena,Tom,Below,Grow marketplace listings 25%,+9%,Launch seller analytics,Not launched,"Missed both goals; feedback given in Q2 and Q3 1:1s, documented" Mo,Tom,Meets,Cut time to first sale for new sellers 21 to 14 days,15 days,Seller NPS +5,+6,"Nearly hit, good stakeholder work" Nora,Sara,Meets,Raise payment success 92% to 95%,95.4%,Cut payment support tickets 30%,-35%,"Quiet, delivered"
What a strong answer doesThe answer key the graders mark against

Judges each PM on outcomes against their goals, as the guidance says. Ravi rated all five of his PMs Exceeds or above, but only Ben and Dev hit both goals; Aisha (Greatly exceeds) missed her conversion goal (8.4% against 11%), Chloe missed API usage (+4% against +20%) and Emma missed expansion (+6% against +15%), and his evidence is output ('shipped 9 features', 'responsive to Sales'). Proposes Aisha, Chloe and Emma at Meets (or lower for Chloe, whose outcome goal was far off), keeps Ben and Dev at Exceeds. Gita (Sara, Below) beat her main goal (47% against 45%) and the Below rests on one September launch, with no documented feedback, which the guidance requires: Meets, with the launch-readiness gap documented now. Nora (Sara, Meets) beat both goals: a candidate for Exceeds. Lena (Tom, Below) has documented feedback and missed both goals: Below stands. Hugo shipped late and missed his outcome: Meets is generous but defensible, or Below needs documentation first. The note to Ravi is direct and respectful: rate on outcomes, not activity, with what evidence he'd need.

Critical failures (cap the score at 40)
  • Keeps Aisha at Greatly exceeds
Task

Staff-level · v1.0

The outputs

8 setups answered this brief.

GPT-6 AstraChatGPT100Usable with a quick editGPT-6.1 SolAPI100Usable with a quick editHaiku 5.5API100Usable with a quick editGemini 3.8 FlashAPI98Usable with a quick editOpus 5.5Claude88Needs more workGPT-6 LunaAPI82Needs more workSonnet 5.5API77Needs more workGemini 3.5 Flash-LiteGemini38Needs more work
GPT-6 LunaAPI

Needs more than a quick edit11 of 13 checks passed

Nora's rating is not correctly addressed; a PM would need to re-evaluate her case for Exceeds. The output also lacks observable behavior gaps, though that is less critical for a calibration pre-read.

Output

Thursday calibration pre-read

Proposed changes below are for discussion, not final ratings. We’ll assess outcomes against goals first; shipping, effort, and stakeholder praise are not outcomes by themselves. Managers can share additional evidence before ratings are finalized.

Ratings I’d change

  • Aisha — Greatly exceeds → Meets. Trial conversion reached 8.4% against an 11% goal. Self-serve billing shipped, but shipping is not an outcome, and the cited nine features, energy, and leadership praise don’t establish goal outperformance.
  • Chloe — Exceeds → Meets. API usage grew 4% against a 20% goal. The partner portal shipped, but no result from the launch is cited; strong specs and on-time delivery don’t establish Exceeds.
  • Emma — Exceeds → Meets. Expansion revenue grew 6% against a 15% goal. The dashboard shipped, and six Sales requests were delivered, but neither is evidence of the targeted revenue outcome or clear outperformance.
  • Ines — Exceeds → Meets. Checkout drop-off reached 21% against a 22% goal, while two payment methods shipped. The checkout result beat its target by one point; for the payment-method goal, no outcome beyond shipping is cited. The evidence supports a solid year, not clear outperformance on outcomes.
  • Gita — Below → Meets on the current record. Online invoices reached 47% against a 45% goal. Reminders v2 launched but was rolled back after four days because Support had not been briefed. That is a serious rollout issue, but the record includes no earlier documented feedback on the gap, as required for Below. Please share any such documentation before calibration closes.

Ratings I’d keep

  • Ben — Exceeds. Onboarding fell to 4.5 days against a 5-day goal, and activation reached 39% against 38%: both measured outcomes surpassed target.
  • Dev — Exceeds. Support tickets fell 27% against a 25% goal, and NPS reached 41 against 40. Both outcomes beat their goals; mentoring two APMs is additional contribution.
  • Femi — Meets. Weekly retention reached its 46% goal, and new-account churn fell 22% against a 20% goal.
  • Hugo — Meets, with a material concern to discuss. The app shipped late and mobile weekly actives rose 3% against a 20% goal. The evidence shows a missed outcome; no earlier documented feedback is cited to support Below.
  • Jon — Meets. Search success reached 66% against 70%, and filters shipped. This is progress, not goal attainment; the record does not establish a case for a non-Meets rating.
  • Kai — Meets. Fraud losses fell 24% against a 20% goal; refund rate was 3.1% against a 3% goal. Mixed results support Meets.
  • Lena — Below. Listings grew 9% against a 25% goal, and seller analytics did not launch. Earlier feedback on the gap is documented in Q2 and Q3 1:1s, meeting the Below requirement.
  • Mo — Meets. Time to first sale was 15 days against a 14-day goal; seller NPS rose 6 points against a 5-point goal. One goal narrowly missed, the other was exceeded.
  • Nora — Meets. Payment success reached 95.4% against a 95% goal, and support tickets fell 35% against a 30% goal. Both outcomes beat target, but the evidence supports Meets rather than clear outperformance.

Note to Ravi

Ravi, several of your proposed high ratings cite launches, feature volume, responsiveness, or praise alongside missed outcome goals. For calibration, please anchor any Exceeds or Greatly Exceeds case in clear outcome outperformance; shipping alone is not enough. Please bring any additional outcome evidence you believe changes these recommendations.

Grades and run details

Decision model 92 · LLM judge 79
Decision model checks
  • passUses the supplied evidence correctly87%
  • passAddresses the actual decision52%
  • passRespects explicit constraints39%
  • partialIdentifies material uncertainty31%
  • passAvoids unsupported claims69%
  • passProduces the required deliverable62%
  • passCatches the inflated team100%
  • passRescues Gita from one bad launch100%
  • partialNotices who's underrated41%
  • passClear and direct, with care85%
  • passJudges outcomes, not activity97%
  • passNames gaps you could see78%
  • passWeighs the whole period64%
Run
Run
#1
API response time
25 s
Submitted
5 Oct 2026
Haiku 5.5API

Usable with a quick edit13 of 13 checks passed

The pre-read gives specific changes, evidence, and a clear note to Ravi; it can be used with light edits and does not require re-checking facts.

Output

Calibration pre-read: 14 PMs, Ravi / Sara / Tom

Purpose: Thursday's calibration. Managers can respond to these proposals before ratings are final.

Rules I applied (from calibration guidance): - Ratings start from outcomes against goals. Shipping is an output, and it doesn't count as an outcome. - Any rating other than Meets needs written outcome evidence. - Below needs documented earlier feedback on the gap. Where that's missing, I've capped the rating at Meets. - Working definition for Exceeds (for discussion): both outcome goals beaten. One beat and one hit is Meets.

Summary: 6 changes, 1 hold, 7 keeps. Proposed spread after changes: 3 Exceeds, 10 Meets, 1 Below.

---

Ratings I'd change

Aisha (Ravi): Greatly exceeds → Meets - Trial conversion reached 8.4% against an 11% target. That's about 13% of the intended 3-point lift. - "Shipped self-serve billing" is an output. Nine features shipped is also an output. - "Great energy" and "leadership loves her" are not outcome evidence. - Greatly exceeds is not supported. Below would need documented feedback on the conversion gap, which the submission doesn't cite.

Chloe (Ravi): Exceeds → Meets - API usage grew 4% against a 20% target, which is one-fifth of the goal. - The partner portal shipped on time. That's an output, and it doesn't offset the miss. - "Strong specs" describes how the work was done, not what it achieved. - Ravi should show what the portal changed in usage, if anything. If the outcome is attributable and I've missed it, I'll revisit.

Emma (Ravi): Exceeds → Meets - Expansion revenue grew 6% against a 15% target, which is 40% of the goal. - The usage dashboard shipped, and six Sales requests were delivered. Both are outputs. Responsiveness to Sales is a good working habit, but it isn't an outcome.

Ines (Tom): Exceeds → Meets (open to revisiting) - Checkout drop-off reached 21% against a 22% target. That's a clean beat, but only by one point. - Two payment methods shipped. The record has no adoption or conversion metric for them, so that goal is an output. - Under the working definition, one beat plus one output is Meets. If Tom has adoption data for the new payment methods, I'll revisit Exceeds on Thursday.

Nora (Sara): Meets → Exceeds - Payment success reached 95.4% against a 95% target. - Payment-related support tickets fell 35% against a 30% target. - Both outcome goals were beaten, the same pattern as Ben. "Quiet" isn't a criterion, and the outcomes support a higher rating than proposed.

---

Ratings on hold (need documentation before Thursday)

Gita (Sara): Below → Meets (Below only if documentation exists) - Invoices paid online reached 47% against a 45% target. That's an outcome hit. - Reminders v2 launched, was rolled back after four days, and support wasn't briefed. That's a real failure in how the work was done, and it needs a written account and a postmortem. - Below requires documented earlier feedback on the gap. None is cited, and one of two outcome goals was hit. - Sara: if you have documented feedback on the launch-readiness issue, I'll reconsider Below.

Hugo (Sara): Meets (hold) - Mobile weekly actives grew 3% against a 20% target, which is 15% of the goal. The app shipped late. - "Delivered a hard project" is an output description, and the outcome miss is large. - Meets is the ceiling the current record supports. If the mobile gap was raised in 1:1s before Q3, Below is defensible. Sara, please confirm before Thursday.

---

Ratings I'd keep

PMRatingWhy
Ben (Ravi)ExceedsBoth beaten: onboarding 4.5 days vs 5; activation 39% vs 38%. Cite the numbers, not "hit both goals."
Dev (Ravi)ExceedsBoth beaten: support tickets per account −27% vs −25%; NPS 41 vs 40. Mentoring two APMs is a useful addition.
Femi (Sara)MeetsRetention exactly at target (46%); churn beat (−22% vs −20%). One beat, one hit.
Jon (Tom)MeetsSearch success 66% vs 70%, about 60% of the gain. Filters shipped. Partial progress is Meets.
Kai (Tom)MeetsRefund rate 3.1% vs 3.0%, a narrow miss. Fraud losses −24% vs −20%, a beat.
Lena (Tom)BelowBoth goals missed (listings +9% vs +25%; seller analytics not launched). Feedback documented in Q2 and Q3 1:1s. Tom, please attach those notes to the file.
Mo (Tom)MeetsTime to first sale 15 days vs 14 target, about 86% of the gain. Seller NPS +6 vs +5.

---

Note to Ravi

Ravi, five of your five ratings are above Meets, and three of them (Aisha, Chloe, Emma) rest on shipping, responsiveness and reputation. The outcome numbers in the same rows don't support those ratings: 8.4% vs 11%, +4% vs +20%, and +6% vs +15%.

Ben and Dev are your strongest calls because their evidence is outcomes, and that's the standard we should use for all five.

I'm not suggesting Aisha, Chloe or Emma had weak years. Nine features and an on-time portal are real work. But ratings have to be tied to the goals we set.

Before Thursday, could you: 1. For Aisha, Chloe and Emma, write the outcome analysis for each goal. Include what was tried, what moved, and what you think caused the gap. 2. Tell us whether any feedback on those gaps was documented earlier in the year. If it was, Below becomes available. 3. Respond to the proposed changes, so we can settle them on Thursday.

Happy to walk through this 1:1 before the session.

---

Expected spread after changes

  • Greatly exceeds: 0
  • Exceeds: 3 (Ben, Dev, Nora)
  • Meets: 10
  • Below: 1 (Lena), with Gita and Hugo pending documentation

This is within the guidance: most PMs at Meets, Exceeds reserved for clear outperformance on outcomes.

Grades and run details

Decision model 100 · LLM judge 100
Decision model checks
  • passUses the supplied evidence correctly70%
  • passAddresses the actual decision86%
  • passRespects explicit constraints33%
  • passIdentifies material uncertainty98%
  • passAvoids unsupported claims67%
  • passProduces the required deliverable85%
  • passCatches the inflated team100%
  • passRescues Gita from one bad launch100%
  • passNotices who's underrated100%
  • passClear and direct, with care100%
  • passJudges outcomes, not activity100%
  • passNames gaps you could see79%
  • passWeighs the whole period79%
Run
Run
#1
API response time
47 s
Submitted
8 Oct 2026
Sonnet 5.5API

Needs more than a quick edit10 of 13 checks passed

The output incorrectly lowers Ben and Dev to Meets and fails to propose Exceeds for Nora, requiring re-evaluation of those ratings before use.

Output

Pre-read: Product calibration, Thursday

Scope: 14 PMs, 3 managers. I propose changing 7 ratings and keeping 7. These are proposals, and managers can respond before anything is final.

Headline: Ravi's team is rated 5 of 5 at Exceeds or above, and the cited evidence is mostly activity: features shipped, responsiveness, energy. Under our guidance, shipping is not an outcome. After my proposed changes, the spread is 13 Meets, 1 Below and 0 Exceeds. Managers who think someone clearly outperformed should bring the outcome evidence.

Ratings I'd change

Exceeds rated on shipping, with the main outcome missed

Aisha (Ravi): Greatly exceeds → Meets - Trial conversion went from 8% to 8.4% against a target of 11%. - Self-serve billing shipped, but no outcome is cited. - The evidence offered is "9 features, great energy, leadership loves her." None of that is an outcome. - "Greatly exceeds" isn't a category in our guidance. - The outcome data would point lower than Meets. I'm not proposing Below because no earlier feedback is documented, which the guidance requires. That gap belongs to the manager.

Chloe (Ravi): Exceeds → Meets - API usage grew 4% against a 20% target. - The partner portal launched on time, but no outcome is cited. - Strong specs and an on-time launch are good craft, not outcomes. - The same Below caveat applies as for Aisha.

Emma (Ravi): Exceeds → Meets - Expansion revenue rose 6% against a 15% target. - The usage dashboard shipped, but no outcome is cited. - "Shipped 6 Sales requests" and responsiveness describe throughput, not results. - The same Below caveat applies.

Hit both goals, but not clear outperformance

The guidance reserves Exceeds for clear outperformance. These three cleared their targets by small margins, and the same result is rated Meets elsewhere (Femi, Nora, Kai).

Ben (Ravi): Exceeds → Meets - Onboarding went from 10 to 4.5 days against a 5-day target. - Activation reached 39% against a 38% target. - This is the closest call of the seven. If Ravi can show that the goals were unusually hard or the business impact was outsized, I'm open to changing my view.

Dev (Ravi): Exceeds → Meets - Tickets per account fell 27% against a 25% target. - NPS reached 41 against a 40 target. - Mentoring two APMs is real and belongs in the written feedback. It doesn't offset margins this thin.

Ines (Tom): Exceeds → Meets - Checkout drop-off fell to 21% against a 22% target. - Two payment methods shipped, but no outcome is cited for them. - Tom's note says "specific evidence" without stating it. I'd need to see it.

Below without documented feedback

Gita (Sara): Below → Meets - Goal 1 was beaten: 47% of invoices paid online against a 45% target. - Goal 2, the reminders v2 launch, was rolled back after 4 days because support wasn't briefed. That is a real execution miss and should be coached. - The guidance requires documented earlier feedback for a Below, and none is cited. One goal beaten and one launch failure on a cross-functional step is not a Below without that record. - The written feedback should name the support-briefing failure plainly.

Ratings I'd keep

  • Femi (Sara), Meets: retention reached 46% (target 46%) and new-account churn fell 22% (target 20%). Both goals were hit.
  • Nora (Sara), Meets: payment success reached 95.4% (target 95%) and payment tickets fell 35% (target 30%). This is a strong result, but not clear outperformance on both goals.
  • Kai (Tom), Meets: refund rate reached 3.1% (target 3%) and fraud losses fell 24% (target 20%). One goal was a near miss and one was beaten.
  • Mo (Tom), Meets: time to first sale reached 15 days (target 14) and seller NPS rose 6 (target 5). One goal was a near miss and one was beaten.
  • Lena (Tom), Below: listings grew 9% against a 25% target, and seller analytics did not launch. Feedback was given and documented in Q2 and Q3 1:1s, so this meets the bar for Below. Tom should confirm the written evidence is in the file before ratings are final.

Two Meets I'm flagging, not changing

  • Hugo (Sara): Mobile app v2 shipped late, and weekly actives rose 3% against a 20% target. "Delivered a hard project" describes shipping, not outcomes. I'm leaving Meets because the same logic that stops me moving Aisha, Chloe and Emma to Below applies here, and consistency matters. Sara, please come ready to discuss the outcome miss and any earlier feedback.
  • Jon (Tom): Search success reached 66% against a 70% target (from 60%), and filters shipped. "Good progress, more to do" is a reasonable Meets but a thin write-up. Please add the actual outcome numbers.

Note to Ravi

Ravi,

Your ratings are the ones I'm changing most, so I want to explain my thinking before Thursday.

I'm not questioning your team's effort or your read on their potential. Ben and Dev hit both goals, and that counts. The issue is that all five ratings sit at Exceeds or above, and for three of them (Aisha, Chloe, Emma) the cited evidence is shipping and responsiveness while the primary outcome was missed by a wide margin. Aisha's 8.4% against an 11% target is the clearest case. "Leadership loves her" will not hold up in calibration, and it's unfair to Sara's and Tom's reports, who are rated Meets on similar or stronger outcomes.

I'm proposing Meets for all five. For Ben and Dev, I'll reconsider if you can show the goals were unusually hard or the impact was outsized. For Aisha, Chloe and Emma, I'd need outcome evidence I haven't seen, such as revenue or usage attributed to the work that shipped.

There is a second issue. If three of your reports missed key outcomes and there was no documented feedback, they didn't hear about it in time to change course. That is a management gap, and it's why I can't propose Below even where the numbers might support it. Please start documenting feedback now, and bring whatever exists on Thursday.

Please respond before Thursday with any evidence I've missed.

What I need before Thursday

  • Ravi: outcome evidence for Aisha, Chloe and Emma, plus any feedback records.
  • Sara: the outcome explanation for Hugo and any feedback history for Gita.
  • Tom: Ines's specific evidence, Lena's feedback documents and Jon's outcome numbers.

Grades and run details

Decision model 88 · LLM judge 71
Decision model checks
  • passUses the supplied evidence correctly57%
  • passAddresses the actual decision81%
  • partialRespects explicit constraints22%
  • passIdentifies material uncertainty95%
  • partialAvoids unsupported claims42%
  • passProduces the required deliverable79%
  • passCatches the inflated team89%
  • passRescues Gita from one bad launch100%
  • partialNotices who's underrated54%
  • passClear and direct, with care96%
  • passJudges outcomes, not activity95%
  • passNames gaps you could see77%
  • passWeighs the whole period72%
Run
Run
#1
API response time
47 s
Submitted
6 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.

Where they differ 3

Catches the inflated teamRightRightMixed
GPT-6 Luna · API

It names Aisha, Chloe, and Emma with their missed outcome goals and proposes lowering them to Meets, while keeping Ben and Dev at Exceeds.

Haiku 5.5 · API

It names Aisha, Chloe and Emma, cites their outcome misses, proposes lowering them to Meets, and keeps Ben and Dev at Exceeds.

Sonnet 5.5 · API

The output lowers Ben and Dev to Meets instead of keeping them at Exceeds, failing to keep the two who hit both goals.

Notices who's underratedWrongRightWrong
GPT-6 Luna · API

It does not propose Exceeds or ask Sara to justify Meets despite Nora beating both goals; it merely keeps Meets, missing the underrating.

Haiku 5.5 · API

It raises Nora from Meets to Exceeds because she beat both outcome goals, using her payment success and ticket reduction results.

Sonnet 5.5 · API

It keeps Nora at Meets without proposing Exceeds or asking Sara to justify Meets, despite her beating both goals.

Names gaps you could seeMixedRightMixed
GPT-6 Luna · API

It does not name specific, observable behaviors or what good looks like for any PM; the calibration pre-read focuses on ratings and evidence, not developmental gaps.

Haiku 5.5 · API

Gaps are described as specific results (conversion 8.4% vs 11%, API +4% vs +20%, app late and actives +3%, launch rollback) and what good would look like is implied by the goals.

Sonnet 5.5 · API

The gaps for Aisha, Chloe, and Emma are described as missing outcome evidence, not as specific observable behaviors with what good would look like.

All got right 10

Uses the supplied evidence correctlyRightRightRight
GPT-6 Luna · API

Every factual claim about the current situation is taken directly from the proposed_ratings.csv or calibration_guidance.md, with no inventions.

Haiku 5.5 · API

All claimed numbers and facts appear in the supplied CSV or are derived by straightforward arithmetic; nothing about the current situation is invented.

Sonnet 5.5 · API

All claims about the current situation are taken directly from the supplied context or follow by arithmetic.

Addresses the actual decisionRightRightRight
GPT-6 Luna · API

The output commits to specific rating changes and keeps, frames them for the calibration discussion, and states that additional evidence from managers could change the recommendations.

Haiku 5.5 · API

The output clearly proposes specific rating changes, keeps, and holds, with conditions (missing documentation) that would alter the calls.

Sonnet 5.5 · API

The output commits to changing 7 ratings and keeping 7, with clear proposals and conditions that would change the call.

Respects explicit constraintsRightRightRight
GPT-6 Luna · API

The pre-read is within 1,100 words, includes the requested sections (changes, keeps, note to Ravi), and follows the calibration guidance.

Haiku 5.5 · API

The output is a pre-read within 1,100 words, addresses the required elements (changes, keeps, note to Ravi), and follows the format.

Sonnet 5.5 · API

The pre-read is within 1,100 words, addresses the managers, and includes the required sections.

Identifies material uncertaintyRightRightRight
GPT-6 Luna · API

It identifies the missing documented feedback for Gita and Hugo as key unknowns and asks for any such documentation before finalizing.

Haiku 5.5 · API

It identifies the unknowns (documentation for Gita and Hugo, adoption data for Ines) and says how they would resolve the ratings.

Sonnet 5.5 · API

It names specific unknowns (outcome evidence for Aisha/Chloe/Emma, goal difficulty for Ben, feedback history for Gita) and how they would be resolved.

Avoids unsupported claimsRightRightRight
GPT-6 Luna · API

No interpretations, causes, or forecasts are presented as established fact; all confident claims are directly supported by the supplied evidence.

Haiku 5.5 · API

Interpretations (e.g., '40% of the goal') are clearly based on arithmetic; no causes or forecasts are presented as established fact without support.

Sonnet 5.5 · API

Interpretations are clearly the author's judgment, and no unsupported facts are presented as established.

Produces the required deliverableRightRightRight
GPT-6 Luna · API

The pre-read is complete, in the right form, within length, and usable for the calibration discussion with at most light edits.

Haiku 5.5 · API

The pre-read is complete with changes, keeps, a dedicated note to Ravi, and could be used by a PM with only minor edits.

Sonnet 5.5 · API

The pre-read is complete, in the right form, and usable as a calibration document with light edits.

Rescues Gita from one bad launchRightRightRight
GPT-6 Luna · API

It raises Gita to Meets, noting she beat her main goal and the Below rests on one launch without the required documented feedback, and asks for that documentation.

Haiku 5.5 · API

It raises Gita to Meets citing her beat on the main goal and the lack of documented feedback, and asks for the launch gap to be documented.

Sonnet 5.5 · API

It raises Gita to Meets, noting she beat her main goal and the Below lacks the required documented earlier feedback.

Clear and direct, with careRightRightRight
GPT-6 Luna · API

The pre-read states changes directly and clearly, and the note to Ravi is respectful, specific, and actionable.

Haiku 5.5 · API

The note to Ravi is direct, names the three PMs and the numbers, asks for specific outcome analyses and feedback documentation, and offers support.

Sonnet 5.5 · API

The note to Ravi is direct, respectful, and states exactly what needs to change and what support is available.

Judges outcomes, not activityRightRightRight
GPT-6 Luna · API

It judges every PM on the outcomes achieved against their goals, treating shipping and activity as context, not the result.

Haiku 5.5 · API

Every rating change is argued from outcome attainment against goals, with outputs and activity treated as context not the result.

Sonnet 5.5 · API

The ratings are judged against goal outcomes, with activity treated as context, not the result.

Weighs the whole periodRightRightRight
GPT-6 Luna · API

It weighs the full review period by using the goal results, and explicitly guards against one recent event dominating (e.g., Gita's rollback).

Haiku 5.5 · API

The review uses the full-year goal results for every PM, and Gita's launch failure does not dominate her rating because the other goal was met and documentation is absent.

Sonnet 5.5 · API

The whole year's goals are weighed, and no single recent event dominates the ratings.

Results

Every setup we’ve tested on this task type, across all its tasks and repeats, graded on the current checklist. Calibrated.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6 AstraChatGPT93.892.32None
2Haiku 5.5API93.888.52None
3GPT-6.1 SolAPI97.984.62None
4Opus 5.5Claude91.885.22None
5GPT-6 LunaAPI94.177.72None
6Gemini 3.8 FlashAPI89.776.92None
7Sonnet 5.5API90.174.22None
8Gemini 3.5 Flash-LiteGemini74.449.52None

What you get for the cost

Haiku 5.5 · API matches GPT-6 Astra · ChatGPT at about a fortieth of the cost.

Task score against relative cost for each setup. The table below has the same figures.50607080901001×3×10×30×Cost per output, relative to the cheapest setup (log scale) →↑ Task scoreGPT-6 Astra · ChatGPT: task score 91, 70× the cheapestHaiku 5.5 · API: task score 91, 1.7× the cheapestGPT-6.1 Sol · API: task score 89, 13× the cheapestOpus 5.5 · Claude: task score 87, 61× the cheapestGPT-6 Luna · API: task score 82, the cheapestGemini 3.8 Flash · API: task score 82, 12× the cheapestSonnet 5.5 · API: task score 81, 31× the cheapestGemini 3.5 Flash-Lite · Gemini: task score 61, 1.9× the cheapest (1 of 2 outputs estimated)GPT-6 Astra · ChatGPTHaiku 5.5 · APIGPT-6.1 Sol · APIOpus 5.5 · ClaudeGPT-6 Luna · APIGemini 3.8 Flash · APISonnet 5.5 · APIGemini 3.5 Flash-Lite · Gemini
Best valueNobody beats these on both score and costCost partly estimated (app runs with no usage meter)

Cost is each output’s tokens at the model’s list price, shown relative to the cheapest setup. Runs done in an app have no usage meter, so theirs are estimated from the same setup’s measured runs. How we compare cost

The PM job

Reviewing the performance of a PM you manage.

Why it matters

Reviews fail by being kind and vague, or by judging a whole year on its worst month. Either way, the person doesn't know what to change.

What good looks like

  • Clear and direct, with care
  • Judges outcomes, not activity
  • Names gaps in observable terms
  • Weighs the whole period, not the last month

Deliberately not measured

  • HR policy compliance beyond what the brief supplies
Capability tested

Feedback and judgement

The failure we’re looking for

Vague, diplomatic feedback, or a verdict driven by one recent event

Grading

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