Humanity’s Last PRD
Latest drop: Gemini 3.8 Flash · 1 Oct
Tasks / Operate

Make the launch call

Can the model make a clear go / no-go call from mixed launch evidence, checked against the criteria agreed up front?

Measures the modelTask v1.2 · 2 casesDifficulty

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

From 14 graded outputs by 7 models. 64% were usable with at most a quick edit.

Reliably right

  1. Finds the Scotland breach100% pass
    Clearly identifies Scotland's refund breach (66% above control, 143% after supplier change) and acceptance dip below 80%, correctly holds Scotland.
    Opus 5.5 · Claude · Go/no-go for AI substitutions, from the rollout data
  2. Produces the required deliverable98% pass
    The recommendation is in the requested form, for the meeting, within the length, and includes necessary steps a PM could act on.
    GPT-6 Astra · ChatGPT · Go/no-go for AI-drafted support replies
  3. Addresses the actual decision95% pass
    The output gives a clear, unambiguous 'no-go' call early, and states what would change it (fixes and retest).
    GPT-6 Astra · ChatGPT · Go/no-go for AI-drafted support replies

Where it slips

  1. Limits the downside of being wrong32% pass
    No specific post-launch monitoring signal or threshold is named; only generic 'monitoring' and a kill switch are mentioned.
    GPT-6 Astra · ChatGPT · Go/no-go for AI-drafted support replies
  2. Catches the duplicate rows50% pass
    Finds and removes the duplicate rows but never states whether they change the results.
    GPT-6.1 Sol · API · Go/no-go for AI substitutions, from the rollout data
  3. Uses the supplied evidence correctly54% pass
    The output states that pilot agents were hands-on, engaged and motivated, which is not supported by the supplied context (only that 7/8 wanted to keep it).
    Opus 5.5 · Claude · Go/no-go for AI-drafted support replies

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 the Staff PM for fulfilment at Basketful. On 30 October the launch meeting decides whether AI auto-substitutions (the model picks a replacement when an item is out of stock, instead of the picker) roll out to every region on Monday 2 November, ahead of the Christmas delivery-slot booking window that opens on 9 November. The rollout workbook is attached as two files: daily_metrics.csv (28 days of the regional test, by region and arm) and incidents.csv. Work from the data, not the summaries people have given you. Write your recommendation for the meeting: go, no-go or go with conditions, region by region, and why. Include a table that checks each launch criterion for each region with the figures. Keep the prose under 900 words.

What the model was given6 items: The test, Launch criteria (agreed in the PRD), What people have said, Engineering, daily_metrics.csv, incidents.csv
The test1–28 October, five regions. In each region orders were split between control (the picker chooses substitutes, as today) and treatment (AI auto-substitutions). The split was not 50/50 everywhere: treatment got 30% of orders in London, 50% in the South East and North West, and 70% in the Midlands and Scotland.
Launch criteria (agreed in the PRD)1. Substitution acceptance (subs_accepted ÷ subs_offered) of at least 80% in treatment, in every region. 2. Refund requests per 1,000 orders in treatment no more than 10% above control, in every region. 3. Zero dietary or allergen mismatches (a substitute that breaks a gluten-free, vegan, vegetarian, nut-free, halal or kosher attribute). 4. Average basket value not lower in treatment than control.
What people have saidHead of Commercial: “Acceptance is up 15 points and refunds are within the guardrail overall. Every day we wait costs us Christmas.” Head of Analytics: “Treatment baskets are £4 smaller across the test. That worries me.” Ops director (Scotland): “The supplier change caused some noise, but it's settling down.”
EngineeringThe dietary-attribute hard filter (never substitute across those attributes) is merged and scheduled to deploy on 4 November after regression tests. Scotland's catalogue re-map for the new dairy supplier is in progress with no date. Rollout can be switched on region by region.
daily_metrics.csv283 rows · Download
date,region,arm,orders,items_ordered,items_out_of_stock,subs_offered,subs_accepted,refund_requests,refund_value_gbp,complaints,avg_basket_gbp
2026-10-01,London,control,1757,52273,2351,1953,1387,32,160.71,6,72.89
2026-10-01,London,treatment,692,20455,871,820,708,12,62.05,2,72.42
2026-10-01,South East,control,853,22846,996,833,593,16,74.49,3,65.62
2026-10-01,South East,treatment,928,25223,1166,1106,952,15,75.93,3,66.59
2026-10-01,Midlands,control,467,10468,465,383,268,8,34.57,1,54.92
2026-10-01,Midlands,treatment,1113,25136,1166,1087,947,21,106.35,3,55.33
…
incidents.csv13 rows · Download
date,region,arm,type,severity,summary,status
2026-10-02,London,treatment,wrong_size,low,"4-pint milk substituted with 1-pint, customer accepted at door",closed
2026-10-04,Midlands,treatment,dietary_mismatch,high,"Gluten-free sliced loaf substituted with standard white loaf; customer coeliac, noticed at home",closed: refund and apology
2026-10-06,South East,control,wrong_item,low,"Picker substituted oat milk with soya milk; customer rejected at door",closed
2026-10-07,North West,treatment,data_export,low,"Daily export job re-ran after a timeout on 8, 9 and 10 Oct; analytics team flagged possible duplicate rows",open
2026-10-09,London,treatment,price,medium,"Substitute priced higher than original; customer charged the difference against policy",closed: policy fix deployed 11 Oct
2026-10-12,Scotland,control,late_delivery,low,"Van breakdown, 14 orders late",closed
…
What a strong answer doesThe answer key the graders mark against

No-go for a national launch on 2 November, because criterion 3 fails everywhere until the dietary hard filter ships: incidents.csv has three treatment dietary mismatches (4 Oct Midlands, 19 Oct South East, 23 Oct London, the last involving nuts and a child with an allergy). Go with conditions, region by region, once the filter is deployed and verified (planned 4 November, still before the 9 November booking window): London, South East, Midlands and North West pass the other criteria (acceptance about 86% against 71% control; refunds 2–5% below control). Hold Scotland: over the month its treatment refunds are about 66% above control, and from the 15 October supplier change about 143% above (43 against 18 per 1,000 orders), with acceptance falling to about 79%, below the 80% bar; it waits for the catalogue re-map and a clean re-test. The pooled refund rate (about 8% above control) hides this, and the criterion is per region. The pooled basket gap (about £65 against £61) is an artefact of the uneven split, since treatment baskets are equal or higher in every region (Simpson's paradox), so criterion 4 passes. Notices that North West treatment rows for 8–10 October are duplicated (an extra 2,194 orders), removes them, and says it doesn't change the result. States what would change the call: the filter slipping past 9 November, or Scotland's refunds not recovering.

Critical failures (cap the score at 40)
  • Recommends any rollout before the dietary hard filter is deployed
  • Recommends rolling out to Scotland while its refund guardrail is breached
Case

Staff-level · v1.8 · synthetic · e-commerce, AI product, data analysis, staff-level

The outputs

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

Opus 5.5Claude91Usable with a quick editGPT-6.1 SolAPI86Usable with a quick editSonnet 5.5API84Usable with a quick editGPT-6 AstraChatGPT82Usable with a quick editGPT-6 LunaAPI64Needs more workGemini 3.5 Flash-LiteGemini50Needs more workGemini 3.8 FlashAPI40Critical failure
Opus 5.5Claude

Usable with a quick edit11 of 13 checks passed

The output is evidence-based, fully meets the brief, and can be used directly by a product manager without re-checking data or redoing analysis.

Output

AI auto-substitutions: launch recommendation

Oct 29, 2026 · @Martin Slaney · For the 30 October launch meeting

Recommendation

No-go for Monday 2 November in any region. The dietary hard filter isn't live until 4 November, and the test produced three allergen and dietary mismatches.

Go with conditions for London, South East, Midlands and North West from Thursday 5 November, once the filter is live and verified. That is still four days before the Christmas booking window opens on 9 November.

No-go for Scotland until the dairy catalogue re-map ships and a fresh week of test data passes the criteria.

Launch criteria by region

Figures are for 1–28 October from daily\_metrics.csv, with duplicate rows removed. T = treatment, C = control.

Region1. Acceptance (T), ≥80%2. Refunds per 1k orders, T vs C, ≤+10%3. Dietary mismatches (T), 04. Avg basket, T vs CVerdict
London86.7% ✅17.4 vs 18.0 (−3%) ✅1: nut-free to hazelnut, 23 Oct, still open ❌£72.70 vs £72.49 ✅Go 5 Nov, after filter
South East86.4% ✅17.5 vs 17.9 (−2%) ✅1: vegan to dairy, 19 Oct ❌£66.64 vs £66.20 ✅Go 5 Nov, after filter
Midlands86.4% ✅17.5 vs 17.9 (−2%) ✅1: gluten-free to standard, 4 Oct ❌£55.66 vs £55.35 ✅Go 5 Nov, after filter
North West86.3% ✅17.6 vs 18.5 (−5%) ✅0 ✅ (same model, so not evidence of safety)£57.51 vs £57.29 ✅Go 5 Nov, after filter
Scotland82.8% overall; 79.1% since 15 Oct ❌30.2 vs 18.2 (+66%); since 15 Oct 43.2 vs 17.8 (+143%) ❌0 ✅£52.98 vs £52.62 ✅No-go until re-map and re-test

Why not 2 November

Criterion 3 requires zero dietary or allergen mismatches. Treatment had three, all rated high severity: a gluten-free loaf swapped for a standard one (the customer is coeliac), vegan cheese swapped for dairy cheddar (the customer complained on social media), and a nut-free multipack swapped for one containing hazelnuts, sent to a child with a nut allergy. The nut incident is still open. Control had none.

The fix is the dietary hard filter, which is merged and deploys on 4 November after regression tests. Launching on 2 November would put every order in every region onto the unfiltered model for two days. North West's zero is luck, because it runs the same model.

Strictly, all four "go" regions fail criterion 3 as it's written. I'm asking the meeting to accept the filter as the remedy, on condition that we show it would have blocked all three incidents. If the meeting won't accept that, the honest answer is no-go everywhere.

Waiting three days costs nothing for Christmas, because bookings don't open until 9 November.

What the summaries get wrong

Head of Commercial: "refunds within the guardrail overall." Pooled, treatment refunds are 7.9% above control (19.5 vs 18.1 per 1,000 orders), which is inside 10%. But the criterion applies to every region, and the pooled figure hides Scotland at +66%. In the other four regions, treatment refunds are below control. The 15-point acceptance lift is real in those four regions (+15.0 to +15.4 points). In Scotland it's +11.7 overall and falling.

Head of Analytics: "treatment baskets are £4 smaller." Pooled, the gap is £61.09 vs £65.15, but it comes from how orders were split, not from the feature. Treatment got 70% of orders in Midlands and Scotland, which have the smallest baskets (£52–56), and only 30% in London, which has the largest (£72). Within every region, treatment baskets are £0.21–£0.44 larger, so criterion 4 passes everywhere.

Ops director, Scotland: "it's settling down." It isn't. Before the supplier switch on 15 October, Scotland's treatment arm matched the other regions (86.3% acceptance, 17.3 refunds per 1,000 vs 18.6 in control). Since then, treatment refunds have run at 25–33 a day with no downward trend: 28.1 a day on average from 15 to 21 October and 28.7 from 22 to 28 October. Refund value per order has tripled, from £0.09 to £0.30. Scotland's control arm is unaffected at 17.8 per 1,000. So the problem is specific to the model reading a mis-mapped catalogue: pickers can see the real pack sizes, and the model can't. The re-map has no date.

Data quality

The North West treatment rows for 8, 9 and 10 October each appear twice, because the export job re-ran (the open incident from 7 October). I removed the duplicates. They don't change any verdict: North West acceptance moves from 86.2% to 86.3%.

Conditions for the 5 November go

  1. The dietary filter deploys on 4 November and passes regression tests. We replay the three incident orders and confirm the filter blocks each of them.2
  2. The London nut-allergen incident is closed with a root cause.
  3. We switch on region by region and keep a 10% control holdout in each region through Christmas, so criteria 2 and 4 stay measurable.
  4. Kill switch: any single dietary or allergen mismatch turns the region off, as does a treatment refund rate more than 10% above its holdout over a rolling 3 days.1
  5. Scotland: ship the catalogue re-map, then run 7 days of test and re-check the criteria. If it can't be ready by 9 November, Scotland takes Christmas orders with picker substitutions.

What a PM had to fix

  1. 1Test or gate too weakTighten the testQuick edit

    What we’d changeSet the refund trigger outside day-to-day noise, with a bigger gap or a longer window, or it will switch regions off for no reason.

  2. 2Test or gate too weakTighten the testQuick edit

    What we’d changeThree replays is thin evidence the filter works. Test it across every protected attribute, including missing or conflicting ones.

Grades and run details

Decision model 81 · LLM judge 14 of 14 checks
Decision model checks
  • failUses the supplied evidence correctly23%
  • passAddresses the actual decision82%
  • passRespects explicit constraints27%
  • passIdentifies material uncertainty97%
  • partialAvoids unsupported claims38%
  • passProduces the required deliverable71%
  • passChecks every criterion, region by region53%
  • passFinds the Scotland breach100%
  • passExplains the basket gap98%
  • passCatches the duplicate rows100%
  • passMakes a clear call96%
  • passLimits the downside of being wrong86%
  • failGets the base of every number right40%
Run
Run
#1
Time to output
2.6 min
Submitted
25 Sept 2026
Gemini 3.5 Flash-LiteGemini

Needs more than a quick edit6 of 13 checks passed

It fails to address the known duplicate data rows, doesn't specify monitoring triggers or rollback criteria, and makes a causal claim about Scotland without hedging, so a PM would need to add those before using it.

Output

Launch Recommendation: AI Auto-Substitutions

As the Staff PM for fulfilment, my recommendation for the 30 October launch meeting is a No-Go for nationwide rollout on 2 November.

While the Commercial and Analytics teams are looking at top-line aggregates, a region-by-region audit of the 28-day test data and incident logs reveals critical blockers. Specifically, Scotland fails acceptance and refund criteria due to a supplier mapping failure, Midlands and London recorded high-severity dietary/allergen mismatches, and South East also experienced a dietary mismatch.

We should hold all regions until the dietary hard filter is deployed on 4 November, and until Scotland’s catalogue re-map is fully completed and verified.

---

Regional Evaluation Table (1–28 October Test Data)

RegionSubs Acceptance ($\ge 80\%$)Treatment Refunds vs Control ($\le +10\%$)Dietary/Allergen Mismatches (Target: 0)Basket Value ImpactRecommendation
:---:---:---:---:---:---
London$86.4\%$ (Pass)Control: 1.77%<br>Treatment: 1.63% ($-7.9\%$)1 High Incident (Nut allergy violation on 23 Oct)Control: £72.33<br>Treatment: £72.71 (+0.5%)No-Go (Hold for 4 Nov filter)
South East$87.0\%$ (Pass)Control: 1.66%<br>Treatment: 1.69% ($+1.8\%$)1 High Incident (Vegan dairy violation on 19 Oct)Control: £66.21<br>Treatment: £66.72 (+0.8%)No-Go (Hold for 4 Nov filter)
Midlands$84.5\%$ (Pass)Control: 1.69%<br>Treatment: 1.68% ($-0.6\%$)1 High Incident (Gluten-free coeliac violation on 4 Oct)Control: £55.22<br>Treatment: £55.51 (+0.5%)No-Go (Hold for 4 Nov filter)
North West$88.4\%$ (Pass)Control: 1.73%<br>Treatment: 1.66% ($-4.0\%$)0 IncidentsControl: £57.34<br>Treatment: £57.65 (+0.5%)Go with Conditions (Deploy 4 Nov)
Scotland$77.0\%$ (Fail)Control: 1.79%<br>Treatment: $3.90\%$ ($+117.9\%$)0 Incidents (High dairy wrong-size volume)Control: £52.48<br>Treatment: £52.88 (+0.8%)No-Go (Hold for catalogue re-map)

Note: Refund request rates are calculated per 100 orders for precise comparison against the 10% relative ceiling.

---

Region-by-Region Breakdown

  • London, South East, Midlands: All three regions easily clear the 80% acceptance threshold and show neutral-to-positive average basket values (debunking the £4 basket drop concern). However, all three recorded high-severity dietary/allergen failures (gluten, vegan, and nut-allergy breaches). Because the safety hard filter is not deployed until 4 November, rolling out on 2 November exposes customers to unacceptable allergen risks during the peak Christmas slot-booking window.
  • North West: Our strongest performer. It meets all four PRD criteria cleanly3: 88.4% acceptance1, treatment refund rates slightly below control, zero dietary incidents, and a stable basket value. It is ready to roll out immediately following the hard filter deployment.
  • Scotland: A clear No-Go. Acceptance dropped to 77.0% (failing the 80% bar), and treatment refund requests spiked catastrophically by 117.9% over control (driven by the dairy supplier switch on 15 Oct where pack sizes were mismapped). Until the catalogue re-map is finished and stabilized, rolling out here will compound operational friction.

Action Plan

  1. Hold Rollout on 2 November.
  2. Deploy Dietary Hard Filter on 4 November as scheduled, and verify regression tests.
  3. Complete Scotland Catalogue Re-map and clear open wrong-size substitution incidents.
  4. Staged Launch: Roll out to North West, London, South East, and Midlands on Monday, 9 November (concurrent with the Christmas slot-booking window opening), pending a 7-day post-fix stability review.2 Scotland follows only after catalogue remediation.

What a PM had to fix

  1. 1Numbers wrongRedo the arithmeticStart again

    What we’d changeRecompute every figure from daily_metrics.csv, removing the duplicated North West rows. Almost every number in the table is wrong.

  2. 2Contradiction missedSurface the contradictionSubstantial rework

    What we’d changeResolve the contradiction: a 7-day review after the 4 November filter can't finish by 9 November, and North West is marked go in the table but held here. Answer the stakeholders' claims rather than dismissing them.

  3. 3Hypothesis stated as factReframe it as a hypothesisTargeted repair

    What we’d changeZero recorded mismatches in North West isn't evidence of safety. It runs the same unfiltered model as the regions that had them.

Grades and run details

Decision model 46 · LLM judge 8 of 14 checks
Decision model checks
  • failUses the supplied evidence correctly49%
  • passAddresses the actual decision32%
  • passRespects explicit constraints8%
  • partialIdentifies material uncertainty63%
  • partialAvoids unsupported claims25%
  • passProduces the required deliverable26%
  • partialChecks every criterion, region by region64%
  • passFinds the Scotland breach96%
  • failExplains the basket gap48%
  • failCatches the duplicate rows100%
  • partialMakes a clear call83%
  • failLimits the downside of being wrong41%
  • failGets the base of every number right77%
Run
Run
#1
Time to output
46 s
Submitted
25 Sept 2026
GPT-6 LunaAPI

Needs more than a quick edit9 of 13 checks passed

The refund figures are inaccurate, so a PM would need to re-check the data and recalculate before using the recommendation. The output is not usable as-is.

Output

Recommendation: No-go for the all-region rollout on 2 November

Do not enable auto-substitutions anywhere on 2 November: the dietary hard filter is not scheduled to deploy until 4 November, so the zero-mismatch safety criterion is not yet protected. After deployment and validation, North West is the only region I’d consider enabling first, with conditions. Hold the other regions for the reasons below.

Figures are for 1–28 October; refund rates are per 1,000 orders. Basket averages are order-weighted within each region. I excluded the three exact duplicate North West treatment rows on 8, 9 and 10 October.

RegionTreatment acceptance (≥80%)Refunds / 1,000 orders: treatment vs control (treatment ≤10% higher)Dietary/allergen mismatches (must be zero)Average basket: treatment vs control (must not be lower)Decision
------:---:------:---
London~86% Pass~17.9 vs ~18.4; Pass1; Fail — nut-containing bar substituted for nut-free item~£72.8 vs ~£72.5; PassNo-go
South East~85% Pass~17.8 vs ~17.9; Pass1; Fail — dairy cheese substituted for vegan cheese~£66.6 vs ~£66.3; PassNo-go
Midlands~86% Pass~17.6 vs ~18.0; Pass1; Fail — standard loaf substituted for gluten-free loaf~£55.6 vs ~£55.4; PassNo-go
North West~86% Pass~17.8 vs ~18.8; Pass0 observed; Pass in test~£57.5 vs ~£57.3; PassGo with conditions, after filter validation
Scotland~79% Fail1~29.8 vs ~18.4; Fail — about 62% higher0 observed; Pass in test~£53.0 vs ~£52.6; PassNo-go

Conditions and rationale

  • North West: After the hard filter deploys, pass regression tests across all six protected attributes (gluten-free, vegan, vegetarian, nut-free, halal and kosher); then enable regionally with close monitoring and a rollback switch. The export duplicates should also be fixed or guarded against before relying on daily reporting.
  • London, South East and Midlands: Do not launch until the filter is deployed and the dietary-mismatch failures are investigated and validated as fixed. Their good acceptance, refund and basket figures do not offset a safety-criterion failure.
  • Scotland: Hold until the supplier catalogue re-map is complete and verified, and a fresh test meets both acceptance and refund guardrails. The sharp rise in treatment refunds and lower acceptance coincide with the dairy supplier/catalogue change; the issue is open, with no completion date.

The pooled “£4 smaller” basket claim should not drive this decision: treatment exposure differed by region, and baskets vary substantially by region. Within each region, the treatment average is slightly higher in these data. Likewise, an overall acceptance or refund result cannot override region-level launch gates. The 9 November booking window is important, but it does not justify launching before the safety control is live.

What a PM had to fix

  1. 1Invented evidenceVerify or remove the claimTargeted repair

    What we’d change82.76%: Pass over 1–28 October. However, acceptance fell to 79.08% over 15–28 October after the supplier change, below the launch threshold.

Claims the judge couldn’t find in the brief

  • London treatment refunds per 1,000 orders is ~17.9.
  • London control refunds per 1,000 orders is ~18.4.
  • South East treatment refunds per 1,000 orders is ~17.8.
  • Midlands treatment refunds per 1,000 orders is ~17.6.
  • North West treatment refunds per 1,000 orders is ~17.8 (after excluding duplicates).
  • Scotland control refunds per 1,000 orders is ~18.4.

Grades and run details

Decision model 69 · LLM judge 9 of 14 checks
Decision model checks
  • failUses the supplied evidence correctly8%
  • passAddresses the actual decision47%
  • passRespects explicit constraints61%
  • partialIdentifies material uncertainty55%
  • passAvoids unsupported claims21%
  • passProduces the required deliverable63%
  • passChecks every criterion, region by region28%
  • passFinds the Scotland breach88%
  • passExplains the basket gap86%
  • partialCatches the duplicate rows69%
  • partialMakes a clear call62%
  • partialLimits the downside of being wrong96%
  • failGets the base of every number right45%
Run
Run
#1
API response time
44 s
Submitted
29 Sept 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 8

Uses the supplied evidence correctlyMixedWrongWrong
Opus 5.5 · Claude

Every factual claim about the current situation is drawn directly from the supplied files or computed correctly from them.

Gemini 3.5 Flash-Lite · Gemini

The output calculates rates from the raw daily_metrics.csv without acknowledging or removing the duplicate North West treatment rows flagged in incidents.csv, so the claimed rates for North West (and potentially others if the duplication affects summaries) are not fully reliable.

GPT-6 Luna · API

Several refund-per-1000 figures are incorrect (e.g., London treatment 17.9 vs actual 17.4, London control 18.4 vs 18.0), so the output does not use the supplied evidence correctly.

Identifies material uncertaintyRightWrongRight
Opus 5.5 · Claude

Names the unknowns (filter deployment date, Scotland re-map readiness, acceptance of filter as remedy) and says each would change the decision.

Gemini 3.5 Flash-Lite · Gemini

Does not explicitly name the unknowns that could change the call (e.g., what if the dietary filter slips past 9 November, or if Scotland’s refunds don’t recover after the catalogue fix) or state the thresholds that would reverse the decision.

GPT-6 Luna · API

Names the filter deployment, duplicate fix, and Scotland catalogue re-map as unknowns that could change the decision.

Avoids unsupported claimsRightWrongRight
Opus 5.5 · Claude

Hypotheses (e.g. 'luck' for NW zero) are clearly labelled as interpretation; otherwise claims are grounded in the data.

Gemini 3.5 Flash-Lite · Gemini

States as fact that Scotland’s refund spike was 'driven by the dairy supplier switch where pack sizes were mismapped' without labelling it as a hypothesis, even though the data only shows a correlation.

GPT-6 Luna · API

Causal claims like the supplier change coinciding with refunds are presented as observations, not unsupported facts.

Checks every criterion, region by regionRightRightMixed
Opus 5.5 · Claude

All four launch criteria are checked for each region with correct computed figures from the files.

Gemini 3.5 Flash-Lite · Gemini

The table checks acceptance, refunds vs control, dietary incidents and basket value for each of the five regions, with figures drawn from the test data.

GPT-6 Luna · API

The table checks each criterion per region but the refund figures are incorrect, so it does not present correct figures from the files.

Explains the basket gapRightMixedRight
Opus 5.5 · Claude

Explicitly explains the pooled basket gap as a traffic-mix artefact from the uneven regional split, with within-region comparisons showing treatment baskets are higher.

Gemini 3.5 Flash-Lite · Gemini

Shows treatment basket values are equal or higher in every region and states the aggregated £4 drop concern is debunked, satisfying the Simpson’s paradox explanation.

GPT-6 Luna · API

Explains the pooled basket gap as a mix effect, noting treatment baskets are equal or higher within each region.

Catches the duplicate rowsRightWrongRight
Opus 5.5 · Claude

Notices the duplicated North West treatment rows for 8–10 October, removes them, and confirms the verdict doesn't change.

Gemini 3.5 Flash-Lite · Gemini

Does not mention or handle the duplicate North West treatment rows on 8–10 October that the incident log flags, so the quality check is absent.

GPT-6 Luna · API

Catches the duplicate North West rows, excludes them, and recommends fixing the export job.

Limits the downside of being wrongRightWrongWrong
Opus 5.5 · Claude

Specifies post-launch kill switch (any dietary mismatch or refund rate >10% above holdout over rolling 3 days), a 10% control holdout, and conditions for Scotland re-test.

Gemini 3.5 Flash-Lite · Gemini

While it mentions a 7‑day post‑fix stability review, it does not specify a signal (e.g., refund rate threshold) or a rollback trigger, and does not flag irreversible actions.

GPT-6 Luna · API

Does not name a specific post-launch signal, threshold, or trigger for rollback; only says 'close monitoring and a rollback switch' without specifics.

Gets the base of every number rightMixedMixedWrong
Opus 5.5 · Claude

All derived figures (acceptance, refunds per 1k, basket averages) are computed correctly from the supplied data with appropriate denominators.

Gemini 3.5 Flash-Lite · Gemini

Rates are computed as sums over the period (acceptance = total accepted / total offered, refund rate = total refunds / total orders), so denominators are consistent and the relative comparisons are valid.

GPT-6 Luna · API

Multiple refund-per-1000 figures are miscalculated (e.g., London treatment 17.9 vs 17.4, London control 18.4 vs 18.0), so the base numbers are wrong.

All got right 5

Addresses the actual decisionRightRightRight
Opus 5.5 · Claude

Commits unequivocally to a no-go for 2 November, go with conditions from 5 November for four regions, and no-go for Scotland, and states what would change each call.

Gemini 3.5 Flash-Lite · Gemini

Commits to a clear No‑Go for 2 November, then a conditional Go by region for 9 November after the filter deploys, and states Scotland’s hold conditions.

GPT-6 Luna · API

Commits to a clear no-go for all-region on 2 Nov, then conditional go for North West after filter, with conditions and what would change it.

Respects explicit constraintsRightRightRight
Opus 5.5 · Claude

Respects the 900-word prose limit, includes the required region-by-criterion table, and is addressed to the launch meeting as requested.

Gemini 3.5 Flash-Lite · Gemini

Prose is well under 900 words, addressed to the launch meeting, and includes a region‑by‑region table with the four criteria.

GPT-6 Luna · API

Includes the required region-by-region table, prose under 900 words, and addresses the launch meeting.

Produces the required deliverableRightRightRight
Opus 5.5 · Claude

The output is a complete, actionable recommendation with a criterion-check table, suitable for the meeting with minimal editing.

Gemini 3.5 Flash-Lite · Gemini

Provides a complete, table‑based recommendation with a region‑by‑region call, within the word limit, that the launch meeting could act on.

GPT-6 Luna · API

Provides a complete recommendation with table, region-by-region, usable for the meeting.

Finds the Scotland breachRightRightRight
Opus 5.5 · Claude

Clearly identifies Scotland's refund breach (66% above control, 143% after supplier change) and acceptance dip below 80%, correctly holds Scotland.

Gemini 3.5 Flash-Lite · Gemini

Identifies Scotland’s refund breach (+117.9%), links it to the post‑15 October supplier change, notes the acceptance drop to 77%, and holds Scotland.

GPT-6 Luna · API

Identifies Scotland's refund breach, ties it to the supplier change, notes acceptance dip, and holds Scotland.

Makes a clear callRightRightRight
Opus 5.5 · Claude

Gives a clear decision per region with dates (no-go 2 Nov, go with conditions 5 Nov for four regions, no-go for Scotland) and states what would change it.

Gemini 3.5 Flash-Lite · Gemini

Gives a clear decision per region (No‑Go for Scotland, conditional Go for others from 9 Nov) with dates and the conditions required to proceed.

GPT-6 Luna · API

Makes a clear call per region with dates and conditions, and says what would change it.

Results

Every setup we’ve tested on this task, across all cases and repeats, graded on the current checklist. Calibrated: the graders match our PM on 87% of checks.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6.1 SolAPI87.388.32None
2GPT-6 AstraChatGPT86.088.32None
3Sonnet 5.5API85.488.32None
4Opus 5.5Claude80.486.42None
5GPT-6 LunaAPI82.177.62None
6Gemini 3.8 FlashAPI71.931.521 capped
7Gemini 3.5 Flash-LiteGemini50.642.22None

The PM job

Deciding whether a feature ships on the planned date, and on what conditions.

Why it matters

Launch meetings reward optimism. A good call checks each agreed criterion, weighs a real risk against a real gain, and says exactly what would change the answer. A weak one rubber-stamps the launch or blocks it on noise.

What good looks like

  • Checks each agreed criterion against the evidence
  • Makes one clear call, with conditions if needed
  • Separates risks that block launch from risks that can be managed
  • Says what would change the call
  • Says how to limit the damage if the call is wrong

Deliberately not measured

  • Rollout engineering detail
  • Project-plan formatting
Capability tested

Deciding against agreed launch criteria

The failure we’re looking for

Rubber-stamps a launch that misses an agreed bar, or blocks it on noise

Grading

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

Variants

A launch memo from supplied evidence · Staff level: a regional call from an attached data workbook