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

Extract discovery insights

Can the model separate evidence, themes and hypotheses without inventing consensus?

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. Frames findings for the decision100% pass
    Each main finding is framed as an opportunity (delivery-to-invoice pain, concentrated supplier need) and linked directly to what to build next quarter.
    GPT-6 Astra · ChatGPT · The loudest ticket
  2. Corrects the draft's claim100% pass
    It corrects the draft's headline by counting interviews (8 describe delivery/invoice pain, 2 mention supplier expansion) and showing the marketplace claim is unsupported.
    GPT-6 Astra · ChatGPT · The loudest ticket
  3. Keeps dissent visible100% pass

Where it slips

  1. Says how many sources support each finding52% pass
    The finding on smaller issues sizes Ordering as 31 tickets, but the actual count is 30, so the finding is not accurately sized.
    Sonnet 5.5 · API · The loudest ticket
  2. Uses the supplied evidence correctly59% pass
    The output states Ordering has 31 tickets, but the CSV contains 30 Ordering tickets; this factual claim is not supported by the supplied data.
    GPT-6.1 Sol · API · The loudest ticket
  3. Avoids unsupported claims63% pass
    The claim 'Ordering (31 tickets)' is presented as an established fact, but the evidence shows 30 Ordering tickets, so it is an unsupported confident claim.
    Sonnet 5.5 · API · The loudest ticket

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. Weights behaviour over opinion

    Does the synthesis distinguish evidence of what customers actually did (specific past instances, observed behaviour, workarounds) from stated preferences, predictions and feature requests, and weight them accordingly?

    Passes when Themes are labelled or weighted by evidence type, and at least one theme resting mainly on stated intent or requests is flagged as weaker.

    Bob MoestaLenny’s Podcast

  2. Frames findings for the decision

    Does the synthesis turn its findings into opportunities tied to the decision or outcome the team faces, saying what each one means for that decision?

    Passes when Each main finding is framed as an opportunity (a need or pain to address) and linked to the decision at hand, with what it implies.

    Teresa TorresLenny’s Podcast

  3. Reports what cuts against the plan

    Does the synthesis report the evidence that cuts against the team's current belief or plan, rather than only what supports it?

    Passes when Names at least one finding that challenges the team's hypothesis or plan, and says what it means for it.

    Judd AntinLenny’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 · Says how many sources support each finding · and 2 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 PM for Larder's ordering product. A colleague has drafted the synthesis below for our CPO, recommending a supplier marketplace. Before it goes, rewrite the synthesis from the evidence, in no more than 800 words: the main findings, how strong the evidence for each is, and what they mean for what we build next quarter. The interviews and the ticket export are attached.

What the model was given4 items: About Larder, draft_synthesis.md, interview_excerpts.md, support_tickets_last_quarter.csv
About LarderOrdering software for restaurants: order from your suppliers in one place, track deliveries and invoices. 1,140 accounts, most of them independent restaurants.
draft_synthesis.md6 lines · Download
# Draft synthesis (PM, for the CPO)

Headline: customers want a supplier marketplace.
- "Add more suppliers" is our top support category by far.
- Most customers we spoke to want a marketplace (see Calder Group and the pizzeria).
- Recommendation: build a supplier marketplace next quarter.
interview_excerpts.md12 lines · Download
# Interview excerpts (restaurant owners and managers, September)

1. Bistro, 2 sites: "Last Tuesday the fish order came short again. I only found out when the invoice came, a week later. I spent Sunday night matching delivery notes to invoices with a highlighter."
2. Café group, 5 sites: "We switched produce supplier twice this year, but finding suppliers isn't the problem. Knowing what actually arrived is."
3. Pizzeria, 1 site: "Honestly I'd love more suppliers in the app. More choice, better prices."
4. Gastropub: "My head chef signs the delivery note without checking. Then I pay for things we never got. Last month that was about £600."
5. Hotel restaurant: "Our finance team chases credit notes for short deliveries every week. It takes days."
…
support_tickets_last_quarter.csv200 rows · Download
ticket,account,account_sites,category,summary
LT-5000,Calder Group,14,Add more suppliers,Can you onboard our meat supplier
LT-5001,Account 1886,3,Ordering,Change delivery day
LT-5002,Account 2004,1,Delivery didn't match order,Short delivery not shown in app
LT-5003,Calder Group,14,Add more suppliers,Add dairy supplier please
LT-5004,Calder Group,14,Add more suppliers,Add dairy supplier please
LT-5005,Account 1858,2,Add more suppliers,Any plans for more suppliers?
…
What a strong answer doesThe answer key the graders mark against

Shows that the marketplace headline doesn't hold. 'Add more suppliers' is the biggest ticket category (75 of 200), but 58 of those come from one account, Calder Group; across other accounts, delivery and invoice problems ('delivery didn't match order' and 'invoice reconciliation', 63 tickets) are far more common than supplier requests (17). In the interviews, 7 of 10 describe a specific recent instance of the same pain (short deliveries, invoices that don't match, credit notes, re-typing paper checks), while the marketplace mentions are wishes (the pizzeria 'would love' it, the tapas bar 'I suppose') or come from Calder. Weights what people did over what they say they'd want, counts how many sources support each finding, and recommends focusing next quarter on checking deliveries and reconciling invoices, with the marketplace treated as one large customer's request to handle separately.

Critical failures (cap the score at 40)
  • Recommends building the supplier marketplace next quarter as the main finding
Task

Staff-level · v1.0

The outputs

8 setups answered this brief.

GPT-6 AstraChatGPT100Usable with a quick editGPT-6.1 SolAPI96Usable with a quick editHaiku 5.5API96Usable with a quick editGPT-6 LunaAPI88Needs more workSonnet 5.5API88Usable with a quick editGemini 3.8 FlashAPI87Usable with a quick editOpus 5.5Claude81Needs more workGemini 3.5 Flash-LiteGemini65Needs more work
GPT-6 LunaAPI

Needs more than a quick edit11 of 12 checks passed

The output fails to show that delivery and invoice issues are spread across many accounts, a key comparison a PM would need to add by re-checking the ticket data.

Output

Synthesis for the CPO

Headline

The evidence points more strongly to a delivery-reconciliation problem than to a general marketplace need. Supplier coverage matters—especially to Calder Group—but the interviews suggest the bigger day-to-day pain is knowing what arrived, resolving discrepancies, and matching invoices and credits. Build toward that workflow next quarter; treat a broad marketplace as an unproven solution.

Findings and evidence

1. Receiving and invoice reconciliation are recurring operational pains. — Moderate evidence Eight of the ten interview excerpts raise delivery checking, discrepancies, invoice matching, credit notes, or the paperwork around them. Examples include paying for undelivered goods, learning about short deliveries only when the invoice arrives, and re-entering clipboard checks into Larder. Several customers describe financial impact or repeated manual work. The support export also contains tickets in both “Delivery didn’t match order” and “Invoice reconciliation.”

This is a consistent signal across the interviews, but ten excerpts cannot establish how common the problem is across Larder’s 1,140 accounts or which part of the workflow is most valuable to solve first.

2. Supplier availability is a genuine need, but the ticket count overstates its breadth. — Moderate evidence for demand; weak evidence for a marketplace “Add more suppliers” is the largest ticket category: 75 of 200 tickets. But 58 of those tickets—most of the category—come from one customer, Calder Group. The remaining 17 are from other accounts, so supplier coverage is not solely a Calder concern, but ticket volume is not evidence that most customers want a marketplace. Calder’s 14 sites and stated intent to move all ordering to Larder make its request strategically important; it is still one account.

The interviews echo that distinction: the pizzeria wants more choice and better prices, and Calder explicitly wants every supplier on-platform. But another multi-site operator says finding suppliers is not the problem, and the tapas bar says paperwork is what hurts day to day.

3. A marketplace is a proposed solution, not a validated priority. — Weak evidence The evidence supports specific supplier gaps and some appetite for choice, but does not show that customers broadly want a marketplace, what marketplace model they mean, or whether it would resolve their more frequent receiving and reconciliation problems. The draft’s conclusion goes beyond what these sources can establish.

What to build next quarter

Prioritize a focused delivery-to-invoice reconciliation improvement: make it easier to record what was received, flag shortages and substitutions against the order, and carry discrepancies through to invoice matching and credit-note tracking. Test the workflow with customers who describe these pains before committing to a larger build.

In parallel, follow up with Calder and a few other customers requesting suppliers. Assess targeted supplier onboarding or integrations for the specific suppliers they need; do not assume this requires a general-purpose marketplace. Use the quarter to measure how widespread each problem is across distinct accounts and whether improved reconciliation reduces manual work and unresolved discrepancies.

Grades and run details

Decision model 100 · LLM judge 85
Decision model checks
  • passUses the supplied evidence correctly62%
  • passAddresses the actual decision61%
  • passRespects explicit constraints77%
  • passIdentifies material uncertainty45%
  • passAvoids unsupported claims78%
  • passProduces the required deliverable78%
  • passSees one account behind the top ticket67%
  • passCorrects the draft's claim53%
  • passWeights behaviour over opinion99%
  • passSays how many sources support each finding55%
  • passFrames findings for the decision90%
  • passReports what cuts against the plan100%
Run
Run
#1
API response time
21 s
Submitted
5 Oct 2026
Opus 5.5Claude

Needs more than a quick edit9 of 12 checks passed

Factual errors in ticket counts (ordering 31 vs 30, credits export 8 vs 10) and word count exceedance mean a PM would need to re-check evidence and edit for length.

Output

Synthesis: what customers need from Larder next (for the CPO)

Headline: The broadest and best-evidenced problem is that customers can't easily tell what actually arrived, and it costs them money and hours every week. Demand for "more suppliers" is real, but nearly all of it comes from one account. That account wants its own four suppliers connected, which is not the same as a marketplace. Next quarter we should build delivery check-in and discrepancy tracking, and handle Calder Group's request as a bounded onboarding project.

Finding 1: Delivery and invoice mismatches are the most widespread pain

Evidence: strong. It is consistent across both sources, comes from many accounts, and customers describe concrete costs.

  • Tickets: "Delivery didn't match order" (34) and "Invoice reconciliation" (29) describe the same problem from two ends: ordered, received and invoiced don't line up. Together they account for 63 of 200 tickets (32%) from about 62 different accounts. By number of customers affected, this is our largest problem.
  • Interviews: 8 of 10 describe it. Examples:
  • About £600 lost in a month on undelivered items (gastropub).
  • A finance team spending days every week chasing credit notes (hotel).
  • Sunday nights reconciling invoices with a highlighter (bistro).
  • Disputes with suppliers that can't be won without proof (bakery).
  • Checking deliveries on a clipboard, then re-typing everything (noodle bar).
  • Paying mismatched invoices because there's no time to check (steakhouse).

Finding 2: "Add more suppliers" is the top ticket category, but it is essentially one customer

Evidence: strong that the demand is concentrated; weak that there is broad demand for a marketplace.

- Calder Group's tickets: 75 tickets (37.5%) fall in this category, but 58 come from Calder Group alone. They repeatedly ask for the same four suppliers: - Brakes (15 tickets) - Their meat supplier (15) - Their dairy supplier (14) - Their wine merchant (14) They want suppliers they already use connected to Larder. They are not asking to discover new suppliers. - Everyone else: The remaining 17 tickets come from 17 accounts and are mostly generic ("Any plans for more suppliers?", "Would like more produce suppliers"). - Interviews: Only 2 of 10 clearly want this (the pizzeria and Calder). The tapas bar is lukewarm ("I suppose") and says paperwork is the real problem. The café group says outright that "finding suppliers isn't the problem."

The earlier claim that "most customers we spoke to want a marketplace" is not supported.

Finding 3: Calder is a real commercial opportunity of a specific kind

Evidence: moderate. This is one account's clear statement, untested.

  • Calder has 14 sites and says it would move all its ordering to us if its suppliers were on the platform.
  • We don't yet know how much of its ordering runs through Larder today or what the account is worth. We need that before sizing the effort.
  • Their need is met by onboarding four named suppliers, not by building a marketplace.

Other signals

  • Login and access (31 tickets) and ordering questions (31) look routine.
  • One subset matters for the recommendation below: 10 tickets are "can't log in on tablet," and the tablet is where delivery check-in would happen.

Limits of this evidence

  • Small interview sample. There are 10 interviews and we don't know how participants were chosen.
  • One quarter of tickets. Ticket volume measures who contacts support, not who has the problem. Customers who silently pay wrong invoices, like the steakhouse, won't file tickets, so mismatch pain is probably under-counted.
  • Data quality. Support staff assign the categories, and a few accounts show inconsistent site counts across tickets.
  • No usage or revenue data yet. We can't weigh Calder against the long tail of smaller accounts.

What this means for next quarter

1. Build delivery check-in and discrepancy tracking. This is our core promise ("track deliveries and invoices") and is where the broad pain sits. Scope: - Check a delivery against the order at the door, on phone or tablet, replacing the clipboard and the re-typing. - Record shorts, substitutions and short weights, with a photo and timestamp to use as evidence in disputes. - Flag invoice lines that don't match what was received. - Raise credit notes and track them until they arrive. - Show credits in the accountant export (8 tickets ask for this). - Fix tablet login as part of the work.

Success measures: - Share of deliveries checked in. - £ of discrepancies recorded and credited. - Fewer tickets in the two mismatch categories. - Time spent on reconciliation, re-checked with these interviewees.

2. Treat Calder as an account project, not a platform bet. - Work with sales to confirm the account's value and how much of its ordering would move. - Then scope onboarding of the four requested suppliers. This is bounded work that doesn't require a marketplace.

3. Don't build a marketplace next quarter. - If we want to test demand, do it cheaply. Follow up with the 17 non-Calder accounts or run a fake-door test in the app. - Revisit with that data.

Claims the judge couldn’t find in the brief

  • Login and access tickets: 31; ordering questions: 31.
  • Show credits in the accountant export (8 tickets ask for this).

Grades and run details

Decision model 100 · LLM judge 69
Decision model checks
  • passUses the supplied evidence correctly39%
  • passAddresses the actual decision95%
  • passRespects explicit constraints52%
  • passIdentifies material uncertainty92%
  • passAvoids unsupported claims47%
  • passProduces the required deliverable95%
  • passSees one account behind the top ticket100%
  • passCorrects the draft's claim98%
  • passWeights behaviour over opinion100%
  • passSays how many sources support each finding72%
  • passFrames findings for the decision99%
  • passReports what cuts against the plan100%
Run
Run
#1
Time to output
72 s
Submitted
5 Oct 2026
Sonnet 5.5API

Usable with a quick edit9 of 12 checks passed

The output is clear, well-structured, and only needs a minor correction to the Ordering ticket count (30, not 31) to be fully accurate and usable.

Output

Synthesis for the CPO: what to build next quarter

Headline: The evidence doesn't support a marketplace. It points to the gap between what we order, what arrives and what we're invoiced. I recommend building delivery checking and reconciliation next quarter.

Corrections to the earlier draft

  • "Add more suppliers is our top category by far" is an artefact of one account. Of 200 tickets, 75 are in that category, and 58 of those come from Calder Group, which repeats four requests: Brakes, a meat supplier, a dairy supplier and a wine merchant. The other 17 come from 17 different accounts. Without Calder, it is our smallest category (17 of 142 tickets, 12%).
  • "Most customers want a marketplace" is wrong. Only 2 of 10 interviewees asked for more suppliers (Calder and the pizzeria). The tapas bar was lukewarm ("I suppose") and went straight to paperwork. The café group said finding suppliers "isn't the problem."

Findings and strength of evidence

1. Customers can't easily verify what was delivered against what they ordered and were invoiced for. Strength: strong. - Interviews: 7 of 10 describe this unprompted, and an eighth (tapas bar) says "paperwork" without detail. Examples are short fish orders found a week later, a chef signing notes unchecked (~£600 in a month), finance chasing credit notes, "my word against theirs," clipboard re-typing, and invoices that never match. - Tickets: "Delivery didn't match order" (34) and "Invoice reconciliation" (29) total 63 tickets, 44% of non-Calder volume, from about 62 distinct accounts. The ticket summaries include "short delivery not shown in app," "invoice includes items not delivered," "credit note missing" and "export doesn't show credits." - Interviews and tickets are independent sources and agree. The combined grouping is my judgement, but delivery tickets alone (34) are still twice the non-Calder supplier requests.

2. Generic demand for more suppliers is small and vague. Strength: moderate. - The 17 non-Calder tickets are "any plans for more suppliers?" (8) and "more produce suppliers" (9). They show mild interest, not a pain point. One interviewee (the pizzeria) wants more choice and better prices.

3. Calder Group has a real, specific need. Strength: strong for Calder, weak as a market signal. - Calder is one account with 14 sites, which says it would move all ordering to us with a marketplace. This is a stated, conditional intent. Its tickets name four suppliers, which is an onboarding request, not evidence for a marketplace. We don't yet know Calder's revenue or how many similar multi-site groups we have.

4. Smaller issues: ordering, logins and tablets. Strength: moderate, low priority. - Ordering (31 tickets) and login/access (31 tickets) are mostly routine. "Can't log in on tablet" matters because delivery checking will likely happen on a tablet at the back door.

What this means for next quarter

Build: a receiving and reconciliation flow. - Staff record what actually arrived at the door: shorts, substitutions and weights. - The app flags mismatches against the order and the invoice. - It produces a timestamped record the owner can send to the supplier as a credit-note claim. - Credits flow into the accountant export.

This addresses the main complaints: finding out a week late, the "my word against theirs" disputes, and double entry. It also helps the people who file tickets about short deliveries and missing credit notes.

Don't build: a marketplace. It's a large investment backed by one account and a few vague requests.

Handle Calder separately. Onboard their four named suppliers through our existing process if feasible. Then ask Calder whether that satisfies their "move everything to you" condition, and get their revenue.

Fix before building: tablet login reliability, since the new flow depends on it.

Caveats

  • The sample is 10 interviews from September, and I don't know how they were recruited. Eight of ten raising the same pain is encouraging but not a measured prevalence.
  • Tickets show what people contact support about, not everything that hurts. Customers with delivery problems may just pay the invoice, as the steakhouse does.
  • Ticket data is one quarter, and the `account_sites` values are inconsistent for a few accounts (for example 1418 and 1625), so I haven't segmented by size.
  • The £600 loss is one anecdote. Before committing scope, we should size the cost of short deliveries and unclaimed credits across a sample of accounts.

What would change my view

If we find a meaningful segment of multi-site groups like Calder, with material revenue that is blocked on supplier coverage, supplier breadth becomes a segment-specific bet. It would then be worth testing through targeted onboarding before any full marketplace.

Claims the judge couldn’t find in the brief

  • Ordering (31 tickets) and login/access (31 tickets) are mostly routine.

Grades and run details

Decision model 100 · LLM judge 77
Decision model checks
  • passUses the supplied evidence correctly36%
  • passAddresses the actual decision100%
  • passRespects explicit constraints53%
  • passIdentifies material uncertainty100%
  • passAvoids unsupported claims56%
  • passProduces the required deliverable95%
  • passSees one account behind the top ticket100%
  • passCorrects the draft's claim99%
  • passWeights behaviour over opinion100%
  • passSays how many sources support each finding87%
  • passFrames findings for the decision99%
  • passReports what cuts against the plan100%
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 6

Uses the supplied evidence correctlyRightMixedMixed
GPT-6 Luna · API

All factual claims about the current situation are directly supported by the supplied interview excerpts or ticket data.

Opus 5.5 · Claude

Two factual numbers are wrong: ordering tickets are 30 not 31, and credits export tickets are 10 not 8.

Sonnet 5.5 · API

The Ordering ticket count is stated as 31, but the CSV contains 30 Ordering tickets, so a factual statement about the current situation is incorrect.

Respects explicit constraintsRightMixedRight
GPT-6 Luna · API

The output is a synthesis under 800 words, includes main findings with evidence strength, and states what to build next quarter, all for the CPO.

Opus 5.5 · Claude

The output exceeds the 800-word limit (approximately 826 words).

Sonnet 5.5 · API

The output is a synthesis for the CPO, well under 800 words, and respects the requested form and reader.

Avoids unsupported claimsRightRightMixed
GPT-6 Luna · API

Interpretations are clearly tied to evidence, and no confident claim goes beyond what the supplied data establishes.

Opus 5.5 · Claude

Hypotheses and forecasts are labelled (e.g., 'untested', 'probably under-counted'), and confident claims are supported.

Sonnet 5.5 · API

The claim 'Ordering (31 tickets)' is presented as an established fact, but the evidence shows 30 Ordering tickets, so it is an unsupported confident claim.

Produces the required deliverableRightMixedRight
GPT-6 Luna · API

The synthesis is complete, within 800 words, addressed to the CPO, and actionable with findings, evidence strength, and a next-quarter plan.

Opus 5.5 · Claude

The synthesis is present but exceeds the 800-word limit, so it is not within the requested length.

Sonnet 5.5 · API

The synthesis is complete, in the right form for the CPO, within the word limit, and actionable with light edits.

Sees one account behind the top ticketMixedRightRight
GPT-6 Luna · API

The output counts Calder's tickets and notes the remaining 17 are from others, but it does not show that delivery and invoice issues are spread across many accounts, only that those categories exist in the export.

Opus 5.5 · Claude

Clearly shows Calder Group accounts for 58 of 75 supplier tickets, while delivery/invoice issues are spread across many accounts.

Sonnet 5.5 · API

Clearly shows that 58 of 75 'Add more suppliers' tickets come from Calder Group, and that delivery/invoice issues are spread across many accounts.

Says how many sources support each findingRightRightMixed
GPT-6 Luna · API

Findings are sized: eight of ten interviews, 75 of 200 tickets, 58 from Calder, 17 from others; no finding is presented as universal without a count.

Opus 5.5 · Claude

Every finding is sized with numbers (8 of 10 interviews, 63 tickets, 58 from Calder, 17 from 17 accounts).

Sonnet 5.5 · API

The finding on smaller issues sizes Ordering as 31 tickets, but the actual count is 30, so the finding is not accurately sized.

All got right 6

Addresses the actual decisionRightRightRight
GPT-6 Luna · API

The output commits to a clear recommendation (prioritize delivery-to-invoice reconciliation, treat marketplace as unproven) and says to measure problem spread to inform future decisions.

Opus 5.5 · Claude

Commits to building delivery check-in, treating Calder as an account project, and not building a marketplace, with conditions to revisit.

Sonnet 5.5 · API

Commits early to building delivery checking and reconciliation, and says what would change the view (finding a meaningful segment of multi-site groups like Calder).

Identifies material uncertaintyRightRightRight
GPT-6 Luna · API

It names unknowns (how common the problem is across all accounts, which part of the workflow is most valuable, what marketplace model) and says to test and measure to resolve them.

Opus 5.5 · Claude

Names specific unknowns (sample size, ticket bias, data quality, no revenue data) and how to resolve them.

Sonnet 5.5 · API

Names unknowns (interview recruitment, ticket bias, inconsistent account_sites, anecdotal £600, need to size cost) and says how they would be resolved or what would change the call.

Corrects the draft's claimRightRightRight
GPT-6 Luna · API

It corrects the draft by counting eight interviews on delivery/invoice pain versus two on marketplace, and shows the draft's headline rests on a minority of interviews.

Opus 5.5 · Claude

Corrects the draft by counting interviews: only 2 of 10 want a marketplace, and 8 describe delivery/invoice pain.

Sonnet 5.5 · API

Counts the interviews (2 of 10 asked for more suppliers) and corrects the headline, showing that most interviews describe delivery and invoice pain.

Weights behaviour over opinionRightRightRight
GPT-6 Luna · API

The output flags the marketplace theme as 'weak evidence' and a 'proposed solution, not a validated priority,' while the delivery/invoice pains are backed by specific instances and financial impact.

Opus 5.5 · Claude

Weights concrete costs and observed behaviour (money lost, time spent) over stated wishes ('would love', 'I suppose').

Sonnet 5.5 · API

Distinguishes specific past instances of delivery/invoice pain from stated preferences and feature requests, and flags Calder's need as 'stated, conditional intent'.

Frames findings for the decisionRightRightRight
GPT-6 Luna · API

Each finding is linked to the decision: the reconciliation pain drives the build priority, the supplier concentration shapes the Calder follow-up, and the weak marketplace evidence justifies not building it.

Opus 5.5 · Claude

Each finding is framed as an opportunity and linked to the decision (build check-in, don't build marketplace, treat Calder as account project).

Sonnet 5.5 · API

Each main finding is framed as an opportunity and linked to the decision, with clear implications for what to build next quarter.

Reports what cuts against the planRightRightRight
GPT-6 Luna · API

It reports that the top ticket category is driven by one account and that the marketplace is an unvalidated solution, directly challenging the draft's plan.

Opus 5.5 · Claude

Reports evidence against the marketplace plan: concentration in one account, interviews mostly about mismatch, café group's statement.

Sonnet 5.5 · API

Reports evidence that cuts against the marketplace hypothesis, including the concentration of supplier requests in one account and the prevalence of delivery/invoice pain.

Results

Every setup we’ve tested on this task type, across all its tasks and repeats, graded on the current checklist. Calibrated. Where our PM checked the graders by hand, they agreed on 74% of 62 checks.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6 AstraChatGPT95.595.82None
2GPT-6.1 SolAPI97.792.02None
3GPT-6 LunaAPI93.284.02None
4Sonnet 5.5API95.567.62None
5Haiku 5.5API86.471.22None
6Opus 5.5Claude86.463.82None
7Gemini 3.8 FlashAPI71.063.12None
8Gemini 3.5 Flash-LiteGemini69.347.42None

What you get for the cost

GPT-6.1 Sol · API scores within a point of GPT-6 Astra · ChatGPT at a ninth of the cost.

Task score against relative cost for each setup. The table below has the same figures.50607080901001×3×10×30×100×Cost per output, relative to the cheapest setup (log scale) →↑ Task scoreGPT-6 Astra · ChatGPT: task score 96, 194× the cheapest (1 of 2 outputs estimated)GPT-6.1 Sol · API: task score 95, 21× the cheapestGPT-6 Luna · API: task score 87, the cheapestSonnet 5.5 · API: task score 80, 62× the cheapestHaiku 5.5 · API: task score 77, 3.4× the cheapestOpus 5.5 · Claude: task score 72, 167× the cheapest (1 of 2 outputs estimated)Gemini 3.8 Flash · API: task score 66, 19× the cheapestGemini 3.5 Flash-Lite · Gemini: task score 56, 6.5× the cheapest (1 of 2 outputs estimated)GPT-6 Astra · ChatGPTGPT-6.1 Sol · APIGPT-6 Luna · APISonnet 5.5 · APIHaiku 5.5 · APIOpus 5.5 · ClaudeGemini 3.8 Flash · 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

Turning a stack of call transcripts into what we actually learned.

Why it matters

Synthesis is where teams fool themselves. A model that smooths away dissent or turns one loud customer into a trend produces confident, wrong roadmaps.

What good looks like

  • Quotes evidence for each theme and counts sources honestly
  • Weights what customers did above what they say they'd do
  • Keeps important dissent visible
  • Labels hypotheses as hypotheses
  • Says what the research cannot tell us

Deliberately not measured

  • Transcript clean-up
  • Persona illustration
Capability tested

Faithful synthesis of qualitative research

The failure we’re looking for

Invents customer consensus or loses important dissent

Grading

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