Tasks / Challenge

Challenge an idea

Can the model find the strongest reason an idea may fail, backed by evidence?

Measures the systemTask 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. Addresses the actual decision100% pass
    It commits early to not committing three squads and authorizing a six-week test, and specifies what result would change that.
    GPT-6.1 Sol · API · The CEO's embedded-payments bet
  2. Identifies material uncertainty100% pass
    It names the unknowns (representativeness of payment data, large-account volume access, actual mix) and resolves them with a bounded test.
    GPT-6.1 Sol · API · The CEO's embedded-payments bet
  3. Uses the interviews faithfully100% pass
    All quotes are accurate and correctly attributed to A05, A08, A04, A07, and A12.
    GPT-6 Luna · API · An AI SDR for small agencies

Where it slips

  1. A cheap test that can actually read out64% pass
    The pre-registered gate requires signed deals and renewal rather than an early signal such as qualified meetings or proposals, so it risks being too late for a cheap read-out.
    GPT-6.1 Sol · API · An AI SDR for small agencies
  2. Re-estimates the revenue correctly71% pass
    It shows working but lands on a central $2.3M and $1M–$6M range, not roughly the $5–10M range required by the grading rubric, and its $9.1M pilot upper bound is not used as the main re-estimate.
    Sonnet 5.5 · API · The CEO's embedded-payments bet
  3. Avoids unsupported claims71% pass
    The memo claims the pilot achieved ~0.38% blended take as a fact, which is not in the evidence and is not derived from it arithmetically; it also treats the pro-rata $1.43B locked value as a hard constraint without flagging the assumption.
    Opus 5.5 · Claude · The CEO's embedded-payments bet

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

We plan to sell an AI sales-development rep, an agent that finds prospects and sends personalised outbound email, to agencies with 5–20 staff. Before we commit a year to it, find the strongest reason this fails. Use the interview material below, including the three transcripts. Write it up for the founding team: the one reason, the evidence for it (quote the interviews), where the evidence cuts the other way, what we would need to see to be proved wrong, and the cheapest test that would show it. Keep it under 600 words.

What the model was given7 items: Scenario, Founders' hypothesis (from our planning doc), Interview summary (14 agencies, 5–20 staff, last six weeks), Interview log, interview-A05-content-agency.md, interview-A04-shopify-agency.md, interview-A08-packaging-studio.md
ScenarioWe are a three-person founding team with £800k of pre-seed funding and about 20 months of runway. Planned price: $600 a month per agency. Two competitors have each raised over $10m in the last year, selling AI SDRs mostly to software companies with 50+ staff.
Founders' hypothesis (from our planning doc)“Small agencies live feast or famine. When a big client leaves they have no pipeline, because the founder is too busy delivering to sell. An AI SDR keeps the pipeline full in the background, so the famine never comes. At $600 a month, one extra client a year pays for it many times over.”
Interview summary (14 agencies, 5–20 staff, last six weeks)11 of 14 get most of their new revenue from referrals and repeat clients (86% on average across those 11). 9 had tried outbound in the last two years; 7 of those stopped within six months, and none of the 7 closed a deal they could attribute to it. The 2 who kept going (A04, A07) have 15–20 staff and a dedicated person for business development. 3 (A08, A12, A14) said they regularly turn work away. 6 described feast-or-famine swings in new business. Average deal size across the 14: $18k; typical sales cycle 6–10 weeks.
Interview logAgency (staff) · services · share of new revenue from referrals and repeat clients · outbound tried? · outcome · representative quote A01 (8) · brand and web studio · 90% · yes, cold-email agency · stopped after 3 months · "We got meetings with people who'd never buy from us." A02 (14) · performance marketing · 85% · yes, LinkedIn automation · stopped after 4 months, account restricted · "LinkedIn shut our founder's account down. That's our best referral network." A03 (6) · PR boutique · 95% · no · — · "Every client we have came from someone vouching for us." A04 (18) · Shopify development · 40% (plus 30% partners, 30% outbound) · yes, in-house BD lead · kept, 30% of new revenue · "It works because of the follow-up, not the first email." A05 (11) · content · 85% · yes, lead-gen agency · stopped after 2 months, no deals · "Content is a trust purchase." A06 (7) · video production · 85% · no · — · "Our clients find us through the videos. Someone shares one and we get a call." A07 (16) · SEO · 45% · yes, outsourced lead generation · kept · "It pays for itself, just. Most of the meetings are a waste of time, but one in ten turns into a retainer." A08 (5) · packaging design · 95% · no · — · "If you sent me ten more leads in October I'd have to say no to nine of them." A09 (12) · paid social · 75% · yes, cold email · stopped after 5 months · "The replies were mostly people asking us to take them off the list." A10 (20) · B2B marketing · 70% · yes, contract SDR · stopped after 6 months · "It cost us about £400 for every meeting, and the meetings didn't close." A11 (9) · WordPress web agency · 90% · yes, cold email · stopped after 3 months · "Our domain ended up on a spam list. Took weeks to fix." A12 (10) · branding · 85% · no · — · "We're booked out till March. I don't need more leads, I need another designer." A13 (15) · HubSpot partner · 48% (plus 40% from HubSpot's partner directory) · yes, cold email · stopped after 4 months · "HubSpot sends us more than we can handle in a good quarter." A14 (6) · UX research · 90% · no · — · "We're small on purpose. We say no to about a third of enquiries."
interview-A05-content-agency.md31 lines · Download
# Interview A05: content agency, 11 staff
Participant: founder and managing director. Interviewer: our co-founder. 34 minutes, lightly edited.

**Interviewer:** Where did your last five clients come from?

**Participant:** Let me think. Two were old clients coming back: one had changed jobs and brought us into her new company. Two were introductions, one from a web agency we partner with and one from a client's CFO who'd seen our work. The fifth found us through a talk I gave at a SaaS marketing meetup. So none of them from anything you'd call outbound.

…
interview-A04-shopify-agency.md23 lines · Download
# Interview A04: Shopify development agency, 18 staff
Participant: head of business development. Interviewer: our co-founder. 29 minutes, lightly edited.

**Interviewer:** Tell me how new business works for you.

**Participant:** We're a bit unusual for an agency our size. I'm a full-time business-development person, and outbound is about 30% of our new revenue. Referrals and repeat work are still the majority, maybe 40% referrals and the rest partners: Shopify's partner directory and a couple of app companies who send us projects.

…
interview-A08-packaging-studio.md23 lines · Download
# Interview A08: packaging design studio, 5 staff
Participant: founder and creative director. Interviewer: our co-founder. 22 minutes, lightly edited.

**Interviewer:** How do new clients find you?

**Participant:** Almost entirely word of mouth. Food and drink brands talk to each other. Someone launches a range, their packaging does well on the shelf, and the next founder asks who did it. I'd say 95% referral. We've never done any outbound.

…
What a strong answer doesThe answer key the graders mark against

Names the load-bearing assumption as 'small agencies' growth is limited by a shortage of new leads': the interviews say it is limited by trust and capacity. 11 of 14 win most work through referrals and repeat clients; 7 of 9 who tried outbound quit within six months with no attributable deals, for reasons tied to trust (A05's 'content is a trust purchase', A01's meetings with people who'd never buy) and reputation risk (A02's restricted LinkedIn account, A11's spam listing); 3 turn work away (A08: 'I'd have to say no to nine of them'). Quotes the interviews accurately and attributes them correctly. Engages the counter-evidence honestly: outbound works for A04 and A07, but both are 15–20 staff with a dedicated BD owner, already use tools like Apollo and Clay, and credit the follow-up rather than the first email, so the segment is small, tooled and not what the product automates. Treats feast or famine as a real pain but notes agencies fix it through past clients and referrals (A05's reactivation emails, and her willingness to pay for help keeping in touch with past clients), pointing at a different product. Says what would prove the challenge wrong (for example, a meaningful share of 10–20-staff agencies with a BD owner and budget, or a concierge outbound test producing qualified meetings that turn into proposals), and designs a cheap test that measures an early signal, since a 6–10 week sales cycle means a six-week test cannot observe won deals.

Critical failures (cap the score at 40)
  • Invents interview quotes or attributes a quote to the wrong agency
  • Treats the founders' feast-or-famine hypothesis as confirmed by the interviews
Case

Staff-level · v1.10 · synthetic · B2B, go-to-market, customer research

The outputs

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

Needs more than a quick edit3 of 11 checks passed

A PM would need to re-check evidence and repair the test before acting. The core argument is useful but not reliable as written.

Output

The One Reason This Fails

Agency services are high-trust purchases that cold automated outreach cannot convert without intensive, bespoke human follow-up—which time-poor founders cannot provide. Automating email generation solves top-of-funnel activity, but the actual point of failure is trust and follow-up capacity.

---

The Evidence

1. Zero conversion for founder-led shops: 7 of 9 agencies that tried outbound stopped within six months having closed zero deals. 2. Cold outreach delivers the wrong buyers: Services require credibility upfront. * A05 (11 staff): “Content is a trust purchase... The people on those calls had no idea who we were and, honestly, most of them didn’t have a budget.” * A01 (8 staff): “We got meetings with people who’d never buy from us.” * A10 (20 staff): “Cost us about £400 for every meeting, and the meetings didn’t close.” 3. The constraint is manual follow-up, not initial emails: * A04 (18 staff, dedicated BD): “It works because of the follow-up, not the first email... What wins deals is that I then send them a teardown of their site, a short video... The owner does sales on the side. They don't have the time to follow up, and the follow-up is the job.” 4. Asymmetric downside risk: Small agencies live on reputation. * A11: “Our domain ended up on a spam list.” * A02: “LinkedIn shut our founder’s account down. That’s our best referral network.” * A05: “I’d worry about our name going out on emails I haven’t read.”

---

Where the Evidence Cuts the Other Way

  • Feast-and-famine is real: 6 of 14 experience painful dry spells. A05 panicked when three clients left simultaneously.
  • Outbound works at scale with dedicated staff: A04 generates 30% of revenue via outbound; A07 makes it pay (“one in ten turns into a retainer”).
  • Willingness to pay exists if positioned right: A04 would pay $600/month if it replaced Apollo/Clay and saved research hours. A05 would pay $200/month for an automated tool to reactivate past clients and referral networks.

---

What We Would Need to See to Be Proved Wrong

  1. A solo founder (with no dedicated BD rep) closing an outbound deal with typical agency pricing ($18k) using standard outbound messaging, without spending >5 hours/week on bespoke pitch work.
  2. Willingness of agency owners to allow an autonomous agent to send cold outbound without manual pre-approval of every single touchpoint.

---

The Cheapest Test to Prove It

A 3-week "Wizard of Oz" Concierge Test with 3 Agencies (Cost: <$500): * Partner with 3 agencies suffering feast-or-famine (e.g., A05, A09, A10) who will commit to running a free, 30-day outbound campaign. * Manually emulate the AI SDR: scrape prospect lists, generate personalized emails, and handle automated sequencing from a secondary domain. * Kill criteria: If after booking 10–15 meetings across these agencies, zero advance to contract negotiations—or founders drop out because they cannot keep up with follow-ups and custom scoping—kill the product.

(Pivot signal: A05 offered to pay immediately for an AI assistant that drafts warm re-engagement notes to past clients and referral partners. That solves the same famine problem without the friction of cold outbound.)

Claims the judge couldn’t find in the brief

  • 7 of 9 agencies that tried outbound stopped within six months having closed zero deals.
  • 6 of 14 agencies experience painful dry spells.
  • A04 generates 30% of revenue via outbound.
  • A05, A09, and A10 are agencies suffering feast-or-famine.

Grades and run details

Decision model 95 · LLM judge 3 of 12 checks
Decision model checks
  • passUses the supplied evidence correctly44%
  • passAddresses the actual decision96%
  • passRespects explicit constraints36%
  • passIdentifies material uncertainty97%
  • partialAvoids unsupported claims23%
  • passProduces the required deliverable71%
  • passFinds the load-bearing assumption26%
  • passUses the interviews faithfully74%
  • passEngages the counter-evidence79%
  • passA cheap test that can actually read out94%
  • passTigers, not paper tigers94%
Run
Run
#1
API response time
22 s
Submitted
1 Oct 2026
Critical failure: Treats the founders' feast-or-famine hypothesis as confirmed by the interviews. Score capped at 40.

Usable with a quick edit11 of 11 checks passed

The memo is well-structured, evidence-backed, and actionable, requiring at most light editing for polish.

Output

Why the AI SDR for small agencies fails

Sep 25, 2026 · @Martin Slaney

The reason

Small agencies win work on trust, and cold outbound only closes when someone does the follow-up. An AI SDR automates the first email, the cheap part. It hands the follow-up, which is the actual job, to a founder our own hypothesis says is too busy delivering to sell. We would be adding meetings to the one resource that is already the constraint.

Evidence

  • Referrals dominate. 11 of 14 get most new revenue (86% on average) from referrals and repeat clients. A03: "Every client we have came from someone vouching for us."
  • Outbound was tried and dropped. 9 tried it; 7 stopped within six months and none of the 7 closed an attributable deal. They used five different channels (lead-gen agency, cold-email agency, contract SDR, LinkedIn automation, DIY email), so this isn't one bad vendor.
  • Meetings weren't the problem; closing was. A05 got six or seven meetings a month and closed nothing: "Content is a trust purchase." A01: "We got meetings with people who'd never buy from us." A10: "It cost us about £400 for every meeting, and the meetings didn't close."
  • Where it works, a human does the work. Both survivors (A04, A07) have 15–20 staff and a dedicated BD person. A04: "It works because of the follow-up, not the first email" and "most agencies our size don't have someone like me."
  • The downside hits their best channel. A02 lost the founder's LinkedIn, "our best referral network"; A11's domain was spam-listed. A04 and A05 both refuse to let anything send unread.
  • Some don't need leads at all. A08, A12 and A14 turn work away. A12: "I don't need more leads, I need another designer."

Where it cuts the other way

  • The famine is real.1 6 of 14 describe feast-or-famine swings; A05 "spent April panicking" after three clients left at once.
  • The maths are forgiving. At an $18k average deal, one closed client a year covers $7.2k of subscription.
  • One buyer exists. A04 would pay $600 today if it replaced Apollo, Clay and three hours a day of research, but only with HubSpot sync and approval of every email.
  • Past failures were generic. A05's buyers wanted proof of work in their niche; better enrichment might fix that. Untested.
  • Small sample. 14 interviews, one interviewer.
  • Warm outreach worked. A05's emails to past clients won two back in a month, and she'd pay "a couple of hundred a month" for a tool that did it. That supports the thesis against cold outbound, but points at a different product.

What would prove us wrong

Agencies without a dedicated BD person closing deals from AI-sent cold email, without damage to their reputation. Concretely, in a 12-week pilot with 5 such agencies:

  • at least 3 reach a scoped proposal and at least 1 closes;2
  • no domain, inbox or LinkedIn penalties;
  • owners spend under 2 hours a week on follow-up;
  • they keep paying when the pilot ends.

Cheapest test

Pre-sell a hand-run pilot before building anything. Offer the 9 agencies with no BD person that aren't turning work away (A01–A03, A05, A06, A09–A11, A13) a 12-week pilot at $600 a month, paid upfront. We run it by hand: Apollo and Clay for prospects, LLM-drafted emails, owner approves every send, separate sending domain.

  1. Gate 1, one week: fewer than 3 pay, stop. There is no demand at this price.
  2. Gate 2, 12 weeks (covers the 6–10 week sales cycle): score against the thresholds above and log owner follow-up hours.

In parallel, offer A05's version (past-client and referrer re-engagement, about $200 a month) as a second arm. If that sells and cold doesn't, we have our answer.

Cost: roughly £1–2k in tools and a quarter of one founder's time3, against a year and most of the £800k.

What a PM had to fix

  1. 1Hypothesis stated as factReframe it as a hypothesisTargeted repair

    Source: Founders' hypothesis (from our planning doc)Small agencies live feast or famine.

    What we’d changeSay 'Famine is real for some: 6 of 14' rather than confirming the founders' hypothesis for the whole segment.

  2. 2Test or gate too weakTighten the testQuick edit

    What we’d changeRaise the bar for proving the memo wrong: one close in five agencies doesn't show the product works.

  3. 3OtherFix itQuick edit

    What we’d changeThe cost looks light for running a hand-run pilot across nine agencies. Build it up line by line.

Grades and run details

Decision model 95 · LLM judge 12 of 12 checks
Decision model checks
  • passUses the supplied evidence correctly42%
  • passAddresses the actual decision95%
  • partialRespects explicit constraints40%
  • passIdentifies material uncertainty100%
  • passAvoids unsupported claims16%
  • passProduces the required deliverable79%
  • passFinds the load-bearing assumption92%
  • passUses the interviews faithfully79%
  • passEngages the counter-evidence97%
  • passA cheap test that can actually read out92%
  • passTigers, not paper tigers98%
Run
Run
#1
Time to output
60 s
Submitted
25 Sept 2026

Usable with a quick edit11 of 11 checks passed

The memo is well-structured, evidence-based, and actionable; a product manager could use it with only light editing.

Output

Founding-team memo: the strongest reason this fails

Reason: The product may solve the wrong bottleneck. Many small agencies are not short of names to contact: they win through trust, referrals and repeat work, and some already lack capacity to take on more. Cold outbound can create low-quality meetings—or reputational risk—without solving the follow-up needed to convert them. At $600/month, an AI SDR could therefore add work rather than prevent famine.

Evidence for it: In 11 of 14 agencies, referrals and repeat clients account for most new revenue (86% on average). Of the nine that tried outbound, seven stopped within six months; none attributed a deal to it. A05’s experience is telling: a lead-gen firm booked “maybe six or seven a month,” but “most of them didn’t have a budget for content.” The founder said, “Content is a trust purchase,” and would pay for prompts and drafts to contact past clients—not “$600 for cold outbound.”

Capacity is also a real constraint: A08 said, “If you sent me ten more leads in October I’d have to say no to nine of them”; A12 needs another designer, not more leads. And even outbound that works appears to require substantial human selling. A04 said, “It works because of the follow-up, not the first email,” then described sending a site teardown, video and making a call. A07 said most meetings were “a waste of time.” The risk is not merely that the AI writes mediocre emails; it is that the agency cannot or does not convert what it sends.

Where the evidence cuts the other way: Six agencies reported feast-or-famine swings, and A05 described panicking after three clients ended together. A04 gets 30% of new revenue from outbound and would pay if the tool saved research time; A07 says its outsourced lead generation “pays for itself, just.” With an $18k average deal, one win could easily cover the subscription. These examples show outbound can work—but the two continuing users are larger agencies with dedicated BD, unlike the founder-led customer we plan to target. A05’s warm reactivation success also suggests the underlying need may be real, but the best channel may not be cold prospecting.

What would prove us wrong: Typical 5–20-person agencies without dedicated BD will pay $600/month for cold outbound, follow up promptly, and generate attributable profitable work—not just meetings—without harming their reputation or exceeding delivery capacity.

Cheapest test: Don’t build the agent. Recruit five founder-led agencies with capacity and a genuine pipeline gap, charge the planned $600/month, and manually run a 12-week, human-approved outbound pilot. Track qualified meetings, founder follow-up time, attributable wins and renewals. Given 6–10-week sales cycles, a strong signal would be at least two attributable wins and three agencies renewing at full price. If interest stops at free trials, meetings fail to convert, or founders cannot follow up, stop or test a warm-client reactivation product instead.

Grades and run details

Decision model 100 · LLM judge 12 of 12 checks
Decision model checks
  • passUses the supplied evidence correctly65%
  • passAddresses the actual decision93%
  • passRespects explicit constraints60%
  • passIdentifies material uncertainty100%
  • passAvoids unsupported claims49%
  • passProduces the required deliverable94%
  • passFinds the load-bearing assumption99%
  • passUses the interviews faithfully75%
  • passEngages the counter-evidence96%
  • passA cheap test that can actually read out73%
  • passTigers, not paper tigers96%
Run
Run
#1
API response time
27 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 correctlyMixedRightRight
Gemini 3.8 Flash · API

It misstates or overreaches several current-situation facts, including 'zero deals', 'painful dry spells', A04's 30% of revenue, and A09/A10 as feast-or-famine agencies.

Opus 5.5 · Claude

Every factual claim about the current situation is drawn correctly from the brief or supplied context, with accurate quotes and attributions.

GPT-6 Luna · API

All factual claims about the current situation are directly supported by the supplied context, with no inventions.

Respects explicit constraintsMixedRightRight
Gemini 3.8 Flash · API

The required sections and length are present, but the proposed test is internally inconsistent (3-week test vs 30-day campaign) and would not reliably enforce the brief's cheap-test requirement.

Opus 5.5 · Claude

The output is a memo under 600 words, covers all requested sections, and the proposals (test, thresholds) enforce the constraints in practice.

GPT-6 Luna · API

The memo is written for the founding team, includes all requested sections, and stays under 600 words.

Avoids unsupported claimsWrongRightRight
Gemini 3.8 Flash · API

It presents unsupported or overconfident claims as fact, especially the feast-or-famine status of A09/A10 and the zero-deal characterisation.

Opus 5.5 · Claude

Interpretations such as 'small agencies win work on trust' are clearly positioned as the reason and backed with evidence; no causal claims are presented as fact without support.

GPT-6 Luna · API

Interpretations and risks are clearly framed as hypotheses, not as established facts.

Produces the required deliverableMixedRightRight
Gemini 3.8 Flash · API

Although structured as requested, it is not usable without correcting evidence errors and the contradictory test design.

Opus 5.5 · Claude

The memo includes the reason, quoted evidence, counter-evidence, a falsification condition, a cheap test, and stays under 600 words; a founding team could act on it.

GPT-6 Luna · API

The deliverable is a complete, actionable memo for the founding team, within the word limit.

Finds the load-bearing assumptionMixedRightRight
Gemini 3.8 Flash · API

It focuses on trust and follow-up capacity but does not make lead-constrained growth the central challenge or use the turning-work-away evidence that directly refutes the founders' assumption.

Opus 5.5 · Claude

The critique directly challenges the assumption that these agencies are lead-constrained, arguing that they lack the capacity to convert leads and that outbound fails without follow-up.

GPT-6 Luna · API

The memo targets the load-bearing assumption that growth is limited by a shortage of new leads, using referral, abandoned-outbound and capacity evidence.

Engages the counter-evidenceMixedRightRight
Gemini 3.8 Flash · API

It names A04 and A07 but does not fully explain what sets them apart or the implication for the addressable segment, such as A07's dedicated BD owner and existing tooling.

Opus 5.5 · Claude

A04 and A07 are named, their shared characteristics (size, dedicated BD, follow-up effort) are identified, and the implication that the addressable segment is very small is drawn.

GPT-6 Luna · API

It names A04 and A07, notes their larger size and dedicated BD, and explains that follow-up, not the first email, drives their success.

A cheap test that can actually read outMixedRightRight
Gemini 3.8 Flash · API

The test is cheap in spirit but inconsistent and not clearly observable within the 6–10 week sales cycle, using contract negotiations after a 30-day campaign.

Opus 5.5 · Claude

The test is a hand-run pilot that measures proposals within 12 weeks (covering the sales cycle) and uses simple gates; it is cheap and provides an early signal before full build.

GPT-6 Luna · API

The test is a cheap, manual pilot with a concrete threshold, tracking early signals (qualified meetings, follow-up time) and acknowledging the 6–10 week sales cycle.

Tigers, not paper tigersMixedRightRight
Gemini 3.8 Flash · API

It does not triage risks or dismiss a plausible worry; it leaves capacity/turning-work-away and the small addressable segment underdeveloped.

Opus 5.5 · Claude

It dismisses the worry that past failures were just due to poor tooling as untested, and surfaces the unspoken issue that a warm-outreach re-engagement tool might be a better product.

GPT-6 Luna · API

It separates the real bottleneck (trust, capacity, follow-up) from manageable risks, dismisses the idea that outbound never works, and surfaces the unspoken issue that agencies lack conversion capacity.

All got right 3

Addresses the actual decisionRightRightRight
Gemini 3.8 Flash · API

It commits early to one failure reason for the founding team and states conditions that would prove the challenge wrong.

Opus 5.5 · Claude

The output commits to a single reason for failure early on, addresses the founding team, and specifies the evidence and conditions that would change the call.

GPT-6 Luna · API

The output commits to a clear answer (the product solves the wrong bottleneck) and states what would prove it wrong.

Identifies material uncertaintyRightRightRight
Gemini 3.8 Flash · API

It names material unknowns—solo-founder conversion and autonomy tolerance—and gives a test and kill criteria to resolve them.

Opus 5.5 · Claude

It names the small sample size, whether better enrichment could fix niche trust, and that the test result would resolve the uncertainty.

GPT-6 Luna · API

It names the key unknown (whether typical agencies without dedicated BD will pay and convert) and says how a test would resolve it.

Uses the interviews faithfullyRightRightRight
Gemini 3.8 Flash · API

The quoted interview material is attributed to the right agencies and is accurate apart from minor ellipses and wording.

Opus 5.5 · Claude

All quoted material is taken from the transcripts or log, attributed to the correct agency, and none is invented.

GPT-6 Luna · API

All quotes are accurate and correctly attributed to A05, A08, A04, A07, and A12.

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 88% of checks.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6.1 SolwithAPI100.095.82None
2GPT-6 AstrawithChatGPT93.295.82None
3GPT-6 LunawithAPI97.791.72None
4Sonnet 5.5withAPI90.975.02None
5Gemini 3.5 Flash-LitewithGemini88.675.021 capped
6Opus 5.5withClaude93.291.721 capped
7Gemini 3.8 FlashwithAPI93.237.521 capped

About the task

The PM job

Pressure-testing a proposal before committing a team to it.

Why it matters

The useful critic finds the one assumption everything rests on. Theatrical negativity is easy to generate and useless in a planning meeting.

What good looks like

  • Identifies the load-bearing assumption
  • Separates the risks that could kill it from the ones that only look scary
  • Uses the supplied evidence, not generic risks
  • Proposes the cheapest way to test the assumption

Deliberately not measured

  • Tone
  • Number of objections raised
Capability tested

Evidence-based critique

The failure we’re looking for

Theatrical negativity without evidence

Grading

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

Variants

Vanilla prompt (core) · With Roast Me skill · Staff level: a company bet with a long evidence pack

This task measures the whole setup. Tools, instructions and skills in the harness do real work here, so read the harness as carefully as the model name.