Tasks / Experiment

Find the growth loop

Can the model find a product's real growth loop, show whether it compounds, and say which lever to pull?

Measures the modelTask type v1.0 · 2 tasksLast changed 2 Oct 2026 · ChangelogDifficulty

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

From 14 graded outputs by 7 models. 33% were usable with at most a quick edit.

Reliably right

  1. Sees the cross-side effect100% pass
    It traces the chain from the tutor bounty to oversupply, thinner bookings, new profiles without reviews, and weaker ranking, and acts on it by pausing broad tutor referrals.
    GPT-6.1 Sol · API · Growing on the surface, decaying underneath
  2. Addresses the actual decision96% pass
    The memo commits early to putting both engineers on idea 4, names the primary loop and its compounding status, and specifies what results would change the call (kill thresholds, quarter-end loop gain).
    Sonnet 5.5 · API · The badge on every form
  3. Produces the required deliverable96% pass
    The memo answers all parts of the brief (primary loop, compounding, engineer allocation, success measurement) in a usable form for the Head of Growth.
    Sonnet 5.5 · API · The badge on every form

Where it slips

  1. The loop maths holds46% pass
    The memo does not give a plain verdict of 'decaying' for the content loop despite showing its decline, and it does not compute a numeric yield or coefficient for that loop.
    GPT-6.1 Sol · API · Growing on the surface, decaying underneath
  2. Uses the supplied evidence correctly57% pass
    The claim that the base settles at 10,700 creators is unsupported by the pack's arithmetic, and the claim that cost per sign-up usually rises with spend is not in the supplied evidence.
    Opus 5.5 · Claude · The badge on every form
  3. Avoids unsupported claims59% pass
    Presents the 10,700 equilibrium and the rising cost-per-sign-up claim as facts without labelling them as hypotheses or supporting them from the pack.
    Opus 5.5 · Claude · The badge on every form

The tasks

Read the brief, then put up to three outputs side by side, each with the LLM judge’s verdict on every check. Highlights mark what a PM had to fix.

The brief

You're a PM at Pollen. Our Head of Growth, Sam Okoro, has two engineers for next quarter and four ideas for how to use them. Write Sam a memo of no more than 800 words that says what our primary growth loop is, whether it's compounding, and where the two engineers should go, with how we'll know it worked. Everything we know is below.

What the model was given5 items: About Pollen, Creators, Where new creators come from (last month), Revenue, The four ideas on the table
About PollenA free form and survey builder. Every published form shows a small 'Made with Pollen: make your own' badge at the bottom. Creators can upgrade to Pro for $20 a month to remove the badge and unlock logic and integrations.
Creators18,000 creators published at least one form last month, publishing 40,000 forms between them. Each form gets 120 respondents on average. 78% of last month's active creators were active again this month.
Where new creators come from (last month)Badge: 0.9% of respondents clicked the badge and 11% of those signed up; 38% of badge sign-ups published a form within 30 days, and 85% of those were active again the next month. Template pages (written by our team, ranking in search): 6,000 sign-ups, 14% published, 71% active the next month. Paid search: 2,100 sign-ups at $38 per sign-up, 21% published, 64% active the next month.
Revenue9% of creators who publish upgrade to Pro, and Pro customers stay for 14 months on average.
The four ideas on the table1. Double the paid search budget (Finance has approved it). 2. 'Build our SEO loop': 200 more template pages. 3. A referral programme: $10 of Pro credit for each friend who signs up. 4. Replace the badge with 'Make a form like this', which opens the editor with a copy of the form the respondent just filled in. A two-week pilot on 500 forms raised badge clicks from 0.9% to 1.6% of respondents; 11% of them signed up, as before, and 52% of those published within 30 days.
What a strong answer doesThe answer key the graders mark against

Names the badge as the primary loop: creators publish forms, respondents see the badge, some become creators who publish more forms. It's chosen because badge creators publish and stay best (38% publish, 85% retained), not because it brings the most sign-ups. Sizes it: 120 × 0.9% × 11% × 38% ≈ 0.045 new publishing creators per form, about 1,800 a month from 40,000 forms. Puts that against churn: 22% of 18,000 is about 3,960 creators lost a month, while all sources together add about 3,090 (1,806 + 840 + 441), so the base is shrinking and the loop contributes rather than compounds. Shows paid isn't a loop: $38 ÷ 21% ≈ $181 per publishing creator against about $25 of revenue each (9% × $20 × 14), so doubling it burns money. Calls the template pages a channel, not a loop, since our team writes them and creators don't. Gives the referral credit little weight, because it only matters to the 9% who pay. Puts both engineers on 'Make a form like this': the pilot gives 120 × 1.6% × 11% × 52% ≈ 0.11 creators per form, about 2.4× today, which would add roughly 2,600 creators a month if it held. Flags that the pilot was small and short, and proposes a full test with a threshold (for example, publishing creators per 1,000 respondents) and a stop condition. Mentions retention as the other lever.

Critical failures (cap the score at 40)
  • Recommends doubling paid acquisition, or calls it a growth loop
Task

Mid-level · v1.0 · synthetic · B2C, PLG, viral loop, unit economics

The outputs

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

Needs more than a quick edit9 of 11 checks passed

The memo contains an unsupported equilibrium number (10,700) and an unsupported claim about rising cost per sign-up, so a PM would need to re-check those parts before using it.

Output

To: Sam Okoro, Head of Growth From: [PM], Growth Re: Our growth loop, and where the two engineers should go next quarter

Summary: Our primary loop is the badge. Today it does not compound, and the active creator base is shrinking. The "Make a form like this" pilot is the only one of the four ideas that could make it compound. Both engineers should ship it properly, with a holdout. We should not double paid search.

1. The loop

Creators publish forms → respondents see the badge → some respondents become creators → they publish forms.

Last month the loop worked like this: - 40,000 forms × 120 respondents = 4.8M respondents - 0.9% clicked the badge = 43,200 clicks - 11% of clicks signed up = 4,752 sign-ups - 38% of sign-ups published = 1,806 new publishing creators

The badge is our largest source of creators. Templates produced 840 publishers and paid search 441. Badge creators also retain best: 85% are active the next month, against 71% for templates and 64% for paid.

2. Is it compounding? Not yet.

Each active creator produces about 267 respondents a month, which yields 0.10 new publishers per creator per month. With 78% monthly retention, a creator stays about 4.5 months. Over that lifetime, each creator brings in about 0.46 new creators.

A loop compounds only when this number exceeds 1. At 0.46, the badge amplifies our other channels by about 1.8×, but it cannot sustain itself.

The overall picture is worse: - We keep 14,040 of last month's 18,000 active creators. - We add about 3,100 new publishers. - That gives about 17,100 active creators, a 5% monthly decline. - At current rates the base settles at roughly 10,700 creators.

This assumes the 78% retention figure covers the whole base, which is how I read the data.

3. The four ideas

Double paid search: no. - Each sign-up costs $38. Only 21% publish, so each publishing creator costs about $181. - Each publisher is worth about $25 in revenue (9% upgrade × $20 × 14 months). - Adding the loop's downstream creators raises that to only about $46. - Doubling spend would add about 440 publishers a month for an extra $80k a month, and cost per sign-up usually rises with spend. - This idea doesn't need engineers. I'd ask Finance to hold the budget rather than spend it because it's approved.

200 more template pages: not now. - This is not a loop, because new creators don't create more pages. Our team writes them, so output grows only as fast as we write. - We don't know how many pages produce today's 6,000 sign-ups, so we can't estimate the yield. - Templates publish at only 14%. - This is mainly content work, not engineering work.

Referral credit: no. - It rewards sign-ups, not publishing, which invites low-quality and fake sign-ups. - $10 of Pro credit means nothing to the 91% of creators who never upgrade.

"Make a form like this": yes. The pilot changed every stage of the loop:

Current badgePilot
Click rate0.9%1.6%
Sign-up rate11%11%
Publish rate38%52%
New publishers per 1,000 respondents0.380.92

At network scale, that would mean about 4,400 badge publishers a month instead of 1,800. Each creator's lifetime yield would rise from 0.46 to about 1.1, so the loop would start to compound. Next month's active base would be about 19,700 instead of 17,100.

The margin is thin. The loop reaches the break-even of 1.0 at 0.83 publishers per 1,000 respondents. The pilot beat that by only about 10%. We also need to treat the pilot result with caution: - It covered 500 forms, which may not be representative. - It ran for two weeks, so check whether the 30-day publish window had fully elapsed. - We have no next-month retention data for these creators. People who copy a form may be less committed than people who build one.

4. What the engineers build

1. Staged rollout with a 20% holdout of forms on the old badge, randomised by form and run for at least 8 weeks so we see 30-day publishing and next-month retention. 2. Creator controls and safety: - Copies take the form structure only, never responses. - Creators can turn off copying for sensitive forms. 3. Instrumentation for each stage: click, sign-up, publish, next-month activity, and Pro upgrade. Results should be split by form type so we learn which forms spread best.

5. How we'll know it worked

Primary metric: new publishing creators per 1,000 respondents, treatment vs holdout. - Ship to 100% if it stays at or above 0.83, the level needed for the loop to compound. - Below about 0.6, the change is still positive but the loop won't compound. We would keep it and look for other levers.

Quality checks: - Next-month retention of copy sign-ups should be at least 70%. If it falls much lower, the extra volume won't last. - The upgrade rate of these creators to Pro should be close to 9%.

Guardrails: - Form completion rate is unchanged. - Creator complaints and copy opt-outs are low. - Pro upgrade rate across the base is unchanged.

Business outcome: the active creator base grows month on month by the end of the quarter, instead of falling about 5%.

Bottom line: paid search and templates rent growth one sign-up at a time. The copy-this-form change is our one realistic chance to make growth compound, and a proper rollout will tell us within a quarter whether it does.

Claims the judge couldn’t find in the brief

  • At current rates the base settles at roughly 10,700 creators.
  • Cost per sign-up usually rises with spend.

Grades and run details

Decision model 91 · LLM judge 9 of 12 checks
Decision model checks
  • passUses the supplied evidence correctly27%
  • passAddresses the actual decision100%
  • passRespects explicit constraints33%
  • passIdentifies material uncertainty100%
  • partialAvoids unsupported claims44%
  • passProduces the required deliverable88%
  • passCalls out the paid maths99%
  • passPicks the lever with the most yield98%
  • passA closed loop, not a channel100%
  • passThe loop maths holds11%
  • partialProposes tests that could fail52%
Run
Run
#1
Time to output
55 s
Submitted
2 Oct 2026

Usable with a quick edit9 of 11 checks passed

The memo is complete and actionable; a PM could add a stop condition and measurement window with a quick edit without redoing any reasoning.

Output

To: Sam Okoro, Head of Growth From: Product Subject: Put both engineers on the respondent-to-creator loop

Recommendation

Assign both engineers to “Make a form like this” next quarter: one to a safe, reliable copy-to-editor experience; the other to attribution, experimentation, and activation measurement. Do not double paid search or build the referral programme. Keep existing template acquisition running, but do not make 200 new pages the engineering priority.

Our primary growth loop

Creators publish forms → respondents encounter Pollen → some become creators and publish → their forms reach more respondents.

The badge closes this loop. Templates and paid search feed creators into it, but are not themselves loops: team-written SEO pages require more team production, and paid acquisition requires more spend.

Using the supplied volumes, last month’s 40,000 forms generated approximately 4.8 million respondent encounters:

SourceNew creators publishingStill active the following month
------:---:
Badge1,8061,535
Templates840596
Paid search441282

The badge is our largest acquisition source by activated creators and our strongest by next-month retention: 85%, versus 71% for templates and 64% for paid.

These badge totals assume the respondent volume is eligible to see the badge. Pro removes it, so we must measure actual badge exposure before treating these as forecasts.

Is it compounding?

It recirculates, but we have not demonstrated self-sustaining compounding.

At current rates, each active creator generates approximately 0.10 new publishing creators per monthly production cycle. Meanwhile, 22% of active creators do not return the next month.

A simplified replacement model therefore gives:

78% returning + 10% newly generated = 88% of the starting base.

The loop alone does not currently replace creator attrition. External acquisition can still grow Pollen, but that is different from autonomous compounding. This calculation is directional: publication, activation, and retention windows are not perfectly aligned, and future cohorts may retain differently.

The pilot is promising. Its click and publication rates imply 2.43 times as many new publishing creators per respondent:

  • Current: 0.376 publishing creators per 1,000 respondents
  • Pilot: 0.915 per 1,000

At unchanged volume, that would mean approximately 4,393 new publishing creators, versus 1,806 today. The same simplified replacement model reaches 78% + 24.4% = 102.4%—potentially crossing the self-sustaining threshold, but narrowly. Badge coverage, cohort maturity, retention, and repeat production could erase that margin.

Why this beats the alternatives

Paid search: Current spend is $79,800, or approximately $181 per publishing creator. Applying our overall upgrade rate and Pro lifetime, expected lifetime subscription revenue per publishing creator is only 9% × $20 × 14 = $25.20, before costs. Channel-specific monetization could differ, but the supplied economics do not justify scaling. Finance approval is not evidence of profitable growth.

More templates: They produce useful acquisition, but weaker activation and retention than the badge. We also have no evidence that 200 additional pages would reproduce existing traffic or conversion.

Referral credits: Paying for sign-ups rewards an intermediate action, invites low-quality referrals, and lacks pilot evidence. Revisit an activation-based incentive later.

Delivery and proof

Engineer 1: Build the contextual editor entry and reliable form-copy flow. Copy only permitted form structure—not respondent submissions, credentials, or integrations—and provide appropriate creator controls. Preserve Pro’s badge-removal benefit.

Engineer 2: Instrument exposure through publishing, next-month activity, downstream form distribution, and Pro conversion; run a persistent randomized test, clustered by source creator.

Our primary success metric is incremental creators publishing within 30 days per 1,000 eligible respondent encounters, not clicks or sign-ups. Target at least 2× control, with a statistically credible lift and no material deterioration in next-month retention.

Track form-completion rate, privacy complaints, Pro conversion, and badge-removal behavior as guardrails. Validate the pilot’s reported 30-day activation on mature cohorts; a two-week pilot alone cannot establish that outcome for every participant.

Finally, follow acquired creators’ subsequent publishing and respondent reach. Declare compounding only when mature production cohorts show that returning creators plus loop-generated creators consistently exceed the starting base without increased external acquisition.

Grades and run details

Decision model 82 · LLM judge 11 of 12 checks
Decision model checks
  • failUses the supplied evidence correctly51%
  • passAddresses the actual decision100%
  • passRespects explicit constraints32%
  • passIdentifies material uncertainty99%
  • passAvoids unsupported claims31%
  • passProduces the required deliverable75%
  • passCalls out the paid maths97%
  • passPicks the lever with the most yield95%
  • passA closed loop, not a channel98%
  • partialThe loop maths holds8%
  • partialProposes tests that could fail92%
Run
Run
#1
Time to output
48 s
Submitted
2 Oct 2026

3 of 11 checks passed

Output

MEMO

To: Sam Okoro, Head of Growth From: Product Management Date: Current Subject: Engineering Allocation for Next Quarter: Doubling Down on Our Primary Growth Loop

---

Executive Summary

Our primary growth engine is our product-led viral loop (forms $\to$ respondents $\to$ badge clicks $\to$ creators $\to$ new forms). While powerful, it is currently amplifying, not self-compounding ($K < 1.0$).

Both engineers should be dedicated to Idea 4 (“Make a form like this”). The pilot proved a 143% lift in publishing creators from our highest-volume, highest-retention channel. Conversely, Paid Search is deeply unprofitable, SEO templates are editorial rather than engineering work, and referrals solve an acquisition problem we can address organically for free.

---

1. Our Primary Growth Loop & Why It’s Not Yet Compounding

Our primary growth loop is driven by the form badge:

$$\text{Active Creators} \longrightarrow \text{Published Forms} \longrightarrow \text{Respondents} \longrightarrow \text{Badge Clicks} \longrightarrow \text{Sign-ups} \longrightarrow \text{Publishing Creators}$$

Current Monthly Metrics (Badge Loop): * 4.8M respondents ($40\text{k forms} \times 120\text{ respondents}$) * 43,200 clicks ($0.9\%$) $\to$ 4,752 sign-ups ($11\%$) $\to$ 1,806 new publishing creators ($38\%$) * Retention: Badge creators have our highest month-two retention (85%, vs. 71% for SEO and 64% for Paid).

Is it compounding? No—not in the runaway sense. * Each active creator publishes $\approx 2.22$ forms, reaching 267 respondents. * At baseline conversion rates, one publishing creator generates only 0.10 new publishing creators per cycle ($267 \times 0.009 \times 0.11 \times 0.38$). Even accounting for creator lifetime retention, the viral factor ($K$) remains well below $1.0$. * The loop is an amplifier of other top-of-funnel acquisition, but it cannot sustain compounding growth on its own without steady top-of-funnel inputs. To move closer to a true compounding state, we must dramatically increase loop velocity and cycle conversion.

---

2. Evaluation of the Four Ideas

IdeaStrategic FitUnit Economics / ImpactRecommendation
:---:---:---:---
1. Double Paid SearchPoorSeverely negative LTV:CAC. Pro LTV is \$280 ($20 \times 14\text{ mo}$). Value per sign-up is \$5.29 ($21\% \text{ publish} \times 9\% \text{ upgrade} \times \$280$). We pay \$38.00 per sign-up, losing \$32.71 on every acquisition.Kill immediately. No eng needed.
2. 200 Template PagesModerate14% publish rate; content-writing task, not an engineering initiative. Programmatic SEO would require eng, but manual templates do not.Pass to Content/Marketing.
3. Referral ProgramLowHigh friction; asks creators to solicit peers for \$10 credit when 4.8M natural form-users already experience the product directly each month.Deprioritize.
4. "Make a form like this"ExceptionalIncreases badge CTR ($0.9\% \to 1.6\%$) and publish rate ($38\% \to 52\%$). Generates 2,587 incremental publishing creators/month.Full Eng Commitment.

---

3. Engineering Allocation: 100% to Idea 4

We should allocate both engineers to productionize and optimize “Make a form like this.”

This is an architectural and UX challenge, not a copy change: 1. Sanitized Form Duplication: Automatically clone questions, logic, and layout while stripping respondent data, webhooks, and private settings. 2. Instant Preview & Onboarding Editor: Drop users directly into a pre-populated editor session before forcing account creation to preserve the pilot's 52% activation rate.

Expected Impact: * Monthly badge publishing creators will increase from 1,806 to 4,393 (+143%). * Revenue impact: 2,587 incremental publishers $\times 9\%$ Pro conversion $\times \$280\text{ LTV} =$ \$65,190 in incremental LTV added per month ($\approx \$780\text{k annualized}$), with zero marginal ad spend. * It raises our loop factor ($K$) from $\approx 0.10$ to $\approx 0.24$ per cycle, dramatically accelerating organic compounding.

---

4. How We'll Know It Worked

We will run a 50/50 randomized split test across all active forms.

#### Primary Success Metric * New 30-Day Publishing Creators generated per 100,000 respondents (Baseline: 37.6 $\to$ Target: $\ge 91.5$).

#### Guardrail & Funnel Metrics * Badge Click-Through Rate: Maintain pilot performance at $\ge 1.6\%$. * Sign-up Conversion: Maintain baseline at $\ge 11\%$. * 30-Day Publish Rate: Maintain pilot performance at $\ge 52\%$. * Form Quality Guardrail: Ensure month-two retention for cloned-form creators matches or exceeds the 85% badge baseline. * Pro Conversion Rate: Confirm downstream Pro upgrade rate stays constant at $\approx 9\%$.

Grades and run details

Decision model 50
Decision model checks
  • failUses the supplied evidence correctly57%
  • passAddresses the actual decision87%
  • partialRespects explicit constraints22%
  • partialIdentifies material uncertainty39%
  • failAvoids unsupported claims35%
  • passProduces the required deliverable58%
  • partialCalls out the paid maths68%
  • partialPicks the lever with the most yield82%
  • passA closed loop, not a channel85%
  • failThe loop maths holds34%
  • partialProposes tests that could fail87%
Run
Run
#1
API response time
32 s
Submitted
2 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 8

Uses the supplied evidence correctlyMixedMixedWrong
Opus 5.5 · Claude

The claim that the base settles at 10,700 creators is unsupported by the pack's arithmetic, and the claim that cost per sign-up usually rises with spend is not in the supplied evidence.

GPT-6 Astra · ChatGPT

All factual claims about the current situation are directly supported by the brief or derived by correct arithmetic.

Gemini 3.8 Flash · API

No reason given.

Respects explicit constraintsRightRightMixed
Opus 5.5 · Claude

Memo is addressed to Sam Okoro, stays within ~800 words, and proposes enforcement mechanisms (holdout, controls).

GPT-6 Astra · ChatGPT

Memo is addressed to Sam, from Product, under 800 words, and covers the required topics.

Gemini 3.8 Flash · API

No reason given.

Identifies material uncertaintyRightRightMixed
Opus 5.5 · Claude

Flags pilot's small size, two-week window, missing retention data, and representativeness, and resolves them with an 8-week holdout test and thresholds.

GPT-6 Astra · ChatGPT

Names badge exposure uncertainty, pilot size limitation, cohort maturity, and retention variability, and says to validate on mature cohorts and measure actual exposure.

Gemini 3.8 Flash · API

No reason given.

Avoids unsupported claimsWrongRightWrong
Opus 5.5 · Claude

Presents the 10,700 equilibrium and the rising cost-per-sign-up claim as facts without labelling them as hypotheses or supporting them from the pack.

GPT-6 Astra · ChatGPT

Forecasts and interpretations are labelled as directional or potential, and no confident claims go beyond the supplied evidence.

Gemini 3.8 Flash · API

No reason given.

Calls out the paid mathsRightRightMixed
Opus 5.5 · Claude

Calculates $181 cost per publishing creator and $25 revenue, states doubling would lose money, and recommends against it.

GPT-6 Astra · ChatGPT

Computes $181 per publishing creator and $25.20 revenue, and states plainly that doubling paid would lose money.

Gemini 3.8 Flash · API

No reason given.

Picks the lever with the most yieldRightRightMixed
Opus 5.5 · Claude

Chooses the copy-as-template badge, sizes its yield from the pilot (0.92 per 1,000 respondents, ~4,400 publishers, ~2.4× current), and notes the pilot's small size.

GPT-6 Astra · ChatGPT

Chooses the copy-as-template badge, sizes its yield from the pilot (0.915 per 1,000, ~2.4×), and flags the small two-week pilot before relying on it.

Gemini 3.8 Flash · API

No reason given.

The loop maths holdsRightRightWrong
Opus 5.5 · Claude

Correctly computes yield per creator (0.10/month, 0.46 lifetime), retention (4.5 months), amplification (1.8×), and break-even (0.83), and says the loop does not compound.

GPT-6 Astra · ChatGPT

Yield (0.10), retention (78%), and replacement (88%) are correctly computed; verdict is 'recirculates but not self-sustaining compounding'.

Gemini 3.8 Flash · API

No reason given.

Proposes tests that could failRightWrongMixed
Opus 5.5 · Claude

Proposes a holdout test with a numeric threshold (0.83 to ship, below 0.6 to keep but seek other levers), an 8-week window, and quality checks with thresholds.

GPT-6 Astra · ChatGPT

Proposes a test with a numeric threshold (2× control) but does not specify a measurement window or what action each outcome triggers.

Gemini 3.8 Flash · API

No reason given.

All got right 3

Addresses the actual decisionRightRightRight
Opus 5.5 · Claude

Commits early to putting both engineers on 'Make a form like this' with a holdout, says not to double paid search, and specifies what metric would change the call.

GPT-6 Astra · ChatGPT

Commits early to putting both engineers on 'Make a form like this', says how we'll know it worked (2× control threshold, guardrails), and states what would change the compounding declaration.

Gemini 3.8 Flash · API

No reason given.

Produces the required deliverableRightRightRight
Opus 5.5 · Claude

Delivers a complete memo to Sam with a clear recommendation, how to measure success, and is within the word limit.

GPT-6 Astra · ChatGPT

Delivers a complete, actionable memo within the word limit that Sam could use to assign engineers and measure success.

Gemini 3.8 Flash · API

No reason given.

A closed loop, not a channelRightRightRight
Opus 5.5 · Claude

Names the badge as the primary closed loop, grounds it in highest-retention creators, and sizes its yield with retention applied.

GPT-6 Astra · ChatGPT

Names the badge as the primary closed loop, grounded in highest retention (85%), and sizes its yield with retention applied (0.10 per creator, 88% replacement).

Gemini 3.8 Flash · API

No reason given.

Results

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

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1Sonnet 5.5withAPI87.396.22None
2GPT-6.1 SolwithAPI89.287.82None
3GPT-6 AstrawithChatGPT82.684.32None
4Opus 5.5withClaude87.164.42None
5GPT-6 LunawithAPI73.764.12None
6Gemini 3.8 FlashwithAPI64.676.92None
7Gemini 3.5 Flash-LitewithGemini40.953.82None

About the task

The PM job

Working out what actually drives growth, and where to push.

Why it matters

Teams tune funnel steps while the loop that compounds goes unmeasured. Mistaking a channel for a loop can cost a year.

What good looks like

  • A closed loop: each cycle's output feeds the next
  • The primary loop, traced from where the best users come from
  • The loop sized: cycle time, conversion, amplification
  • Retention in the maths
  • One lever, with a test that could fail

Deliberately not measured

  • Building a full growth model in a spreadsheet
  • Channel-level media planning
Capability tested

Growth systems thinking

The failure we’re looking for

Calls a channel a loop, or a referral button a viral loop

Grading

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