What does TAM Desk compute that a spreadsheet cannot?

The disagreement. A spreadsheet holds one build; TAM Desk holds both and measures the distance between them. If your top-down slide says $28B and your segment table says $948M, it says so as a 30x gap, then solves for what would have to change: the blended ACV would have to be $921K instead of $31K, or there would have to be thirty times as many accounts. It then asks the question the model itself cannot: at your rep count, your quota and your win rate, how many years does the SOM actually take?

Do I have to sign in?

No. The parsing, both builds, the triangulation and its reconciling solve, the capacity test, the unit economics, the sensitivity sweep, the provenance ledger, the complete sizing memo, the sourcing checklist, the driver CSV, the measurement JSON, the comparison between two sizings and every check run in this tab with no account and nothing charged. Only the judgement pass is metered, and the three bundled examples replay a saved run for free.

What counts as a sourced number?

A line in the sources box that names the driver. A bare list of report titles proves nothing about which number came from which report, so it does not count. Every driver ends up marked sourced, stated or assumed, the counts are reported, and any value this engine supplied as a default is listed as an assumption in the same table as the number it produced.

What is checked about the AI pass?

Every measured driver must get exactly one review - a driver reviewed twice fails as loudly as one left out, so the partition is counted rather than sampled. Source verdicts are compared against the ledger, so calling a driver unsourced when a source line names it is a failure. Any figure the pass recommends is re-derived against the measurement and shown beside the current value. And a driver the engine itself measured as absent is never held against the pass, because the browser already reported that gap.

What does a run cost, and what if my balance is low?

A worst-case amount is reserved and only what the run actually uses is charged, and the amount is quoted beside the run button before you submit. Signed out, the button asks you to sign in rather than failing after you press it. Signed in with a balance below the model's minimum, the button stays disabled and names the shortfall with a top-up link. A reply cut short by a low balance says so rather than presenting itself as a complete review, and the sections that did arrive are kept.

Where does the model go afterwards?

Out of the browser. The memo downloads as markdown for the data room, the drivers, segments, triangulation, sensitivity rows and checks as a CSV for a spreadsheet, and the whole measurement as JSON. The sourcing checklist copies out as ready-made source lines, so the drivers nobody has traced yet can be filled in and pasted straight back. Sizings are saved with the assumptions that produced them, so the next re-size starts from the last one with the numbers back in the boxes, and any past sizing can be re-measured and compared against what is on screen now.

TAM Desk — assumptions in, a triangulated market-size model out

Stage — what the number has to survive

Measured as of — the date the model is a statement about

Anything to steer the review? — optional

The market — one key: value per line. A total market value and a target share are what the top-down build needs

The segments — CSV, TSV or a markdown table. Drop a .csv on the box

Go-to-market — what has to deliver the SOM

Sources — one driver: source per line

Everything above — both builds, the triangulation, the memo, the sourcing checklist, the CSV, the JSON and the comparison against a past sizing — is free and needs no account. Only the judgement pass calls the model. It reserves a worst-case amount and charges only what the run actually uses; the line beside the button quotes that amount before you submit. Signed out, the button asks you to sign in rather than failing after the fact; if a signed-in balance is below the model’s minimum the button stays disabled and names the shortfall with a top-up link, and a reply cut short by a low balance says so instead of presenting itself as complete.

Result

How it works

Nothing to hand? Load the — two builds that triangulate and a plan the sales team can carry — the , where the slide says $28B and the segment table says $948M, or the , credible but thin, with almost nothing sourced. All three replay a saved run for free. Or press to watch the checker name its findings in a memo with no model call at all.

A market size is arithmetic before it is a story

The number on a market slide is not wrong because the prose is weak. It is wrong because the total came from a press release, the filters were picked to land on a round figure, and the segment table underneath it was never multiplied out. So this app multiplies it out: the top-down build is a waterfall where every filter shows what it removed, and the bottom-up build is accounts times ACV per segment, narrowed by a reachable share and an attach rate. Both are computed here, for free, before anything is charged.

The useful number is the distance between the two builds

One method is an estimate; two that agree is a market size. TAM Desk measures the gap at TAM, SAM and SOM and calls it converged, soft or contradictory — and then does the part that makes the gap arguable rather than embarrassing: it solves for the single driver value that would close it. “Your blended ACV would have to be $921K instead of $31K” is a sentence a founder can answer. “These are 30x apart” is not.

A SOM is a claim about a sales team, so it is tested as one

The SOM is compared against the capacity that would have to produce it, computed two ways — reps times ramped quota, and reps times meetings times weeks times win rate times blended ACV — and the lower of the two is used, because a quota does not create pipeline. The result is a number of years against your horizon and a rep count that would close the gap. The same pass computes the LTV, the LTV/CAC ratio and the CAC payback the claim implies, ranks every driver by how far it moves the SOM, and records whether each number was sourced, merely stated, or supplied by this engine as a default.

The metered pass is judgement, and it is held to the measurement

What a model is for: whether a 47% segment filter is defensible or convenient, what source would settle a churn rate, which of the two builds is the one to fix, and the paragraph that carries the argument. It returns exactly one verdict per driver, and the app counts them — a driver reviewed twice fails as loudly as one left out. Its source verdicts are checked against the provenance ledger, every figure it recommends is re-derived against the measurement and shown beside the current value, and a driver the engine itself measured as absent is never counted against it.

A derived work of @sickn33/startup-analyst: its market-sizing method — TAM/SAM/SOM built bottom-up and top-down, validated by triangulation, with assumptions documented, conservative defaults and benchmark comparison — is what this app implements and measures. The skill is an agent persona that asks a model to do the arithmetic; this is the same method with the arithmetic computed and the model held to it.