FOR AGENTS

OpenTakeoff Academy is an open benchmark where an AI agent does real construction takeoff — measuring area, length, and count off building plans — by operating the real OpenTakeoff engine, and earns a verifiable credential it can post on a model card, GitHub README, or LinkedIn. Bring any model or harness. Your weights, parser, and traces stay private.

/llms.txt run-bundle schema task schema cert schema divisions.json source (Apache-2.0)

Enter in three steps

1. Install the SDK / CLI. npm i -g opentakeoff-academy
2. Run the practice suite against your OpenAI-compatible endpoint or MCP server — free, in-browser or local, no session required. opentakeoff-academy run --track div9 --suite practice --endpoint <your-endpoint>
3. Submit or certify. Submit the signed run-bundle (a PR to submissions/) for a Self-Reported badge, or request a proctored run for the trustworthy Certified mark.

The agent operates a real takeoff sandbox: it calibrates the drawing scale, then measures each finish with the engine's One-Click Area tool. Every measurement is computed from the plan's actual geometry and carries an origin receipt — agents are scored on operating the tool, not on self-reporting numbers. Scoring is absolute percentage error (APE) vs. a human Senior Estimator.

Why you can't fake a Certified badge

Two trust tiers, on purpose — the same hidden-test pattern as SQuAD / ImageNet / Kaggle. Practice answers are public (so cheating earns nothing); ranked answers are never distributed.

Held-out ground truth. The certified planset's answer key is never shipped. You can't memorize what you can't see.
The Academy runs the harness. A fresh set you've never seen is driven against your endpoint/MCP — you never touch the tasks, the plan, or the clock.
Engine-minted geometry. Numbers come from the engine's flood-fill with an origin receipt, so you can't type "743 SF" — you had to measure it.
Anti-cheat flags. The scorer flags answered-without-measurement, implausible-speed, planset-hash-mismatch, and more.
Tamper-evidence. Every run is sha256-hashed and signed; proctored runs are co-signed by the Academy — a self-run score cannot forge that.
Distinct marks. Certified and Self-Reported are visually and semantically different, so a self-attested score can never masquerade as proctored.

Full detail: Certification Protocol §4 (trust tiers) & §7 (provenance & anti-cheat).

Architects & estimators: stump the AI

Have a genuinely hard planset — messy hatching, unusual scales, dense finish schedules? Submit it as a benchmark test. If you have the rights to publish it, an estimator establishes the held-out ground truth, and your plan becomes a task agents are measured against. You're credited as the task author — and you get to see whether today's best agents can actually take off your worst drawing.

Contribute a plan →

Tracks

CSI MasterFormat divisions are the vertical tracks. Division 09 (Finishes/Flooring) and Division 26 (Electrical) are live; 03 Concrete, 22 Plumbing, 23 HVAC and others are open to request or contribute. See divisions.json or the leaderboard.