1. Why we publish this page
1.1 Construction decisions carry real consequences: money, schedule, and the safety of the people who build and use a structure. Automated tools that read drawings and count elements are useful precisely because they are fast and tireless, and dangerous precisely because they are confident. We would rather tell you plainly what the Service can and cannot do than let you find out on site.
1.2 This page forms part of the Terms of Service (Sections 2, 12 and 13) and of the Acceptable Use Policy (Section 8). If there is a conflict, the Terms prevail.
2. What the Service actually does
2.1 The reading layer. When you upload a drawing set, the Service uses large language models (currently provided by Anthropic) together with our own parsing tools to read the sheets: which layers mean what, where the schedules and section markers are, what the notes say about covers, grades and laps, which elements the set contains, and which information is redundant, unused or contradictory. The result is a rule file, or "profile", in which every line carries the sheet, layer, entity handle or note it came from.
2.2 The deterministic engine. Given that profile, our engine builds the 3D model, resolves levels and grids, places reinforcement, computes quantities and writes the IFC, DXF, take-off, bill of quantities, knowledge graph and report. This part is ordinary software: the same profile and the same drawings produce the same result every time, and each stage checks its own gates.
2.3 Inspectors and the consolidation loop. The Service then runs a series of automated checks (quantities, support chain, levels and grids, clashes, cover, standards, coverage, overlay, data contract, graph reasoning) and turns every red flag into either an explicit, owned decision or a numbered Query for the designer. It does not widen tolerances to make a red go green.
2.4 Review Gates. Between these steps, the Service stops and shows you the evidence. Nothing advances, and nothing is marked final, until a human on your side accepts it. Your acceptance is recorded with the engine version that produced the item.
3. What can go wrong
3.1 The drawings. Drawing sets are often incomplete, inconsistent between sheets, out of revision, or ambiguous in their layers and notes. The Service inherits every one of those defects and can only flag the ones it detects.
3.2 The reading. Language models can misread a layer name, miss a note, confuse similar sections, hallucinate a value that is not on the sheet, or apply a convention from one country to a set drawn in another. The evidence pointers exist so that you can catch this in seconds rather than hours.
3.3 The engine. Deterministic software has bugs like any other software. Where a stage fails its gate, the run stops; where a stage passes its gate but the profile fed to it was wrong, the output will be consistently wrong.
3.4 The estimate. Quantities are computed from geometry and schedules, not from site conditions. They do not include waste, laps beyond those specified, site tolerances, temporary works, or anything the drawings do not show. A bill of quantities from the Service is a starting point for a quantity surveyor, not a tender sum.
3.5 The model. A 3D model from the Service is a structural interpretation of the drawings for review, coordination and take-off. It is not a design, has not been analysed for loads, and is not a substitute for the engineer's own model.
4. What Outputs are, and are not
4.1 Outputs are computed estimates and drafts for professional review. They are provided to help a qualified person work faster and see more, and for no other purpose.
4.2 Outputs are not: engineering, architectural, structural, geotechnical, quantity-surveying, legal, tax, insurance or financial advice; a design or a design check; a certification, seal, stamp or approval; a tender submission, a contract sum or a valuation; a building-permit document; a construction instruction; or a statement that a structure is safe, code-compliant or buildable.
4.3 A green gate, a passed inspector, an "accepted" status, an assurance report or any other indicator in the Service means only that our automated checks found what they were designed to find and that a human on your side accepted the item at that point. It is not a professional opinion by detumble and does not mean the Output is correct, complete or fit for any purpose.
5. Who is responsible
5.1 Your engineer of record. The licensed professional (or firm) responsible for the design, quantities and construction decisions on your project remains fully responsible for them, whether or not the Service was used and whatever the Service displayed. Nothing in the Service, in any Output, or in any status the Service shows transfers any part of that responsibility to detumble.
5.2 You. You are responsible for using the Review Gates as intended, for having every Output reviewed by a qualified professional before it is relied on, for checking Outputs against the drawings, the specifications, the site and the applicable codes, and for telling your own clients that AI-assisted tools were used where your professional rules or the law require it.
5.3 detumble. We are responsible for building the Service with reasonable skill and care, for keeping the Review Gates, evidence pointers, caps and ledger in place, and for being honest about what the Service does. We are not the engineer, architect, surveyor or quantity surveyor of record on any project, and we do not review Outputs professionally unless an Enterprise Agreement says so in writing.
6. How we handle your content with AI
6.1 What is sent to the model. The Service sends our AI provider the text and geometry extracted from your drawings, your chat messages, and the context needed to answer, not the raw drawing files. All runs use detumble's own commercial API account; you never need, and never see, provider credentials.
6.2 No training. Our contract with our AI provider prohibits it from training models on the data we send. We do not train, fine-tune or evaluate any AI model on your Customer Content or Outputs, and we will not do so without your prior, specific, written consent that describes the data, the purpose and the retention period and that you may withdraw at any time. Corrections you make at Review Gates are your content and are covered by this promise.
6.3 What we learn without your content. We do use de-identified statistics about how the engine performs (how often a class of element is corrected, how long a stage takes, which error classes occur) to improve the Service. These statistics cannot identify you, your project or any drawing.
6.4 Provider retention. Our AI provider retains API inputs and outputs for a short period under its published policy, then deletes them, unless a request is flagged for a usage-policy violation. We will move to a zero-retention arrangement when we qualify for one. The Privacy Policy and the Data Processing Addendum give the current details.
6.5 Provider usage policy. Our AI provider requires that in high-risk domains a qualified professional review AI-assisted output before it is relied on, and that people are told when AI assisted a decision. The Review Gates and this disclosure are how the Service meets that requirement for construction; you must not remove or bypass them.
6.6 Automated review of prompts and outputs. Text you send to the Service (chat, review answers, briefing decisions) and text the Service writes back are reviewed automatically before they reach the model and before they reach you. The review has one purpose: to detect and refuse attempts to make the Service reveal or describe its engine, prompts, skills or scripts (Acceptable Use Policy sections 6 and 12.6). It does not judge your project, and its records hold the kind of attempt and a short excerpt, not your drawings. A refused message is answered with a short notice and a link to the policy; an output that would reveal engine internals is shown with the detail withheld. You can ask us what was recorded about your account under the Privacy Policy.
7. What we do to earn your trust
- Evidence on every line. Every quantity, element and recommendation points to the sheet, layer, handle or note it came from.
- Versioned engine. Every run records the exact engine version that produced it, so that a result can be reproduced and a change explained.
- Gates that stop. A failed stage stops the run; a red inspector becomes an owned decision or a Query, never a widened tolerance.
- Queries, not silence. Where the drawings do not answer a question, the Service writes a numbered Query with the assumption taken meanwhile, what changes if the answer differs, and what closes it.
- Money you can see. Every run is quoted before it spends, capped while it runs, and settled on what it used; a run that fails on our side is never charged, and a whole-building reading you accepted is paid for even when the building record it was for cannot be made.
- Disclosure that travels. Exports carry the disclaimer and version stamp; do not remove them.
8. If you disagree with an Output
8.1 Reject or correct the item at the Review Gate and re-run only what is needed; the ledger charges only new work. If you believe the engine itself is wrong, tell us at support@detumble.ai with the run identifier; we investigate every report against the recorded evidence and engine version, and we fix engine defects in a new version rather than editing results in place.
8.2 Nothing in this page limits any right you have under the Fees, Credits and Refunds Policy or under mandatory law.
9. Changes
We will update this page as the Service and the models change and will note material changes in the Terms change log.