A manifesto went up at typesafe.ai this month under a line that repudiates three years of industry noise: “We’re building prod, not God.” TypeSafe — founded by Diogo Almeida, a contributor to the original ChatGPT work, per The Rundown AI — argues the bottleneck in applied AI is not raw intelligence but that today’s intelligence is hard to build on. Its first model, Jev, doesn’t write prose: it answers typed questions against a state and returns a value, a probability distribution and a confidence figure. MarkTechPost put the price at $0.042 per 1 million input tokens; Tom’s Hardware carried the vendor claim of 193× faster and 445× cheaper than an LLM call for the same class of judgement. Vendor numbers earn vendor scepticism — the shape of the claim is what matters.

The TypeSafe docs name three primitives: Choice (pick from a list), Score (rate against a rubric) and Noul (is this statement true, 0–1). Every question is evaluated in parallel and in isolation against the same state, so a fourth question barely moves response time and does not poison the others with context rot. The guidance that follows is the interesting part: decompose. Instead of “rate this tender submission”, ask separately about completeness, pricing plausibility and programme risk, then combine the scores with your own coefficients in your own code.

←TODAY: In 2026 most “AI in BIM” still means a human reading a chat window and re-typing the answer into Archicad or Rhino. →3012: Semantic judgement is a callable primitive with an error bar, buried five layers deep in a checking pipeline nobody thinks about. Fulcrum: Only a judgement you can audit, version and constrain is one you are allowed to automate — the argument openBIM made about file formats thirty years earlier.

The whip socket in your coordination workflow

The manifesto’s central image is the 1903 Ford Model T as a horseless carriage — high seats, buggy springs, a whip socket bolted to a machine with no use for one. Models trained with RLHF to be helpful and articulate make sense if a human sits on the other side; the foreseeable consequence, TypeSafe argues, is AI that requires a human in the loop instead of running in the background. AEC has its own whip socket. Watch a BIM coordinator today: paste an IFC property set into a chat, read a paragraph, decide, re-type the decision into the coordination tool. Every loop is a human acting as a parser. The work is not the judgement — it is the transcription around it.

Composability is an argument our industry already won once. Our openBIM concept panel makes the same move about file formats: IFC 1.0 shipped in June 1996, IFC2x3 aligned to ISO/PAS 16739 in 2006, and IFC4.3 Add2 stretched the schema into rail, roads and bridges in 2023 — stewarded throughout by buildingSMART International. A capability becomes infrastructure only once it is dependable enough to bury in someone else’s dependency tree.

The trade-off, plainly: a typed decision hands you a number you can branch on, and a number invites you to stop reading the reasoning. A confidence of 0.87 on “this wall type conflicts with the fire compartment strategy” is more actionable than a paragraph — and easier to rubber-stamp. Calibration is a property of a distribution, not a promise about the case in front of you.

Atelier: For a twelve-person Zürich studio living with AI rather than demoing it, the shift is from prompting to plumbing, and that changes who owns the work: a prompt lives in one person’s chat history, a typed question with a rubric lives in the repository and can be argued about in a team meeting. Monday move: take the one recurring judgement your team already makes by hand every week — model-issue triage, tender completeness, whatever it actually is — write it as three atomic questions with explicit rubrics and commit that file to the office repo. Don’t call an API yet.

Where it lands on a Swiss desk

The application is unglamorous: does this BCF issue duplicate an existing one, does this room name map to the SIA classification the BEP demanded, does this datasheet meet the Minergie value it claims. None need extended reasoning; all currently need a person, because the answer has to come back in a form the next script can use. The discipline from PAZ’s digital identity architecture panel transfers exactly — keep the decision plane separate from the enforcement plane, verify per request, not per session.

Hack: Split one fuzzy office judgement into three atomic ones and keep the weighting in code your colleagues can diff. Sketch it as a pure function with the model call stubbed — the hard part was never the API.

W = {"completeness": 0.5, "price_plausibility": 0.3, "programme_risk": 0.2}
def triage(score, conf, floor=0.7):
    if min(conf.values()) < floor: return "HUMAN"   # low confidence never auto-routes
    return "PASS" if sum(score[k]*w for k, w in W.items()) >= 0.6 else "REVIEW"

Change 0.2 to 0.35 when programme risk starts hurting, and you have re-tuned your office’s judgement in a commit with your name on it.

A manifesto is also a fundraising document — TypeSafe’s own benchmark, 3% global TFP growth held for ten years, is a bar no small team controls. Whether typed judgement becomes infrastructure won’t be settled by the founders quoted in TechCrunch, but by the unnamed engineers who make a primitive reliable enough that someone else dares depend on it five layers down. Buy nothing this quarter on the strength of a 445× figure. Do write down the three judgements your office repeats most often, and see whether any survives contact with an explicit rubric.

Source: typesafe.ai