KIR

A building is a program.
Typed, verifiable IR for Autodesk Revit — Python brains, verified hands.

Install · Why · The idea · Code · History · Measured results · Invariants · Examples · Repository state

---

The same building, held two ways. As the verified C# that KIR emits for a single Revit version, the 60-storey tower from examples/tower_numpy.py weighs 3 902 141 characters — roughly 1M tokens, beyond any model's context window. As the typed KIR program a model actually edits, it weighs 15 508 characters — roughly 4k tokens, 252× less. That is the difference between a model that can re-plan a floor or bend a facade with every change verified, and a model that cannot even read the building it is asked to change. (Sizes re-measured 2026-08-31 by running the example in this repo — the KIR side has not moved since 2026-07-28, the emitted C# has grown; the render is a visual companion, not the measurement.)

Install and run it

KIR is a Python package. Nothing below needs Revit, a licence, a service to stand up, or sudo — only Python 3.12 or newer. From a clone of this repository:

python3.12 -m venv .venv && . .venv/bin/activate
pip install .

python -c "from kir import spec, sdk; print(len(spec.OPS), len(sdk.builders()))"
python examples/tower_numpy.py

The python -c line prints 82 82 — one generated builder per registered op. The example after it builds a 60-storey twisted tower and compiles it for every supported Revit version, offline:

100 строк питона → 6 опов написано → 840 после экспансии → 780 элементов (программ KIR: 3)
этажей 60, талия 30%, закрутка 120°
ломаная против синуса: 217 мм по радиусу
компиляция: 6/6 версий ['2021', '2022', '2023', '2024', '2025', '2026']

(Every command in this section was run verbatim on 2026-08-31 in a venv created from scratch: pip install . pulled four direct and fifteen total dependencies, exit 0, nothing outside the venv touched. The shipped examples currently print in Russian.)

Writing into a live model additionally needs a Revit 2021–2026 installation, its API reference assemblies, the .NET Roslyn compile service and a separately running bridge — none of which are part of this repository.

Why

Language models write code well. They do not write buildings well, and the reason is structural rather than a matter of training scale. A building does not fit in a context window — one of the models we measure against holds 90 758 elements. The work is stateful: a wall must exist before a door can be hosted in it, and that state lives inside an application, not in a file you can diff. It is re-entrant, because the same instruction issued twice must not produce two doors. And it is versioned six ways: the same intent compiles differently against Revit 2021 and Revit 2026.

So the practice today is to have the model write Revit C# directly, and the arithmetic of that practice is available to us. Across 85 374 production compile errors logged over seven weeks, two codes account for 48.4% — CS1061 (32.3%) and CS0117 (16.1%) — and both are the same failure: a member of the Revit API that does not exist. Another ~10% (CS0104/CS0012) are namespace and assembly-reference problems. Roughly 60% of all production compile errors are spent fighting the surface of an API, not describing a building.

KIR is an attempt to build it: the model concentrates on 3D, geometry and composition; units, transactions, API versions, hosts, witnesses and rollbacks belong to a compiler.

The idea in one picture

flowchart TB
    subgraph FWD["FORWARD — intent becomes a building"]
        direction LR
        SDK["Python SDK<br/>numpy · shapely · the model"]
        PROG["Typed program<br/>KIR JSON, ops + refs"]
        REG["Registry<br/>one source of truth:<br/>schema · grammar · docs · emitters"]
        GRND["ground<br/>symbolic selectors to ElementIds"]
        PLAN["typecheck + plan<br/>mm only · DAG · txn partitions"]
        EMIT["emit<br/>per-version C#"]
        GATE["Roslyn gate<br/>2021–2026 · 1056 checks PASS"]
        RUN["live Revit<br/>one TransactionGroup"]
        SDK --> PROG --> REG --> GRND --> PLAN --> EMIT --> GATE --> RUN
    end

    RUN --> WIT["witness<br/>read the RESULT back, not the call"]
    WIT --> OUT{"exactly two<br/>typed outcomes"}
    OUT -->|ok| OKN["geometry_ok · semantic_ok · topology_ok"]
    OUT -->|refused| REFN["diagnostics + candidates + route<br/>KIR-G/T/L/E/C/X/W codes"]
    REFN -.->|"IR-level repair, max 3 rounds"| PROG

    subgraph REV["REVERSE — a building becomes a program"]
        direction LR
        DOC["live model"]
        L0["extract to L0<br/>48 categories · full-model census"]
        L1["lift to typed ops<br/>or a typed ATOM with a reason"]
        FOLD["fold to canon<br/>template-canon/4 · fidelity-canon/1"]
        MAT["materialize<br/>chunked programs, host-atomic"]
        DOC --> L0 --> L1 --> FOLD --> MAT
    end

    RUN -.-> DOC
    MAT -.->|"rebuild · edit · diff"| PROG

Two properties matter more than the boxes.

Every call ends in exactly one of two typed outcomes — an ok carrying a read-back proof, or a machine-readable refused with candidates and a route to a fallback path. A silently-wrong answer is the single forbidden state; ok:true wrapping a nested error is a permanent regression case with its own test.

The IR runs in both directions. Anything the lifter cannot express becomes a typed atom with a reason code, never a silent drop. That reverse direction is what turns "we can build" into "we can edit what already exists".

A stage-by-stage account of both pipelines — the registry as one source of truth, isolation modes, the census, fold, rebuild verification — lives in the language contracts under kir/specs/ (SPEC_V1.md, the decompile spec) and the design specs under docs/laws/.

Repository layout

Everything here is the compiler and its instruments; the product that hosts them is a separate repository. Revit is one target of several, and the package is agnostic about which one it emits for.

• kir/ — the forward compiler: registry, typecheck, plan, per-version emitters, typed diagnostics, witness/acceptance machinery;
• kir/decompile/ — the reverse pipeline: extract, lift, fold, materialize, and the version-control half (merkle, journal, rebuild, merge3);
• kir/checker/, kir/clash/ — the judge and the clash pipeline;
• kir/instruments/ — instruments that run offline, without Revit: compile_gate_offline, bounds_audit, unwired_census, scope_audit, snapshot_janitor, canon_state, and others;
• kir/course/, kir/skill.py — the authoring course and the skill an LLM is handed;
• kir/specs/ — the language contracts: SPEC_V1.md, the decompile spec;
• kir/bridge/ — the client side of the bridge protocol, plus a mock server for offline runs;
• kir/tests/ and kir/*/tests/ — 621 test files: unit, contract, golden and property tests;
• examples/ — small SDK programs that compile offline, before any Revit is connected;
• docs/laws/ — the design specs behind the honesty, merkle and verified layers.

The Revit bridge itself, the .NET Roslyn compile service, the product's admin routes, runtime data, credentials and logs are intentionally not part of this repository.

Show me code

A real program, printed verbatim from the SDK:

{
  "ir_version": "1.0",
  "intent": "one bay: wall + door + window",
  "ops": [
    {"op": "create_wall", "id": "wall1",
     "p0_mm": [0, 0], "p1_mm": [6000, 0],
     "level": {"by": "name", "value": "Level 1"}, "height_mm": 3300},
    {"op": "create_door", "id": "door1",
     "host": {"by": "ref", "value": "wall1"},
     "offset_mm": 1200, "sill_mm": -100,
     "symbol": {"by": "name", "value": "0915 x 2134mm"},
     "mirrored": false, "hand_flipped": false, "facing_flipped": false}
  ]
}

Note what is not expressible. A door has no xyz: it is host + offset_mm along the host + sill_mm. "A window floating in the air" is not a case we validate — it is a sentence the language cannot form. Every length is millimetres; feet do not exist in the IR.

The Python surface is generated, not written: 82 builders are born from the registry at import time, one per registry op (len(sdk.builders()) == len(spec.OPS), re-checked 2026-08-31), so a signature cannot drift from the spec. The SDK adds no semantics of its own — it cannot express anything the registry lacks, and it cannot hide a refusal.

import numpy as np
from kir import sdk

p = sdk.program(intent="tower with a waist")
with p.stack(levels=20, h_mm=3600,
             transform=sdk.transform(scale_xy_top=[0.8, 0.8],
                                     twist_deg_total=24)) as floor:
    for a in np.linspace(0, 2 * np.pi, 12, endpoint=False):
        floor.add(sdk.create_column(xy=[20000 * np.cos(a), 20000 * np.sin(a)],
                                    level=sdk.BY_MACRO, symbol="К 300x300"))

p.stats()   # {'ops_written': 1, 'ops_expanded': 260, 'elements': 240}

snap = {"column_symbols_structural": [{"id": 6011, "name": "К 300x300"}]}
out = p.compile(version="2023", snapshot=snap)
out.ok, len(out.csharp)   # (True, 1105378)

The snapshot is what a live Revit document would answer; offline you can hand the compiler the one pool this program grounds against. Get the name wrong and the refusal names the pool it looked in — KIR-G101 column_symbols_structural: «К 300x300» не найден — rather than failing somewhere later.

The shipped example goes further. In tower_numpy.py, numpy computes a sinusoidal waist and twist and KIR repeats the storey — run on 2026-07-28:

100 lines of Python -> 6 authored ops -> 840 after expansion -> 780 elements (3 KIR programs)
60 storeys, 30% waist, 120 deg total twist
piecewise vs. true sine: 217 mm along the radius
compilation: 6/6 versions ['2021','2022','2023','2024','2025','2026']

That third line is the point. stack.transform interpolates linearly; the requested curve is a sine. Saying "sine" and building a polyline without naming the divergence would be a silently wrong answer, so the example prints the error in millimetres.

Both demonstrations — the numpy tower and a shapely-cut curved floor plate — ship in examples/ with their measured outputs.

What happens to one op

sequenceDiagram
    autonumber
    participant M as Model — the LLM
    participant K as KIR compiler
    participant R as Revit — live document

    M->>K: create_wall — p0, p1, level by name, height 3300 mm
    K->>K: parse · typecheck · mm to feet · plan the DAG
    K->>R: snapshot query — resolve selectors
    R-->>K: level id, wall type pool

    alt selector missing or ambiguous
        K-->>M: refused KIR-G102 + candidates + route
    else grounded
        K->>K: emit C# for THIS Revit version
        K->>R: TransactionGroup — Wall.Create, op_id stamped
        R-->>K: new element id
        K->>R: read the element BACK
        R-->>K: LocationCurve, height param, level id
        alt postconditions hold
            K-->>M: ok + witness — geometry_ok, semantic_ok, topology_ok
        else violated
            K->>R: RollBack
            K-->>M: refused KIR-X004 + which axis failed
        end
    end

That stamp is why a retry is idempotent: op ids are written into the model inside the same transaction, so re-running a program skips what is already stamped, and resume-from-op-K is simply "skip what carries a stamp".

A building has a history

A single write is not the whole story. A real project gets read, edited, and read again dozens of times, and for most of this project's life nothing here remembered that revision N and revision N+1 were the same building — every rebuild started from zero. Four modules close that gap, and together they are exactly a version-control system, mapped onto concepts a reader already knows:

git concept	KIR module	what it actually is
content-addressed object store	merkle	a Merkle DAG over the folded building tree — a repeated floor is one node, not N
a history you can replay	journal	append-only per-building revision log; replaying it reconstructs any past revision exactly
applying a diff	rebuild	materializes only the delta between two revisions instead of the whole building
three-way merge	merge3	base / current document / target revision, with a real conflicting verdict — never a silent overwrite

It does not replace the addressed building graph shown above. The merged state is deliberately identity-free — a canonical op carries no ElementId — which is exactly what makes it mergeable; a program is still materialized from one concrete revision, never from the merge itself. The routes that drive these four live against an open Revit document (/admin/kir/rebuild, /admin/kir/decompile) are part of the private product and are not in this open-core slice, but their offline half ships and runs right here, no bridge and no Revit needed:

pytest kir/decompile/tests/test_merkle.py kir/decompile/tests/test_rebuild.py \
       kir/decompile/tests/test_journal.py kir/decompile/tests/test_merge3.py
98 passed, 19 subtests passed   -- run 2026-08-31, Python 3.12, this repository, no Revit

pytest is not pulled in by pip install . — it arrives with the dev extra; see Repository state for the one line that installs it.

Measured results

Everything below is derived from an instrument on the date shown, not from memory.

Every row below was produced by running the command in the right-hand column, on commit 3211555. Rows marked historical are older runs whose instrument or corpus is not shipped here; they carry their date and cannot be re-run from this repository alone.

Fact	Value	Date / how
Ops in the registry	82 — 77 writing, 5 query	2026-08-31, len(spec.OPS)
Generated SDK builders	82 — one per op	2026-08-31, len(sdk.builders()) == len(spec.OPS)
Source size	284 production modules, 634 test files	2026-09-01, python3.12 tools/readme_numbers.py --check
Python syntax check	947 files parsed, 0 errors	2026-09-01, ast.parse over every *.py outside build/
Install into a clean venv	PASS — exit 0, 15 packages, no sudo	2026-08-31, pip install . in a fresh python3.12 -m venv
Offline six-version compile of the shipped example	PASS — 3 programs, 840 ops, 780 elements, 6/6 versions	2026-08-31, python examples/tower_numpy.py
Version-control half, offline	98 passed, 19 subtests, under 10 s	2026-08-31, pytest kir/decompile/tests/test_{merkle,rebuild,journal,merge3}.py
Emitted C# for the 60-storey tower, one version	3 902 141 characters against 15 508 for the KIR program — 252×	2026-08-31, Revit 2023
Historical live baseline	31 of 31 writing ops had a witnessed run — of the 31 that existed then	2026-07-28 local telemetry; telemetry is not shipped here
Historical six-version compile gate	1 056 Roslyn compilations, PASS	2026-07-28 local gate run
Historical reverse-direction coverage	48 categories; 92.83% on the 90 758-element R2026 model	2026-07-27/28 local runs
Historical production compile errors structurally inexpressible in KIR	≈60% of 85 374 over seven weeks	local report; completeness caveats apply

Checked invariants

The reverse pipeline enforces five checked invariants: full document coverage, a reason recorded for every element it cannot express, witnesses that match the data actually read, contamination marking for partial reads, and neutral identifiers. Each invariant is backed by a test or verification run; a violation stops the build or run.

flowchart LR
    subgraph LAWS["§18 — five invariants"]
        direction TB
        L1["1 · CENSUS<br/>lifted + atoms + not_read<br/>= the whole document"]
        L2["2 · RECEIPT<br/>every cut emits<br/>element_id + typed_reason"]
        L3["3 · WITNESS AXIS<br/>sign only the axis<br/>you actually read"]
        L4["4 · CONTAMINATION<br/>a partial read marks<br/>everything derived from it"]
        L5["5 · NEUTRALITY<br/>no device ids, install paths,<br/>or locale-only keys"]
    end

    subgraph CORE["what they enclose"]
        direction TB
        C1["compile"] --> C2["execute in Revit"] --> C3["read back"] --> C4["publish a number"]
    end

    L1 -->|"run-fatal identity check + CI fixture"| CORE
    L2 -->|"rows + failures contract on every side stage"| CORE
    L3 -->|"certificate lint: kind to legal reader"| CORE
    L4 -->|"header round-trip test + A5 gate on partial data"| CORE
    L5 -->|"CI pattern lint"| CORE

    CORE --> R1["built, with a witness"]
    CORE --> R2["typed refusal, with a route"]
    CORE -.->|forbidden by construction| R3["silently wrong"]

Full account: Five Conservation Laws of Honesty for an AI Agent in CAD.

Read more

• A Building Is a Program — the long-form technical case: what the IR expresses, why it runs in both directions, and the measured number behind each claim.
• Five Conservation Laws of Honesty for an AI Agent in CAD — the five laws, the catch behind each, and how each one is mechanically enforced.
• A Day of Measured Revit API Traps — fourteen Revit API behaviours as symptom, wrong hypothesis, measurement, rule. Useful even if you never touch KIR.
• One Day Inside an AI-led Compiler Team — a first-person account of one working day on this project.

Repository state

The compiler, the reverse pipeline, the instruments, the authoring course and the whole test suite live in this repository and run without Revit. To work against the tree instead of an installed copy:

python3.12 -m venv .venv && . .venv/bin/activate
pip install -e '.[dev]'
PYTHONPATH="$PWD" python -m pytest \
    kir/decompile/tests/test_merkle.py kir/decompile/tests/test_rebuild.py \
    kir/decompile/tests/test_journal.py kir/decompile/tests/test_merge3.py -q

Three traps worth naming, all bought by measurement rather than reasoning:

• Name the four files; do not point pytest at the directory. kir/decompile/tests collects 171 files, and one of them allocates about 5 GB and takes the rest of the run down with it. The four named above are the ones the number in the table refers to.

• PYTHONPATH must be absolute. A bare PYTHONPATH=. is shadowed by any stray kir.py beside you, and the refusal then blames the wrong module.

• Four direct dependencies, and httpx is one of them. It looks like bridge-only machinery, but compile_client.py sits in the package import chain; dropping it into an extra made import kir fail in a clean venv. The dependency list is settled by a run, not by an argument.

Live execution additionally needs a Revit 2021–2026 installation, its API reference assemblies, the .NET Roslyn compile service and a separately running bridge.

Machine-specific paths: measured, not asserted. An AST sweep over every string literal outside docstrings finds 0 absolute machine paths in the 274 production modules. The test suite is a different matter and is named rather than glossed: 17 test files pin absolute paths of the machine they were measured on, and the tests ship inside the built package. No credentials, keys or device identifiers were found anywhere in the wheel.

For auditors

Beyond the compiler core in kir/:

• Language contracts: kir/specs/ — SPEC_V1.md, KIR_DECOMPILE_SPEC.md
• Laws of honesty / verified / merkle: docs/laws/
• LLM authoring skill + sandbox course: kir/skill.py, kir/course/
• CI evidence & secret boundary: .github/workflows/kir-evidence.yml, kir-security.yml
• Offline instruments, runnable with no Revit: kir/instruments/ — each one is expected to name what it could not read rather than report a zero it never measured

---

_{Apache License 2.0 — see LICENSE.}