Add the project channel link

Point the reproducibility section at the now-published benchmark/

Publish DEV-RC-100: the 100 tasks, the scorer, the frozen manifest and the full per-item score report

Correct the headline metrics to the frozen release artifact (22% compile / 21% pass@1 / 0.713 / 93.2%)

Point the closing link at the KIR compiler repository

Ship training_args.bin so the reported hyperparameters are checkable

Rewrite as a technical report: work performed, results measured, direction taken; correct hyperparameters from training_args.bin

Tighten to results: drop the development-time scorer episode and the per-tier corpus breakdown

examples/: record per-version compile results for both training rows

Compile-verify both showcase rows against real Revit assemblies; state the EN row's measured 2024-2026 version lock

examples/: two complete training rows verbatim (42,685 and 17,147 chars) so the corpus scale is visible

Drop tier/weight annotations from the examples; link the full-size training rows

examples/: the 336-char KIR program and its full 193-line emitted C#, compile-verified 6/6 against real Revit assemblies

Strongest examples: EN rebar-forensics (refuses the brief, falsifiable anchor) + RU whole-building trunk; verifiable examples/ folder

Add the English gold example (grid-nudge unit-slip diagnosis) and state plainly that the gold was not the problem — the arithmetic was

Add the IFC pipeline: 214 real building models enriched with reasoning + a real RU record, and the RU/EN split

Add closing section: put the API in a compiler, not the weights — real KIR vs emitted C# (132 vs 3125 tokens), and the v2 direction

Precision pass: validation split was a deliberate choice (grade-the-curve early-stop), meta.json scope, honest reproducibility caveat

Rewrite card as a first-person account: full pipeline story + real dataset samples showing the 3%-gold / 40%-breadth gap

viveCADer_revit_9B: LoRA adapter (r=128) + compile-verified model card / honest negative result

initial commit