89 of 103
Sealed held-out rooms, room number
Rows sealed before any work began and never read during development, against a range committed before the run.
Everyone claims accuracy. Here is ours.
Every gate below states the engine, the value, the date it was measured and the caveat.
- AS 4254.2
- AS 1668.2
- DW/143
- DW/144
- NCC
- ISO 19650
- IFC
Register
The damper register, with the basis on every row.
How it is built
Assets on plan
Assets on plan, ringed where the position disagrees.
How it is built
Drawing to schedule
One damper on a sheet, one schedule you can defend.
How it is built
Assistant
Ask which dampers hold no signed record, and read where it came from.
How it is built
Take-off
A layout crop, with every asset Archie read off the sheet.
How it is built
Scope gaps
Specification against the priced bill, with every gap traced.
How it is built
Illustrative photography. The screens are Archie's own interface on a demonstration project, photographed in use rather than reproduced at reading size. Every one of them is live and interactive on the platform page.
Numbers that carry their own working.
Tagged assets on a live rail station project
Every row carries its basis.
2,459
Of those, placed in a room
The other 4 percent are reported as withheld, with the reason.
96%
Numbers written by a language model
Every quantity comes from a deterministic engine.
0
Archie is not a tool your team uses. It is a member of your team.
Every gate, its value, and the date it was measured.
79 of 103
Sealed held-out rooms, full room code
Same sealed set, scored on the full room code rather than the number alone.
328 of 339
Open review set, full room code, after rulings applied
The visible development set, not the sealed one. Diagnostic accuracy is always the optimistic one.
0 of 148 raised
Unresolved position-versus-room flags at close-out
148 assets whose computed position sat in a different architectural room than the one assigned. Each ruling became a rule, and the count went to zero.
Within 0.06 percent
Independent clean-room reconstruction
Rebuilt twice from scratch by independent agents given only the raw inputs and the review notes, with a blindness audit.
220 of 220 across 10 documents
Full review round actioned, comment by comment
A chartered mechanical commissioning engineer's complete review. Every comment became a corrected rule or a permanent regression case.
191 rulings
Engineer and client rulings applied to the data
Each ruling is labelled on the row it changed, and a ruled row is excluded from every score.
99.5 percent recall, 97.4 percent precision
Identification across a full building level, against a real project model
This gate measures identification, finding and classifying the asset, not placing it in a room.
784 against 775, a 95.4 percent match
Cross-source count check, model against schedule
Across 25 level-and-category cells, cross-checked against an independently computed figure. This validates counts, not lengths.
Agreement to 0.004 percent
Two-codepath internal consistency, pipe length
Internal consistency between two independently coded aggregators, not correctness against an external source, because no external length source exists.
11 gaps of 25
Specification systems assessed against the priced bill, on a live tender
Every gap is traced to a specification section and page. A gap is an exposure to investigate, not automatically an unpriced item.
Measured on a live rail station project, with nothing retained afterwards.
How a sealed exam is run
- A fifth of the register sealed before work starts
- A scored run needs a milestone registered in advance
- Nothing changes until a fresh diagnostic gate
- Hypotheses are pre-registered, and failures stay recorded
- Every run logged with its code version and leak check
- A ruled row leaves every score
- Model output is never presented as human-reviewed
- Roadmap: a blind proof on a second project