The AI lives by three rules · skills · audit · provenance all serve them. No clever exception. No "AI helpfulness" workaround. The rules are the product.
If an engineering value isn't in the inputs, the AI doesn't make one up. Ever. It writes [ENGINEER TO CONFIRM] and moves on. 0 invented values per Demo A run is not aspirational — it's the contract.
Every interpretive call · every value the AI can't ground in source · every cross-discipline handoff · gets surfaced for the engineer to confirm before signing. Doubt is loud. Confidence requires source.
No "the investigation revealed" · no "notably" · no praise for the firm's own diligence. Engineering reports are evidence chains. The AI writes like a senior engineer, not a marketing intern.
Single-tenant deployment in the firm's own Azure tenant. No data leaves. Compliance-friendly by construction · not by review.
Cowork agents · Tier-A sources · Australian primary law / standards / DA filings · every finding cited. Below: the missions that changed the product.
16-row voice-pattern table from Douglas Partners J1 exemplar · derived 4 skills (calibrated-commitment · hedge-ladder · scope-boundaries · structure-interpretive).
21 verified AU geotech cases · 12 documented failure patterns (Thredbo · Woolcock · Bryan v Maloney · Mascot · Opal Tower).
14-item self-audit + 10 hard disqualifiers · QLD/NSW/WA convergence map · standards review-readiness criteria.
Only 6 of 16 councils are prescriptive · Toowoomba prohibits limiting clauses (collision with skill 05).
Same engineer / project · two registers (council vs sworn evidence) extracted side-by-side from Watercare Herne Bay project.
Hedge ladder vanishes in design lane · deontic mood replaces epistemic · 8 patterns transfer · 6 modify · 2 absent.
One council manual overrides AS verbatim · RAG cannot be national-only · 3-scope retrieval required.
AS 4067 doesn't exist · superseded standards still cited · year missing on 1 in 4. THIS is a product feature, not internal hygiene.
Built-in. Out of the box. Customers can use as-is · upload their own (template scanner extracts schema) · or contract us to build new ones. Three revenue lines from one engine.
Trust isn't something we tell prospects. It's something we build into every layer of the pipeline. By the time a report reaches the engineer's desk, six independent mechanisms have asked "is this honest?".
Every value the AI writes is registered against one of five source kinds: input · lookup · computed · boilerplate · engineer_to_confirm. A trail per report.
Any interpretive call · any specific value the AI can't ground · highlighted in yellow in the .docx · counted on the run record · audit blocks ship if missed.
18 items run after every draft · 10 hard disqualifiers · IDE-style modal if any trip · state × discipline aware (won't false-block WA civil for missing RPEQ).
Auto-scans every section for stale years · missing editions · standards that don't exist (AS 4067) · suggests corrections. Differentiator no generic AI has.
Database-enforced minimum 2 distinct reviewers per report · no exceptions · prevents the "engineer signs their own work" failure pattern.
Skill 02 blocks the model from emitting numbers without input source · skill 03 forces AS 2870 class to be flagged not computed · skill 05 locks limitations boilerplate · 41 skills in total.
A single day's commits (2026-06-08). Each one a discrete improvement to one of the six layers above.
Built on the assumption that engineers reading the output have professional indemnity exposure · disciplinary frameworks · and a duty to court. The product can fit nowhere else but inside a profession that takes evidence seriously. That's the moat.