Firm settings

Business profile v3 · history

Firm-wide voice applied to every report type. Changes create a new version — each report records the version it was generated with.

Tone & voice

Terminology 13 pairs · learned from your reports

non-engineered fill→uncontrolled fill water table→groundwater level recommend that→recommend it is considered→we consider

Boilerplate library inserted verbatim — never AI-paraphrased

Limitations (standard)used in every report
Report reliance & third partiescommercial reports
Groundwater disclaimerwhen < 2 monitoring rounds

Risk posture

Flag assumption when groundwater within 2.0 m of founding depthtrigger
Always recommend additional sampling when QA rules flag index anomaliestrigger
Site classification: adopt more reactive class when borderlineconservatism

Report types

Each type trains separately from its own example reports — structure, per-section guidance and exemplars. Firm voice applies on top.

DSI Interpretive 3 examples · ready

9 sections · per-section guidance ✓ · 14 reports drafted with this type · avg consensus 4.2

Pavement Investigation 1 example · usable

7 sections · CBR/subgrade emphasis · guidance generic in places — add examples to sharpen

Contamination Assessment low confidence — 1 example

Drafting enabled but flagged · add 2 more examples to train reliably

Knowledge base

Your firm's corpus — retrieved per section while drafting, filtered by report type. Signed-off Auger reports join automatically (tagged by provenance).

Milton Depot DSI Rev CDSIsigned off in Auger · human-edited · high weight
Kelvin Grove Residential Stage 1DSIuploaded · human-authored
Toowong Commercial TowerDSIuploaded · human-authored
Gold Coast Pavement InvestigationPavementuploaded · human-authored
QLD residual soils guidance noteguidanceinternal memo · 2024
Deleting a document purges its embeddings immediately.

Rubric & judges

Independent judges score every section against these criteria — no debate, dissent surfaced to reviewers. Dry-run any change against past reports before committing.

Rubric v2 target consensus ≥ 3.5

Data fidelity — claims match cited data×1.8
Standards compliance×1.5
Conservatism of interpretation×1.5
Completeness vs section guidance×1.2
Firm voice & terminology×1.0

Model panel

ClaudeDrafter+ self-critique pass
ClaudeJudge 1
GPTJudge 2
GeminiJudge 3
Judges score independently and never rewrite. Panel composition is config — reduce to one judge if your validation data supports it.

Team

Sign-off authority is role-based; every sign-off is stamped to the audit trail.

CR
Chris Rchris@geotechpartnersAdmin · CPEngcan sign off
DV
Declan Vanderhordeclan@geotechpartnersSenior reviewer · RPEQcan sign off
JM
Jess Mjess@geotechpartnersEngineerdraft & edit