Hand your interpretation to a pipeline of five specialist agents — Ingest & QC, Geometry, Structural Intelligence, Uncertainty, Review — and watch them measure, cross-examine, write it up and sign it off. The agents orchestrate; deterministic, published rules decide. Not a chatbot.
Each agent finishes its brief and hands a stricter object to the next. You watch the whole exchange, live.
Parses every stick, then puts the network on trial: orphan segments, kinematically impossible crossings, abrupt throw steps, one fault digitized as two — the errors reviewers find, found first.
hands over → a screened network + an interpretation-quality factorMeasures corrected throw on every fault–horizon pair, builds SGR at reservoir level, and resolves the regional stress tensor on every plane — slip tendency, dilation tendency, and the critical ΔPc.
hands over → profiles, seal columns, geomechanics per faultApplies the twelve published rules — relay classification and breach risk, throw partitioning, bed-parallel slip, SGR seal, juxtaposition, reactivation, dilation targeting — and files a finding for every hit, citation attached.
hands over → findings with claims, rules, citationsRuns Monte-Carlo over pick error, fault position and Vshale, stamps P10/P50/P90 on every metric, and assembles the deliverables — QC grid, cited PDF report or CO₂ dossier in four languages, share link.
hands over → the evidence pack your reviewer opensRe-checks the finished run against an explicit release rubric — citations on every finding, four-language parity, verified P10/P50/P90 ordering, cross-reference integrity — and runs a revise pass before sign-off. The fifth agent reviews the other four.
hands over → a signed-off run: rubric 5/5, in the log and on the PDFSlip and dilation tendency on every plane, and the number a CO₂ reviewer asks first: ΔPc — pore-pressure headroom before Coulomb failure. Hermerswiller: −1.1 MPa. Thebe Main: 7.4 MPa.
Two interpretations of the same volume, both re-read by the team and compared: faults added or removed, geometry deltas, findings that appeared or resolved.
Findings, live agent logs and PDFs render natively in English, French, German and Greek — not just the menus. Your reviewer reads evidence in their own language.
One click renders the structural screening a Directive 2009/31/EC permit expects — inventory, integrity QC, juxtaposition & membrane seal, reactivation — with stated scope limits.
Auto-detected interpretations welcome. Declare a project machine-built and the agents screen it for auto-picker failure modes — fragmentation, over-connection, orphan noise — with provenance carried through to the report.
The same screening, written the way your reader's regime expects it: EU Directive 2009/31/EC, UK NSTA, US EPA Class VI, or Greece's HEREMA — chosen at export, not re-run.
Paste what a report claims — throw, dip, SGR, ΔPc, counts — and every line comes back match, mismatch or unverifiable against the measured value, with the tolerance it was judged on.
Autopilot mints the reviewer's link the moment Agent 05 signs a run clean — and one zip carries the dossiers in four languages, the QC grid and the findings.
Invite the rest of the interpretation team, assign owner, admin or member, and share projects inside the account. Invites are single-use, bound to one address, and expire in fourteen days.
Runs open on a verdict — submission-ready or not, with the critical count and a named review path — then triage rows worst-first, each expanding to its claim, citation and P10/P50/P90.
| FAULT | CROSSING | SPLIT | RELAY | MIN SGR | ΔPc MPa |
|---|---|---|---|---|---|
| Soultz — Nord | ok | ok | relay | 0.13 | 1.6 |
| Kutzenhausen | × | split | ok | 0.13 | 3.0 |
| Hermerswiller | × | ok | ok | 0.13 | −1.1 |
“Resolved on Hermerswiller (strike 155°, dip 64°) at 2 577 m under a normal-faulting stress regime with SHmax 170°: slip tendency Ts = 0.63 and a critical pore-pressure increase of ΔPc = −1.1 MPa before Coulomb failure at μ = 0.60. ΔPc ≤ 0 means the plane sits at or beyond the Coulomb criterion at ambient pressure — treat as critically stressed today.”
Most “AI for the subsurface” puts a model between you and the answer. Here the answer is geometry and published rules; the model is kept to the two jobs it is genuinely good at, each behind a check you can inspect.
Paste a page of a competent-person report and the model turns sentences into candidate claims. Each is re-parsed by the same rule parser and graded by the engine, and each is shown beside the verbatim sentence it came from — so an invented claim has nowhere to hide.
Ask a question and the answer is drawn from that run's own fact sheet, citing the finding ids it rests on. Anything the run does not contain comes back as “not in this run” — not as a guess, and never as advice on whether a site should be permitted.
No model measures throw, classifies a relay, sets a threshold, computes a percentile, or writes a line of a regulatory dossier. Switch the assists off and every finding, number and PDF is identical — that is the test, and it is the reason the assists ship off by default.
Every threshold is a published number; every measurement is geometry. A language model may read your report prose or answer questions about a finished run — it never decides a finding, by construction.
Yielding 1997. Peacock & Sanderson 1994. Allan 1989. Morris 1996. Each finding names its rule and the paper it stands on — inline, not in a footer.
Monte-Carlo over pick error, position and Vshale. P10/P50/P90 on every number — including the ones that hurt the story.
Agent 05 re-checks every run against a five-point rubric and revises before sign-off — the check appears in the log and on the PDF.
Rhine Graben, NW Shelf, Utah FORGE — public geometries, yours to run the moment you sign in. Disagree with an agent in one click; that feedback is the beta.
design-partner beta · 500 credits free · private projects anonymized on ingest