Your Reservoir Model Needs a Memory. Sticky Notes Don’t Count.
CLARISSA · · Road to ATCE
The model contains a number. The number has survived three revisions, two handovers, and a folder called FINAL_final. Nobody remembers why it is there. Naturally, it now has seniority.
Reservoir engineers know this problem. A value can be perfectly readable while the decision behind it has disappeared. Was it supplied by the asset team, constrained by a correlation, or borrowed from a reasonable analog while better information was on its way?
One of CLARISSA’s most useful strengths is its deliberate treatment of that missing context. The architecture described in our final ATCE paper preserves engineering intent and records the origins of assumptions alongside the work of generating a simulation deck. The goal is practical: give the next engineering question somewhere solid to start.
Give the assumptions name tags
The paper describes a shared ledger: a structured record that checks the kind and consistency of information being entered. Recorded values carry provenance tags identifying them as user-supplied, constrained by literature or correlations, or defaulted from an analog.
Those distinctions matter. An analog default may be useful for a preliminary case, but it should remain visibly an analog default. A number supplied by a user also needs scrutiny; knowing its source does not make it correct. The benefit is that reviewers can see what kind of support a value has, rather than treating every populated field as equally established.
In the paper’s described workflow, checks on physical consistency come before commissioning deck generation. If a value is rejected, CLARISSA explains the violated condition conversationally. A repaired value requires the engineer’s confirmation before it enters the record. The provenance tags then follow through to the released deck.
That is a valuable habit to build into an engineering tool: make assumptions visible while they are being made.
Save the why, as well as the file
The architecture treats the deck as an output generated from preserved intent. The paper describes recording rationale during generation and deriving it again when the model is regenerated. This is a design for keeping the explanation connected to the actual model, instead of leaving it in a separate document that quietly ages.
Consider an illustrative review question: “We have new fluid information. Which assumptions should we revisit?” A visible record of supplied data and provisional defaults gives the discussion a starting point. It does not decide whether the new information is representative or whether the forecast is fit for use. Those remain engineering questions.
The distinction is useful precisely because engineers do change their minds when evidence improves. A good model record should make that easier. It should not require an archaeological expedition through someone else’s desktop.
A clever AI still benefits from a clear record
Capable frontier models such as Opus 5.5 can tackle simulation workflows. Anthropic’s release documents its coding and tool-use capabilities. Generating a deck is therefore only part of the story we want to tell about CLARISSA.
The useful question is whether the whole workflow helps engineers understand, inspect, and revise the result. Preserved intent, documented assumptions, deterministic checks, and evaluation through RIGOR address different parts of that question. The final paper is not a CLARISSA-versus-Opus-5.5 benchmark, and this article makes no superiority claim.
On the Road to ATCE: make the explanation travel
The October 3 Road to ATCE plan assigns October 10–11 to consolidation, bug repair, and planning the next phase. That is the plan’s schedule, not a report that those tasks are finished.
For a workflow built around engineering intent, a useful review question during that preparation is simple: can another engineer tell what was specified, what was assumed, and what still needs judgment? That is a proposed test of usefulness, not a new result claimed for the paper.
The paper’s demonstrations run on OPM Flow at reference-problem and single-asset scale. Its architecture and reported cases provide evidence to examine; they do not establish fleet-wide performance or automatic fitness for every decision. The engineer remains the final acceptance gate.
That is the direction behind our vision for subsurface AI: make more of the model’s reasoning available to the people responsible for using it.
A sticky note can remind you to buy coffee. Your reservoir model deserves a more durable explanation.
Explore CLARISSA and request a demonstration →
Sources: final manuscript SPE-234136-MS, A Conversational User Interface For Autonomous Reservoir Simulation Deck Generation And Execution, Sections 3, 5.1, and 7; Road to ATCE plan dated October 3, 2026. Architecture descriptions and demonstration scope are attributed to the paper; no independent replication is claimed.