Your Reservoir Model Came With Baggage. CLARISSA Unpacked It.
CLARISSA · · Road to ATCE
Someone hands you a reservoir model. It runs. There are folders inside folders, a collection of include files, and assumptions nobody remembers making. Congratulations: you have inherited a small mystery.
The next request sounds innocent: “Could you just make one change?”
Of course. Somewhere between finding the right files and working out which results belong to which version, “just” acquires a rather ambitious job description.
This is where CLARISSA has a particularly useful story to tell. In the custody-transfer case reported in our final ATCE manuscript, it gathered an inherited model’s dependencies, reproduced the predecessor’s results within tolerance, and only then accepted the requested modification.
First, account for the luggage
A simulation deck can depend on data held in several include files. Having the main file is therefore only the beginning. Before an edit means anything, the working set needs to be complete and the starting result needs to be understood.
In the paper’s reported case, CLARISSA parsed the inherited deck, located and resolved its dependencies, and assembled a self-contained file set. It then reproduced the predecessor’s summary vectors within tolerance. After that check, it accepted the requested change, which passed the validation gates and ran to completion.
That sequence is the achievement: establish the baseline, then change it. It gives an engineer a way to separate the effect of the requested modification from a problem introduced while assembling the model.
“It ran” is a promising first sentence
Successful execution is welcome news. It does not, by itself, tell you whether the inherited case has been reproduced or whether a forecast is appropriate for the decision in front of you.
The CLARISSA architecture described in the paper includes deterministic validation and audit artifacts. RIGOR provides a framework for evaluating model-generation and modification work against executable checks. These are practical ways to make a persuasive answer inspectable.
The distinction matters just as much when the AI is excellent. Capable frontier models, including Opus 5.5, can tackle simulation workflows; Anthropic’s release describes substantial coding and tool-use capabilities. The question for CLARISSA is what its engineering workflow adds. This paper reports the custody-transfer example; it does not establish an advantage over Opus 5.5.
Keep the engineer in the conversation
There is a lovely practical ambition here: make an existing model easier to work with, so the next engineering question has a better chance of being asked. A new fluid study or an unexpected well result should invite investigation, rather than a long search for the person who remembers the folder structure.
The reported demonstration is a concrete step toward that ambition. It is also a bounded result. The paper’s demonstrations use OPM Flow, and they do not establish performance across an entire fleet of field models. Passing the checks shows consistency against those checks; the engineer still decides whether the model is fit for its intended use.
What we are packing for ATCE
Our October 3 roadmap calls for testing, consolidation, and a worked demonstration as part of the Road to ATCE. Those entries describe planned preparation, not confirmation that every milestone has been completed.
The inherited-model case gives that preparation a clear purpose: show the starting point, show the evidence, and make the change understandable. That is a useful expression of our vision for subsurface AI—more room for engineering judgment because less effort disappears into reconstructing the setup.
Your reservoir model may arrive with baggage. CLARISSA’s reported demonstration shows it can do something helpful with it before anyone starts rearranging the contents.
Explore the CLARISSA workflow and request a demonstration →
Source: final manuscript SPE-234136-MS, A Conversational User Interface For Autonomous Reservoir Simulation Deck Generation And Execution, Sections 5.1, 5.4, 6, and 7; Road to ATCE plan dated October 3, 2026. Results described here are reported by the paper, not an independent replication.