MasterControl Seventeen Every Time
MasterControl AI Lab
MasterControl AI Lab
Published September 2, 2026 · on the feed September 9, 2026 · ▲ 3 · arXiv 2609.03209
The abstract
We study a governed approach to enterprise analytics: a language model interprets the question, while deterministic policy selects and runs a pre-approved analytical program that returns both results and evidence. We show that this restriction can remain expressive within a defined analytical class, using relational operations plus aggregation, comparison, windows, ranking, and similarity. Fixed meaning, policy, data, and execution rules also make results replayable. Across 440 runs, three 8B models generated SQL and selected tools at runtime, while Qwen3-8B interpreted intent only and policy executed the approved program. None of 330 runtime-planning episodes matched the full answer-and-evidence contract across all test datasets; the policy-executed analyzer matched 110 of 110. This is a configuration-specific result, not evidence that runtime agents cannot succeed under other designs.
The Brief
1:45A short podcast summary of the paper. Written and voiced by AI.
Transcript
MasterControl Seventeen Every Time. By MasterControl AI Lab, from the MasterControl AI Lab. Published on the second of September. No code repository is linked to it.
The setting is enterprise analytics, and the approach studied is a governed one. A language model interprets the question, and nothing more. Deterministic policy then selects and runs a pre-approved analytical program, which returns both results and the evidence behind them. The authors show that this restriction can remain expressive within a defined analytical class: relational operations, plus aggregation, comparison, windows, ranking, and similarity. Because meaning, policy, data, and execution rules are all fixed, results also become replayable.
The comparison ran over four hundred and forty runs. In one arm, three 8B models generated SQL and selected tools at runtime. In the other, Qwen3-8B only interpreted intent, and policy executed the approved program. Of three hundred and thirty runtime-planning episodes, none matched the full answer- and-evidence contract across all test datasets. The policy-executed analyzer matched one hundred and ten out of one hundred and ten.
The authors draw the boundary themselves. This is a configuration-specific result. It is not evidence that runtime agents cannot succeed under other designs. What it shows is how these particular runtime-planning setups fared against this governed configuration, on this contract, which demands both the answer and its evidence.
The trade on offer is flexibility for guarantees. Letting the model plan at runtime gives it room, and here that room produced no fully contract-matching episode. Confining it to interpretation, and handing execution to fixed policy, produced a perfect match and results that can be replayed.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a number that looks right but may not mean what you think.
You run analytics for a regulated company. Someone asks how many approved temperature-excursion deviations happened last week, and the system answers 17.
The road ahead
- 1 · The problem ~6 min Getting a number out of company data is easy; making sure everyone knows exactly what that number measures is the hard part.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Let the language model only interpret the question; fixed policy picks and runs a reviewed program that returns the answer with its evidence.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Four steps: a model interprets the question, policy selects an approved program, the program runs on the data, and results return with evidence.
- 6 · The experiment ~6 min You’ll guess: On the full answer-and-evidence contract, how did three 8B runtime-planning agents compare with the policy-executed analyzer?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.