MTR SIGMA
An agent that answers from computation, not memory.
MTR SIGMA, the Strategic Intelligence Game Modeling Assistant, works over MTR simulation output. Ask a strategic question in plain language and it plans the work, retrieves the scored worlds it needs, each one a complete version of your market with every carrier's move in every county, does the math, and answers with full provenance. It is not a chatbot with opinions about your market; it reports what the engine computed.
Grounded
Every claim traces to computation.
The agent answers from the county-level reads the MTR engine computed. If the engine did not compute it, the agent does not say it.
Agentic
Plans, retrieves, computes.
Give it a question and it decides what to look up, what to compute, and in what order, working across your MTR Latitude Studies before it answers.
Auditable
Show the work.
Each answer carries its provenance: the run, the model version, and the data that produced it. Your analysts can verify every step.
In scope
- Northwind Health Plan · Texas
- Northwind Health Plan · New Mexico
The agent reaches only the studies on your licence.
Threads
- Rio Grande Valley expansionNow
- Where our posture and the model disagreeYesterday
- Counties carrying the thinnest margin4 days ago
This vintage
- Released
- 9 August 2026
- Next release
- October 2026
A vintage is the CMS data release the study is built on.
We are weighing a service area expansion into the Rio Grande Valley for plan year 2028. Does it hold up?
SIGMA
Across the four counties it splits two and two. Hidalgo and Cameron support expanding. Starr and Willacy do not.
| County | Leading action | Of worlds | Support | Durability (0 to 1) |
|---|---|---|---|---|
| Hidalgo | Develop | 71% | High | 0.78 |
| Cameron | Develop | 64% | High | 0.71 |
| Starr | Hold Out | 58% | Medium | 0.62 |
| Willacy | Hold Out | 66% | Medium | 0.66 |
In Starr, the worlds where you enter are mostly worlds where Gulf Star plays Prepare Enter alongside you, and your position there carries 0.19 less durability than the Hold Out row.
Support tier follows durability, not how often an action wins. Willacy holds a higher share than Cameron and a lower tier because its position is less robust to what the field does.
Read 12,480 scored worlds in 2.4sShow the steps
Why is Cameron weaker than Hidalgo if both are frontier-supported?
SIGMA
Because 0.71 is an average, and in Cameron the average is hiding the shape. The worlds are close to bimodal on one rival's action.
Hidalgo does not do this: its worlds cluster tightly around 0.78 whatever the field does. So Cameron is not a moderately good bet. It is a good bet or a poor one, decided by a single rival, and the mean sits in a region almost no world actually occupies.
Rio Vista displays as aggressive on 13 years of observed county decisions, which is context for how to read that 22%, not a weight I have applied to it.
Durability distribution across 3,050 Cameron worlds
Is there an action that is less exposed to which way that goes?
SIGMA
Yes, and it costs you the ceiling.
Develop is the leading action and reaches 0.84 in the favourable branch. Enter with a defensive benefit posture is second by mean, but it holds 0.79 and 0.74 across the two branches. Narrower spread, lower peak.
Restricting to rivals that anticipate your move before making their own, the gap between those two actions narrows further. If you think Rio Vista reasons that way, the defensive posture is close to free.
Both actions scored on the same world set
Before we commit to the Valley, is there anything in the rest of the book we should be looking at first?
SIGMA
Outside the question
One county you already hold carries more gap at stake than the entire Valley expansion.
I scored your marked posture against the leading action in all 41 Texas counties you operate in. You diverge from the model in 12 of them. Eleven are small. Bexar is not.
You have Bexar marked Hold. The leading action there is Develop, at 68% of consistent worlds, frontier-supported, 0.80 durability. It is the largest single divergence in your book, and the gap at stake on that one county exceeds the four Valley counties combined.
That does not tell you the Valley is wrong. It tells you the cheapest available move is in a market you already have.
41 counties scanned · 12 divergences · sorted by gap at stakeOpen divergence list
Rio Vista is in Bexar too. Are they likely to exit it?
SIGMA
I can name that risk, not score it. Rio Vista sits on the exit watchlist in two counties, Bexar among them. I will not give you an exit probability: that model has not cleared our validation bar, and a number from it would be worse than no number.
What I can score is the branch. In the 14% of Bexar worlds where Rio Vista plays Prepare Exit, your Develop position carries 0.88 rather than 0.80. The downside case for the recommendation is favourable, which is worth putting in front of the committee precisely because it is unusual.
Watchlist framing · no exit probabilities reported
What is the timing on each?
SIGMA
Two clocks, and that is the part that bites. The Valley is a footprint decision: it needs a service area expansion, so the Notice of Intent to Apply falls in mid-November 2026 and the application on the second Wednesday of February 2027.
Bexar is a bid decision in a county you already hold, so it runs on the April pricing window. Different deadline, larger number, and no filing required.
The filing is also cheap and reversible: approval is an option, and the June bid decides which approved counties actually launch.
Decision windows from the CY2028 cycle
Draft it for the bid committee.
SIGMA
Done. The four Valley counties with a recommendation each, the Cameron split and the defensive alternative, the Bexar divergence as the largest gap at stake, the two watchlist notes with their limitation stated, and both deadlines called out separately.
Every figure carries the run and dataset it came from, so your actuary can trace any line back to the worlds underneath it.
Open draft memoStamped user-requested · vintage 2027
Ask about your markets…
Sources
- Scored worlds read
- 28,900
- Counties scanned
- 45 of 254
- Rivals modeled
- 3
- Model version
- mfn-2026Q3-v5
- Dataset
- cms-2026-05
Open any figure to see the worlds it came from.
Guardrails
- Exit questions return a watchlist, never a probability.
- Equilibrium output, meaning which market configurations hold once every rival has responded, is characterized, not forecast.
- Nothing is asserted that the engine did not compute.
- Questions are processed on MTR systems and logged under the AI Use Disclosure.
Take it with you
- Export as a memo
- Add to a study addendum
- Share with named readers
The reasoning layer
The agent reasons over a model of your market, not a pile of documents.
Counties, carriers, and every position the engine has computed, held as one connected structure. Questions resolve by walking it.
Context is your market, not your last message
It starts from your footprint rather than your chat history. Ask about your exposure in any part of that footprint and it already knows which counties that means, and who else is in them.
Provenance is the route, not a footnote
The citation is the path the agent took, so your analyst can walk back up it. A plausible citation and a traceable one look identical until someone checks.
It knows where its evidence ends
Uncovered market, year past the horizon, quantity the engine does not compute: it says so, and says what would have to run. Refusal is architecture, not a warning label.
The question can be one we never anticipated
Relationships, not rows. Who competes with us where, what is reachable from the position we hold, which neighboring counties resemble the ones we are winning.
An agent that talks about your market is a writing tool. One that has your market loaded can be checked.
Join the MTR SIGMA early list
MTR SIGMA is in development for MTR Latitude Study clients.