Oklahoma Municipal Natural Gas Coalition

Dekacern

How the system decides — and how much of the decision
it is trusted to make.


A walk through the autonomy ladder, what goes into each decision,
where we are today, what's next, and how we prove it works.

Prepared for Bill & Rob · 2026-07-01 · point-in-time, no peeking

What we'll cover

Six things

1 · The autonomy ladder

How the system moves from watching to doing — safely, one rung at a time.

2 · What goes into a decision

The full picture the system weighs before it recommends anything.

3 · Where we are today

What is built and running right now.

4 · What's next

The near-term road to acting on your behalf.

5 · Performance & how we test it

The scoreboard — and why the way we test it means the numbers are honest.

6 · How we decide, pipeline by pipeline

The detailed playbook for OGT, Southern Star, and everyone else — for after the scoreboard.

01

The autonomy ladder

View → Suggest → Act. The system earns trust before it takes it.

The principle

We never flip a switch from "off" to "autopilot"

The system climbs a ladder. Each rung hands over a little more, and only after the rung below it has proven itself on your real months.

Phase 1View & Suggest

The system recommends. A person executes. Every decision arrives with its reasoning, alternatives, and what it would cost to do nothing. You stay fully in control.

Phase 2Approve → Act

The system prepares the action; you approve it, then it executes. One click instead of a spreadsheet. The judgment is still yours.

Phase 3Interrupt → Act

The system acts inside a window unless you stop it. For routine, reversible moves only. You get notified with time to intervene.

The safety mechanism

The rung you're on is not the whole story

Every individual action carries its own reversibility class — a ceiling on how much autonomy it can ever be given. The system always takes the lower of the two.

how much it may do  =  min( system phase , what this action allows )
REVERSIBLE

In-tolerance nomination, no-fee pull from a sister town.

Ceiling: Phase 3 — may auto-act with an interrupt window.

MEDIUM

A deliberate storage withdrawal.

Ceiling: Phase 2 — must be approved first.

CRITICAL

Locking in baseload at a panic price; answering an OFO.

Ceiling: Phase 1 — always a human's call.

Safety property we can state plainly: even at full autonomy, the system will never auto-commit a critical, hard-to-unwind move. Those always come to you.

02

What goes into a decision

First the full picture. Then what must happen, and why.

The shape of every plan

Two halves, kept honest by keeping them separate

The Picture — facts

Everything true about this town this month, that anyone could verify:

  • Every tool available & its state
  • How towns are connected (who can lend to whom)
  • The price environment
  • Contract & tariff limits in force

The Plan — recommendation

What the system believes must happen, and why:

  • Each action, typed and dated
  • Its reasoning & confidence
  • Its reversibility class → autonomy ceiling
  • What happens if no one acts

We call this the Monthly Operating Plan. It's the single artifact a scheduler reviews — and the same object the system would eventually act on.

The Picture · every lever, every month

The tools at the system's disposal

For each tool we track four things: what you're entitled to, what's available this month, its current state, and its limits.

Baseload

The monthly commitment bought at first-of-month price.

Storage

A finite account — inject in summer, draw on cold days. The catastrophe hedge.

No-notice

Same-day withdrawal to cover a spike without a market buy.

Pool transfer

Move gas between sister towns — often no fee.

Park & loan

Borrow or lend against the pipeline within terms.

Imbalance bank

The running over/under position and its cash-out rules.

Facts we know from your contracts are shown plainly; anything we're still estimating (live inventory, today's imbalance) is flagged as estimated until the pipeline feeds are wired.

The Plan · the eight decisions

Every recommendation is one of eight kinds

  • D1  Nomination — per cycle
  • D2  Daily purchase / sellback
  • D3  Pool transfer — pull / put
  • D4  Storage touch — inject / withdraw
  • D5  Imbalance response — overs / unders
  • D6  Baseload sizing — the big monthly call
  • D7  Storage trajectory plan
  • D8  Contract elections — renew / amend / MDQ

When it needs to cover a shortfall, it spends in this order:

no-fee transfer park & loan storage draw market buy

Cheapest cover first — the same instinct a good scheduler has, made explicit and applied every time.

The one thing it's optimizing for

A prudent supply manager — not a trading desk

Minimize loss, not chase profit

The objective is to protect the coalition from bad outcomes, weighted toward the tail — the cold-snap and Uri-scale months that do real damage. It is not trying to beat the market on an average day.

No peeking

Every recommendation is built from only what was knowable at the time — the prices, weather forecasts, and contract terms in hand at bid-week. Never a number that hadn't happened yet.

This is the discipline that makes the test results in Section 5 trustworthy: the system in a backtest sees exactly what it would see live, with the clock moved back.

The data behind the picture · weather

Weather tells us how much gas each town will burn

What we receive

  • Daily observations back to 1994 (Open-Meteo): temperature, humidity, wind, snow — turned into Heating Degree Days, a single measure of how cold the day was.
  • As-issued forecasts — what the forecast actually said on each day, archived since 2024 and never rewritten after the fact.
  • Each town mapped to its own weather station(s).

How we use it

  • Each town's usage rises a fixed amount per degree-day, fit from 12 years of its own meter data.
  • A rolling 30-day base tracks its current level; steep industrial growth is tracked separately.
  • Cold snaps and price spikes are replayed together — the engine never separates a hard freeze from the prices it brings.

Honest limit: daily forecasts are trustworthy out to ~16 days — there is no reliable 30-day. So forecasts sharpen day-to-day execution, not the month-ahead commitment. And because forecasts are stored as-issued, a replayed decision only ever sees what was known that day.

The data behind the picture · prices

Three prices from NGI — each drives a different decision

Daily  ·  GDD

“Gas Daily” — today's spot price. What you pay to cover a shortfall on a cold day, or receive when selling back extra. Live from NGI, reaching back to 2008.

First-of-Month  ·  FOM

Set once at the start of each month (Inside FERC). The price your baseload is locked to — the big monthly commitment.

Forward

The market's price today for a future month. How we value and plan months that haven't priced yet.


How the engine uses them: baseload is committed at the monthly price; a shortfall is bought at the daily price (a little more on cold days, when it spikes); leftover gas is sold back at a discount. If the monthly price spikes far above its normal range, the engine trims the baseload — a panic price usually reverts, and you don't want the whole month locked to it.

Every price is stamped with the instant it was published — a replayed decision can never see a price that hadn't printed yet. Same no-peeking discipline as the weather.

03

Where we are today

The whole chain is built — from raw data to a plan on a scheduler's screen.

Built end to end

From your data to a decision — the full chain runs today

real data state of every tool cheapest-cover math Monthly Operating Plan scheduler screen

Real data, no stand-ins

12 years of daily usage across 19 towns; real point-in-time daily & first-of-month prices (NGI, live); verified tariff rulebooks for Southern Star, PEPL, and Oklahoma pipelines.

The decision engine

Sizes each town's monthly baseload, sizes storage-sharing pools collectively, and carries a storage account across months so the savings are actually bankable.

The Monthly Operating Plan

Generated automatically per pool & month — Picture + Plan, with confidence flags and autonomy ceilings on every action.

Live & on-screen

Running behind a secure OMNGC-themed site; schedulers can open a plan and see the picture, the recommendation, and the reasoning.

Honestly: which rung are we on?

Phase 1 today — View & Suggest

Phase 1View & Suggest

Live now. The system produces a complete plan with reasoning; a scheduler reviews and executes. This is where trust is earned.

Phase 2Approve → Act

Next build — the approve-then-execute wiring.

Phase 3Interrupt → Act

Later — reversible moves only, always with a stop window.

Deliberately. We want a season of the system's suggestions sitting next to what your schedulers actually did before anything acts on its own.

04

What's next

The road from suggesting to acting — and sharpening the picture.

Near-term priorities

Four moves

Climb to Phase 2

Build the approve-then-act workflow: a scheduler approves a recommendation and the system carries it out. The first real hand-off.

Wire the live feeds

Connect pipeline inventory & imbalance so the "estimated" flags come off — a firmer picture, especially the storage number.

Widen the flexibility map

Enter more towns' storage & pooling entitlements. The dollar value scales directly with how many towns we can represent — only two carry storage in the model today.

Close the expert gaps

A short list of questions for you & Rob — sellback terms, the OGT pooling agreement, and the risk-appetite dial — each worth real money in the results.

Also in flight: regulatory watch (OCC scanning + calendar) and continued model sharpening on individual towns.

One decision is genuinely yours

How much tail insurance to carry

The system can be tuned to buy more or less protection against catastrophic winters. More protection costs a small premium in mild years and pays off enormously in a Uri.

~$2–3Kper town, per winter-month
cost of full cover in a mild year
~$167Ksaved across four towns
in one cold January (Jan-2026)
$2.07Msaved on two storage towns
in the Uri event alone

This is a coalition call, not a modeling one. We've made it priceable so you can set the dial deliberately.

05

Performance & how we test

The scoreboard — and why the way we keep score makes it honest.

How we test · the method

We replay six years with the clock moved back

Backtest = live

We run the system across 2020–2026, month by month, feeding it only what was knowable on each decision day. The backtest and the live product are the same code — one just has the clock rewound.

Three yardsticks, not one

Every month is scored against naive (buy the same as last year), perfect (best possible in hindsight), and — where we have it — what your schedulers actually did.

naive
worst you'd tolerate
—— the system & your team
the real contest
—— perfect
hindsight ceiling

Success = how much of the naive→perfect gap we capture with information available at the time.

How we test · the discipline

Nothing ships unless it passes a fixed bar

Every change to the model must clear a pre-registered gauntlet — set before we look at the result, so we can't move the goalposts:

Beats humans

Improves the town-vs-scheduler comparison on real decisions.

Doesn't break the fleet

Stays within the fleet-wide guardrail across all six years.

Preserves the tail

Keeps the Uri-event protection — the whole point of the exercise.

Levers that fail are documented as rejected right in the model, so we don't relitigate them. Several promising ideas have been tried and turned down this way.

Where the numbers stand

Versus your schedulers, on real prices

TownPipelinevs. scheduler
CopanSouthern Star+$49.2K▲ system ahead
GraniteEnable / EGT+$8.2K▲ system ahead
RipleyOGT−$11.0Kresidual shoulder over-buy

The system wins the cold months decisively — where the money is — and gives a little back in mild shoulder months. The remaining gap on Ripley is understood and being worked.

The tail is preserved: in a rerun of the Uri event, the system captures +7.9% more than naive — that protection was not traded away to win the mild months.

Model v0.6.0 · forecast era · scored against actual scheduler decisions on real point-in-time prices.

The clearest win we can attribute

Sharing & storage, on two real towns

Drumright + Mannford, sized together with their shared storage account, over 16 months:

$9.26Mnaive cost
$7.39Msystem cost
+$1.87Msaved · 36% of the gap captured

The value is storage, not diversification — a firm backstop lets the system size the monthly commitment lean without fear, so it stops over-buying in warm months and still covers the cold-day spikes. The savings are bankable: we track the account across months, and it never runs dry.

Honest caveat: only two towns hold storage in the model today, so the fleet-wide figure grows as we enter more entitlements (Section 4).

Take it with you

The full briefing packet

Everything in this deck, written up long-form for a non-scheduler — the worked Ripley replay, the storage finding, the four-town test, the no-peeking methodology, a plain-language glossary, and the tariff + decision-record appendices.

Regenerated under model v0.7.0 (OGT anchor) on real point-in-time prices · dekacern.org/reports

06

How we decide — pipeline by pipeline

One objective, three playbooks. The tools on each pipeline differ — so the decision does too.

How we decide · OGT

OGT — buy what you'll burn, cover the spike spot

What we have to work with

  • Baseload at the monthly index + a small supplier adder
  • The daily spot market for anything extra
  • A tight ~3% imbalance tolerance
  • No on-system storage · pooling unconfirmed

So the decision is

  • Size baseload ≈ expected usage — what the town will actually burn
  • Cover cold-day spikes on the spot market
  • Decline standing tail-insurance in baseload — with no storage, over-bought gas is stranded and dumped at a loss

This is exactly what your schedulers do — they buy ~1.0× usage; the model used to buy ~1.6×. We measured it and taught the engine to match, with an expected-usage anchor that recovers the OGT over-buy losses. The honest catch: buying lean accepts a surprise-cold hit — the proper OGT tail hedge isn't baseload, it's storage (the ONEOK Gas Storage question). How lean to buy is a risk dial.

How we decide · Southern Star

Southern Star — size with confidence, the tools cover you

What we have to work with

  • A real tolerance band — greater of 1,000 units or 5%
  • Member pooling — move gas between our towns at no fee
  • On-system storage — inject in summer, draw on cold days
  • Park-and-loan against the pipeline

So the decision is

  • Size baseload to cover the cold months confidently — we can lean slightly long because excess is cheap to unwind
  • Draw storage on cold days instead of buying at the spike
  • Net imbalance monthly inside the free-swing band; pool a long town into a short one
  • Size storage-sharing towns collectively (Drumright + Mannford)

This is where we beat your schedulers (Copan +$55K) — the flexibility covers cold months cheaply without a warm-month penalty, and sizing storage towns together is worth +$1.87M. Autonomy tracks reversibility: in-band nominations & no-fee pool pulls can auto-run with an interrupt window; a storage draw needs approval; locking baseload at a panic price is always a person's call.

How we decide · everyone else

Enable/EGT & Panhandle — rule-driven, one hard line

Enable / EGT

Zero tolerance for shorts since Dec 2025 (Critical Notice 6224). The rule writes itself: never be short. Size adequately, tight nomination discipline, treat every day like an emergency order. (Granite +$9K.)

Panhandle / PEPL

Monthly balancing — the greater of 1.5× the daily contract quantity or 1,000 units — and post-month trading through the 17th business day. More room to true up cheaply after the fact.

The pricing layer

On every pipeline the molecules are bought from a supplier at the monthly index + a small adder (e.g. Southwest Energy). The pipeline is how it's delivered and balanced — the supply contract is the price.


The gap we're closing: OGT and EGT terms are contractual and not yet verified against posted tariffs — today we run from the schedulers' memory, flagged on every decision. Getting those agreements sharpens all three playbooks. This is a standing ask.

In summary

Three playbooks, one scoreboard

OGT

No storage → buy expected usage, cover spikes on the spot market, skip the baseload tail-insurance. Now matches the scheduler — the real tail hedge is storage.

Southern Star

Full toolkit → size long with confidence, draw storage on cold days, pool and net monthly. This is where we beat the scheduler.

Everyone else

Rule-driven → never short EGT, true up Panhandle monthly. On every pipe, the price is the supply contract.


Where we stand: we beat your schedulers where we have the tools — Southern Star, Enable, every real cold snap. The OGT fix just recovered +$86K of the gap. Across all towns over the last 6 months we're not ahead yet — the remainder is storage not-yet-scored plus the fast-growing towns — but the path to parity is in hand.

The other half of every decision

The demand side — when a town's future isn't in its past

Every buy has two unknowns: the price of gas and how much the town will burn. Our sharpest fixes so far were about price and flexibility — this is the load side.

The issue

  • We forecast a town's use from its own history — a weather-normalized base plus how hard it leans on cold.
  • That's right for a stable town. But a fast-growing one's next winter isn't in its history — a new plant or customer is a fact the schedulers know and the meters haven't seen yet.
  • So the model trails the ramp: Pryor Creek's cold-day load is up +31% year-over-year; Tuttle's weather-sensitive load nearly tripled. This is the biggest piece of the residual OGT gap.

The fix — now built

  • Load-event tickets: a scheduler records "+Y/day (or per cold-degree) from date D, because…" and the forecast uses it immediately.
  • Point-in-time & self-fading: it only affects decisions made after it's entered, and fades over ~270 days as real usage absorbs the new load.
  • Growth-watch: the system now flags towns whose weather-adjusted load is drifting and pre-fills a ticket — so the gap gets caught, not discovered in hindsight.

Evidence: on Tuttle's ramp winter, giving the model that one ticket cut the loss nearly in half — −$54.8K blind to the ramp → −$28.6K with it, a $26K save. Human knowledge in; the system supplies the prompt, the scheduler supplies the "because."

In one breath

A prudent supply manager that shows its work,
earns autonomy one rung at a time,
and is tested the way it runs.

Today: the full chain runs and suggests. It beats your schedulers where the money is, and keeps the catastrophe hedge intact.

Next: the first hand-off (approve → act), firmer live data, and more towns' flexibility on the map.

Two things we need from you: the short expert-question list, and a decision on how much tail insurance to carry.

Dekacern · decide in dekatherms · 2026-07-01