Firm vs. Estimate:
How to Read the Numbers
in an Energy Study

An energy study is a stack of numbers set in the same font, and the font is the problem. Some of those numbers are documented facts, some are expiring quotes, some are borrowed averages, and some are assumptions with a decimal point, and the page almost never says which is which.

Section 01Every number wears the same font

Open the savings table in any energy proposal and read across one row at a time. A tariff rate that a regulator approved on a documented date sits beside an equipment price that expires in six weeks, which sits beside an availability figure describing a fleet the seller does not operate, which sits beside an escalation rate a modeler chose because the template had always used it. Four different kinds of knowledge, one typeface, no labels.

The arithmetic in these documents is almost always right. Spreadsheets do not make multiplication errors, and the people who build them are competent. What fails is the layer underneath the arithmetic: the status of each input. A model can be internally consistent and externally unmoored at the same time, and the energy decisions that go wrong at scale usually trace back not to a wrong formula but to a number that was treated as firm when it was somebody's assumption.

The remedy is not more modeling. It is a labeling discipline that any owner can impose and any honest analyst can satisfy: every load-bearing figure in a study carries one of four tags, and a figure that cannot be tagged from the document alone is treated as the weakest class until proven otherwise. This paper defines the four tags, walks through the four places where estimates most often pass as firm, and closes with an audit an owner can run in about fifteen minutes without an engineering degree.

Section 02The four tags

The tags answer one question: what would have to be true, and documented where, for this number to deserve trust. There are only four answers that matter.

Two properties of this system matter more than the definitions. First, firmness decays. A tariff is firm as of its date and no later; a quote dies at its expiry; reference data ages. The date is part of the number, which is why the strongest tag is sourced-and-dated rather than merely sourced. Second, illustrative is not a slur. Every multi-year model requires assumptions, and a study with no illustrative lines has simply hidden them. The defect is never the estimate. The defect is the missing label.

Section 03The four places estimates pass as firm

Mislabeling is not evenly distributed. Across technologies and across sellers it concentrates in the same four cells, because those are the cells where an optimistic assumption does the most commercial work. An owner who knows the four addresses can audit a document quickly; the rest of the table tends to be honest because the rest of the table is checkable.

Availability: the fleet is not your site

Every generation proposal carries an availability or output figure, and it is usually the least examined number on the page. Start with a distinction the sales document may blur: availability describes the share of hours a machine could have run; capacity factor describes the energy it actually produced against its nameplate potential. The two can sit far apart, and a model that multiplies nameplate by availability has quietly promoted a maintenance statistic into a production forecast.

Reference data exist and are worth demanding. The U.S. Energy Information Administration publishes average capacity factors for utility-scale generators by technology, monthly and annually.¹ Those tables are the right starting point, and they carry the indicative tag on their face: fleet averages blend old units with new and well-run plants with neglected ones, and a single-site installation is not a fleet. One machine that loses a major component sits at zero until the part arrives; the average contains that story and its opposite, and predicts neither. The honest use of fleet data is screening and bounding. The honest form of a site-level availability claim is a contract.

Each technology class earns its own scrutiny here, in both directions. Reciprocating engines and turbines carry the deepest fleet histories, which makes their indicative data unusually good, and they carry scheduled-overhaul calendars that belong in the model as visible downtime rather than in a footnote. Fuel cells run quietly at high availability between stack events, and stack replacement is a known lifecycle cost that a fair model shows as a dated line rather than a surprise. Solar arrays rarely break, and their production swings with weather and season, so the operative uncertainty is resource rather than repair. Storage responds in milliseconds, and its usable capacity is a moving target that degradation and cycling assumptions must track honestly. None of those sentences disqualifies any technology. Each one identifies which cell in that technology's column deserves the auditor's pen.

A written availability commitment in an executed service agreement is the only version of this number that approaches firm, and even it deserves two readings: the measurement definition, with its exclusions for scheduled maintenance and external causes, is the real number; and the remedy for missing it is paid in money, not in electricity, so the model should also show what the site does during the shortfall the remedy compensates.

Escalators: the quietest cell in the model

No single input moves a multi-decade model like the escalation rate, because it compounds. One percentage point of annual escalation carried for twenty years lifts the terminal-year price by roughly a fifth; that is arithmetic, not opinion. The commercial temptation is symmetrical. A model selling against the utility bill wants the utility escalator high, so the avoided cost grows every year. The same model wants its own fuel and maintenance escalators low, so the lifecycle cost stays flat. Both choices can hide in one spreadsheet, and each is defensible in isolation, which is what makes the pair effective.

History, at least, is knowable. The Public Advocates Office at the California Public Utilities Commission reports that average residential electric rates rose 101 percent in Pacific Gas and Electric territory between January 2016 and January 2026, 98 percent in San Diego Gas & Electric territory, and 76 percent in Southern California Edison territory over the same decade.² The first of those compounds to about seven percent a year.

Here is the discipline in one sentence: the 101 percent is sourced-and-dated, and the moment anyone projects it forward it becomes illustrative, whoever's history it extrapolates. A decade that doubled rates is neither a floor nor a ceiling on the next one; the same regulatory system that produced those increases is now under explicit pressure to restrain them, and rate trajectories are a contest, not a formula. A study that wants trust shows its recommendation at the chosen escalator, at half of it, and at zero real escalation, and says plainly whether the answer survives all three. If the recommendation flips, the recommendation was the escalator.

Incentives: statutes are firm, qualification is not

Federal incentive lines invite a specific confusion between two different numbers. The statutory rate is genuinely firm: under current federal law as of this writing, the investment tax credit for qualifying energy property is 30 percent.³ But qualifying is a fact about your project, and your project does not exist yet. Property-class eligibility, construction timing, and placed-in-service tests are all future events, so even the cleanest statutory figure enters a model as a conditional line, and an honest study says so.

The adders are where conditional becomes fictional. Statutory bonus provisions can raise the credit meaningfully above the baseline for projects that qualify, and each bonus must be individually established: the energy-community adder turns on location determinations against published federal criteria, and the domestic-content adder on certifications under detailed and evolving guidance.^3,4 A model that books adders into its base case has converted a statutory maybe into cash. The tag ladder is strict here: the statutory rate is sourced-and-dated; your project's qualification is indicative at best until qualified tax counsel has examined it; an unexamined adder is illustrative, whatever the cover letter implies. State and utility program lines deserve the same treatment plus one further test, the as-of date, because programs step down, fill up, and close, and incentive lines from closed programs have a documented habit of outliving the programs inside circulating sales templates.

The operating rule for an owner is short. Incentive numbers move from illustrative toward firm only through named counsel and dated determinations, never through repetition.

Timelines: a date is not a fact because it has a month in it

Dates read as binary, which makes them feel firm; they are usually the softest numbers in the document. Audit them by family. Equipment delivery is quoted, and therefore firm only inside a quote's validity window and its conditions: deposit terms, production slots, and what happens to the date if the deposit moves. Permit durations are indicative at best, because air-quality treatment is district-specific and site-specific; a schedule that shows a permitting bar without naming the district has not studied your site, whatever else it has studied. Energization dates are firm in exactly one form: a written, dated statement from the utility for your specific request. Nothing recalled from a phone call qualifies.

The connection frameworks themselves are moving, which sharpens the point. In July 2025 the California Public Utilities Commission approved an interim rule to streamline connections for very large new loads, with applicants funding transmission work up front and key terms still being finalized.⁵ A timeline quoted under last year's framework can describe a process that no longer exists in that form. That is not a criticism of anyone's diligence; it is why the sourced-and-dated tag has a date in it.

The general test for any date is to ask what document stands behind it. There are only three good answers: a quote within its validity, a determination with a docket, or a utility statement in writing. Every date without one of the three is an assumption wearing a calendar.

Section 04Why the labels disappear

None of this requires bad faith, and the charitable explanation is the accurate one. Sales models are assets. They are built once, at real cost, and reused for years; assumptions harden into defaults; the analyst who chose the escalator left the firm two winters ago and the cell has been copied forward ever since. Sensitivity tabs get trimmed because clean tables close faster than honest ones. And the consequences are asymmetric: when an untagged estimate misses, the gap lands in the owner's operating budget, not in the seller's commission statement. Incentives shape documents the way water shapes stone, slowly and in one direction.

An unlabeled estimate is not a lie. It is a risk with no owner, and unowned risk always settles on the buyer's side of the table.

The same drift operates on every technology's sales channel, on both sides of any comparison, and on independent analysts too, including us. That is the argument for making the tags structural rather than trusting anyone's culture: a tag column turns a model into a document that someone other than its author can audit.

Section 05The fifteen-minute audit

This audit requires the study, a pen, and no engineering background. Two minutes a step is enough, because the goal is not to fix the numbers. The goal is to find out whether the document knows what its own numbers are.

1. Go to the one table the decision rests on.Usually the savings or lifecycle-cost summary. The prose was written after the table and inherits its numbers, so audit the table and let the prose wait.
2. Tag every load-bearing cell.For each input, ask whether a stranger could find the document and its date from this study alone. Sourced, quoted, indicative, or illustrative. Any cell you cannot tag from the page is your first finding.
3. Circle every percent sign.Escalators, degradation, discount rate. Ask for the model at the stated escalator, at half of it, and at zero real escalation. A recommendation that cannot survive the low case is an escalator with a cover page.
4. Trace the incentive line to statute and counsel.Separate the statutory baseline from booked adders. Ask who qualified each adder, on what date, and whether tax counsel has been named. Unqualified adders move to illustrative.
5. Demand the document behind every date.Quote validity for delivery, a named district for permits, a written utility statement for energization. A date with no document is an assumption; move it and watch what happens to the schedule.
6. Match availability to its paper.Executed service terms with a measurement definition and remedies, published fleet data, or a brochure figure. Check that the model's downtime assumption matches the contract's definition rather than the pitch's number.
7. Find the oldest source date in the pile.A study is as current as its stalest load-bearing input. If the controlling tariff, quote, or determination predates the study by a year, the study is a year old, whatever its cover says.

Fifteen minutes of this does not tell you whether the recommendation is right. It tells you something more useful at the decision stage: where the document's knowledge ends and its hopes begin, and how much of the case rests on cells nobody has to stand behind.

Section 06What to demand in writing

The fix costs the analyst one column. Any organization commissioning an energy study, from any party including ours, can require five things in the engagement letter: a tag on every load-bearing input; an as-of date on every source; an assumption register listing each illustrative value, who chose it, and why; the low case printed beside the base case rather than available on request; and a plain statement of which new fact would change the recommendation. An analyst who resists the column is telling you something about the column's contents.

Be equally clear about what tagging does not do. It does not make estimates accurate, and a fully tagged study can still be wrong; engineering and market risk do not go to zero because the labels are honest. What the discipline buys is different and, at the commitment stage, worth more: every uncertainty is visible, every number has an owner, and the board spends its argument on the cells that deserve it. A study with many clearly labeled illustrative lines is more trustworthy than one with none, because a real site contains real uncertainty, and if the uncertainty is not on the page it has been moved, unpriced, into your operating budget.

Our own stake in this is structural and worth stating once. We sell fixed-fee decision studies with no equipment margin behind them, which means no cell in our models improves our economics by being optimistic. That does not make us right. It makes us auditable, and auditable is the standard this paper is asking you to hold everyone to, starting with us.