Methodology

What the grade is

One grade per project, on one scale, forever. The grade is a public, versioned, deterministic function of the frozen evidence record. It is never recomputed on the fly for anything public, never silently restated, and never issued where the record cannot support one - that state is Not rated, a designed first-class output, not an absence.

The grade claims one thing only: the probability that the project's headline promise resolves. It does not claim the project is good, the sponsor is honest, or the economics clear.

What is rated

The rated unit is the project's headline promise: its stated nameplate capacity on its stated completion date, exactly as printed in the record at evaluation time. One rating, one atomic unit, one language - a sponsor with ten projects carries ten independent grades.

What resolved means

Every grade maps to a published probability band over a single binary event: the project reaches energization at 90% or more of its stated capacity within six months of its stated date, measured against grid-operator energization prints, state construction records, permit dockets, and first-party filings. This definition is pinned at first publication and can never move; methodology versions may change how the probability is estimated, never what the probability means.

Resolution is scored against the promise as it stood when the grade claimed it. If a sponsor restates capacity or date, the ledger records a promise-restated action - yesterday's grade is never scored against today's promise.

The scale

Five coarse grades plus Not rated. Coarse on purpose: fine granularity implies a calibration the record does not yet support, and coarse bands are defensible under scrutiny. The bands below are proposed pending the calibration backfill described under Governance; they pin where grades actually separate resolved outcomes.

Every published grade carries a coverage qualifier - Deep, Mixed, or Thin - computed from the project's evidence record. The qualifier never softens the number; it constrains publication.

When a project is not rated

Not rated is a published state with a reason, not a missing row. A project is not rated when any of five triggers holds at evaluation time:

• The promise is unmeasured. The project's stated capacity or completion date is not on the record, so there is nothing to grade against.
• The evidence floor. The record is too thin for the delivery model to produce a defensible distribution, or coverage classifies below the evidence floor.
• An unresolved conflict. A load-bearing fact - capacity, date, or status - carries an open conflict between sources.
• Thin coverage at initiation. A first grade is never published on thin coverage, however promising the record looks.
• A stale promise. A load-bearing input is past its re-verification horizon and re-verification has not landed.

Every not-rated state publishes its reasons. A grade is never withdrawn into silence: if the record thins, a not-rated action is appended with the reason.

How a grade is computed

Six deterministic steps. No analyst touches the number at any of them.

1. 1
   Freeze the record. Every input is pinned to an exact snapshot: the project's evidence rows, the sealed model runs, and the corpus state are hashed together. The grade is a function of that frozen manifest - never of whatever happens to be true at page-load time.
2. 2
   Check the gates. The five not-rated triggers are evaluated first. If any holds, the output is Not rated with its stated reasons, and no probability is estimated at all.
3. 3
   Read the delivery distribution. The sealed delivery model's frozen distribution for the project gives the probability that it energizes at 90% or more of stated capacity within six months of the stated date. Read directly, never re-simulated, smoothed, or adjusted by hand.
4. 4
   Apply two published adjustments. Dependency stress: each unproven critical dependency moves the probability by a fixed, published multiplier; dependencies that fail together are stressed once, not stacked. Track-record prior: where the sponsor's class has enough resolved outcomes, the observed base rate blends in at a fixed published weight; below that count the term is inert. Both adjustments are computed from observed inputs only.
5. 5
   Map to the scale. The adjusted probability maps to a band. Boundary values fall to the higher-risk grade. The coverage qualifier attaches and travels with the grade everywhere it appears.
6. 6
   Append to the ledger. The result is written as a rating action with its one-sentence reason, the frozen inputs, and the methodology version. Same frozen inputs and same methodology version always produce the identical grade.

The ledger

The grade ledger is append-only. No grade row is ever updated or deleted: every initiation, upgrade, downgrade, move to or from not-rated, reaffirmation, and promise restatement appends one action with a published reason. The current grade is simply the latest action. A wrong action is reversed by a new action citing it - errors stay on the record, because the errors are themselves the calibration record.

Recompute happens on every evidence intake, on material events (a date move, a capacity change, a cancellation, construction start or stop, a sponsor change), and on a fixed review cadence for every graded project. A recompute that changes nothing writes nothing - rating actions are state transitions, not ticks. A due review with no movement appends a reaffirmation, so the record shows the grade is watched, not just untouched.

Governance

Before the scale pins, the method is run backward over every resolved project on record, using only evidence that existed before each outcome - no look-ahead. The retroactive ledger is then scored against what actually happened: projects the method would have placed on the bottom rungs slipped or died at what rate; projects on the top rungs landed at what rate. That calibration table publishes with this methodology, and the constants and band boundaries pin where grades actually separate outcomes. This is the difference between a grade and marketing.

The methodology is versioned openly with a dated amendment log. When the method changes, the book is re-run and the re-grades publish as methodology-regrade actions citing the amendment - never a silent restatement. The scale's meaning is invariant; the methodology is versioned; the ledger shows every move and why.

Once grades are public, the failure mode is never-downgrading. A downgrade with a named reason is the product working. Misses stay on the page.

Amendment log

• September 12, 2026 - Amendment 1. The public page shows the current grade and latest change per project; the full historical ledger lives on the Atlas terminal. The record is unchanged; the publication surface narrowed.
• September 12, 2026 - Amendment 2. Every project on record enters the ledger and the public page - graded where the record supports it, not-rated with stated reasons everywhere else. Graded projects list first; not-yet-rated records follow, collapsed by default with reasons on expansion. Grading gates are unchanged.