· Methodology

How we measure real vs. phantom

Method open on purpose — a black box never becomes the reference. What we measure, how, from which sources, how you check it, and — honestly — what we do not claim. LOADSTAR's asset is credibility: here's the proof, not the promise.

1. What we measure

For each U.S. region, a 0–100 number that answers one question: is the announced AI-infrastructure buildout there real or phantom? Real (68–100) · Unsure (45–67) · Phantom (0–44). It's descriptive: it measures today's formation signals — not a promise of a future hit (that's the track record, section 5).

2. The formula (exact, reproducible)

The Compute Formation Index (CFI) is the weighted average of 4 signals, each 0–1:

35%	Formation — convergence of independent signal types (permit + contract + hiring + energy)
30%	Energy — interconnection queue + % of committed MW + EIA demand (log scale)
20%	Momentum — signal acceleration (30-day vs. prior 30-day window)
15%	Corroboration — share of the queue that is committed, not just speculative

The Reality Index starts from the CFI and applies honesty adjustments: it flags an announcement with no active request in the interconnection queue, and discounts when a queued request withdraws (a real, observable signal — the phantom dying). We treat these as weighted signals, not a verdict. Each region shows its exact derivation at /r/<region>/how-it-works.

On queue nuance (we know it isn't a binary): interconnection status is graded — application → study → interconnection agreement executed → in service — and withdrawal rates and delays differ by ISO (e.g., PJM's withdrawal rates vs. CAISO's chronic delays). FERC's 2023 cluster-study reform also changed what "queue position" means. We read the queue as evidence, weighted by status and corroborated against EIA and filings — never as a naive "committed vs. delivered" flip.

3. The sources

We prioritize primary sources — filings and official records, not just news:

Grid queues	ISO/RTO interconnection (PJM, ERCOT, CAISO, MISO, SPP, ISO-NE, NYISO) — who's in the queue, how many MW, and who withdrew. Note: ERCOT operates outside full FERC jurisdiction; its large-load rules differ (connect-and-manage + the 2025 transparency law).
Financial	SEC EDGAR (full-text) — capex and commitments declared to investors
Energy	EIA — electricity demand by balancing authority (EIA-930), mapped to our regions
Federal	SAM.gov — federal contracts
Context	Global news (multilingual), classified as official vs. press — we don't treat a press announcement as a brick

4. The three seals (how you verify)

Verifiability > trust. Three seals, and we're clear about what each one proves — and what it doesn't:

Seal 1 — the prediction hash. Every call is frozen with a sha256 of the locked fields (region, CFI, components, date, horizon) before the outcome. Recompute and check at /api/v1/ledger/<id>/verify. It proves immutability — that we didn't edit it after the fact. It doesn't prove the call is right; it proves there's no hindsight.

Seal 1b — external witness (new). The root of the entire ledger is submitted to the independent OpenTimestamps calendars — servers we don't control that attest when the ledger contained those calls — and upgrades to a Bitcoin-anchored proof once the next block confirms. Check the live status (pending calendar vs. Bitcoin-confirmed) at /api/v1/ledger/root. The timestamp isn't ours to move — that's what closes the "you could have backdated it" objection.

Seal 2 — validated track record. The proof of physical reality — whether the calls come true — comes from the verdict when each horizon (180 days) matures. Firm sample from ~Dec 2026. See The Index Journal.

Dating without accuracy is just honesty; accuracy without dating is cherry-picking. We need both — that's why we keep them separate.

5. What we show — everything

We show ALL resolved predictions — hits AND misses. Hiding a miss would be cherry-picking. While the validated sample is small (<30), the Journal carries a preliminary seal: you can verify by hash, but the firm accuracy number only holds from ~Dec 2026. We don't claim accuracy we don't yet have.

6. What we do NOT claim (the honest limits)

A "phantom" verdict is a probability read from the signals — not a certainty that nothing will be built.

Behind-the-meter is a known edge case. A data center powered off-grid (dedicated gas, nuclear, or SMR) can be real and absent from the interconnection queue. We flag this and don't treat "no queue counterpart" as proof of phantom.

The bands (68 / 45) are conventional thresholds, not yet empirically calibrated — calibration follows the validated track record (§5). The CFI weights (35/30/20/15) are expert-set priors, not fitted — also revisited against the track record.

We measure formation signals (queue, capex, energy, permit), not brick-by-brick physical confirmation. Where we have satellite imagery, it's a shadow signal, not a verdict.

The predictive track record isn't validated yet (starts ~Dec 2026). Until then, the value is descriptive: the honest picture of real vs. phantom today.

Coverage is honest: if a region isn't in the index, we say "we don't measure it" — we don't invent.

7. Don't trust — verify

Every number links to a checkable source. Recompute the hashes, check the external anchor, read the misses in the Journal. Found something wrong? That's exactly what we want to hear — tell us. The method changes when the evidence changes.