How EarnIQ works

EarnIQ measures how a company's investment thesis changed after earnings and whether management delivered on measurable promises. It does not issue buy or sell signals.

1. Sources and coverage

Phase 1 uses SEC EDGAR 8-K exhibits, 10-Q/10-K filings and SEC XBRL company facts. Transcripts are excluded unless separately licensed. Consensus, prices and implied moves are displayed only when the applicable vendor data is licensed and a timestamped pre-report snapshot exists.

2. Six equally weighted dimensions

Demand, Pricing, Margins, Guidance, Capital spending and Management confidence each receive -1, 0, +1 or null. Each metric is compared with the prior quarter and the company's own four-quarter average. A change must exceed a noise band equal to 0.5 × the standard deviation of the company's last eight quarterly changes. Fewer than four historical quarters produces null, not zero.

3. Total and vetoes

The raw scores are summed. +3 or more is Strengthened; -2 through +2 is Unchanged; -3 or less is Weakened. One null is scaled to six dimensions. Two or more nulls produce Inconclusive. A guidance cut or withdrawal cannot produce Strengthened. An 8-K Item 4.02 event, going-concern language or a material weakness forces Weakened.

4. Missing data states

Not yet analyzed means the required history has not been processed and no score exists. Inconclusive is reserved for a completed analysis with insufficient or conflicting evidence. Missing evidence is null and is never silently converted to zero.

5. Expectation gap and fiscal periods

Actual revenue and EPS are compared with timestamped pre-report consensus and the company's own eight-quarter surprise range. Guidance midpoint is compared with consensus when definitions match. Fiscal labels use the company's fiscal calendar and display the period-end date. Before comparison, EarnIQ identifies quantified one-time items disclosed by the company. The GAAP figure remains visible; a separate comparable figure is calculated only when the adjustment amount and direction are explicit, with the arithmetic and source shown. Consensus is never compared with GAAP when the estimate is defined on an adjusted basis.

6. Promise ledger and evidence

Only management statements that are attributable, measurable and dated are entered in the promise ledger. Every scored claim must point to a filing location and show its calculation. Short source anchors are used instead of republishing documents.

The AI Read judgment card

Company pages may carry an AI Read: a short judgment written by a language model (currently Google Gemini) that sits above the evidence sections. The name is deliberate: the synthesis is machine-generated, and everything around it is the deterministic machinery on this page.

What the model sees. Only the structured inputs already shown on the page: decisive XBRL changes, the earnings-quality score and trend, reported EPS and revenue, the promise ledger, Form 4 insider aggregates, material 8-K items, risk-factor changes, quantified guidance thresholds and excerpts of the earnings release. It never sees prices, analyst estimates, news or social content.

What it may say. A verdict (strengthened, weakened, mixed or unchanged), one score per rubric dimension, the single sharpest disagreement between sections, and one watch condition. Insider activity is reported as dollar facts, never as intent. It may not issue recommendations, price targets or predictions, and its vocabulary is screened for them.

Verification before publication. Every number in a draft read is checked programmatically against the exact input values it cites; every score must be supported by the direction of the evidence it cites; the verdict must follow mechanically from the dimension scores. A draft that fails is regenerated once, and if it fails again the read is withheld entirely — the page shows the evidence sections with no card. When fewer than two dimensions can be independently supported, the card says so instead of rendering a verdict.

Limits. Verification proves a read is consistent with its cited evidence; it cannot prove the judgment is complete. The inputs are point-in-time snapshots of SEC data and can lag new filings. The AI Read is research synthesis, not investment advice.

Rubric version history

v1.0 — current. Six equal dimensions, company-specific noise bands, explicit null handling, guidance and critical-event vetoes, and promise-delivery-only management confidence when transcripts are unavailable. Historical results are not silently rescored when the rubric changes.

7. Track record and limitations

The track-record page discloses the sample and frozen rules. Vendor definitions can differ from SEC GAAP figures, fiscal calendars require careful normalization, and extraction can be imperfect. Check linked primary sources before making decisions.

Signals are informational and not investment advice. Report a data discrepancy.