Methodology

1 — Citation verification

Every fact statement in every SpotlightAICore report carries a citation to a specific page or paragraph of a specific source document. Before a report ships, a built-in verifier confirms that each citation actually exists in the source — that the cited page contains the cited language, that the cited document is in the production, and that the citation has not been hallucinated.

How verification works

1. The draft report is parsed. Every citation in the report is extracted into a structured list.
2. For each citation, the verifier opens the cited source document and confirms the cited text is present on the cited page. If the source is a Bates-numbered production, the verifier confirms the Bates range matches.
3. Citations that fail verification are flagged in the QC log and are either corrected (when the underlying fact is supportable from a different cite) or removed from the deliverable (when no valid support exists).
4. Failed-verification rates and resolution actions are recorded in the QC log row for the deliverable, so the audit trail is complete.

Why this matters in court

AI hallucinations of citations are now a well-known professional hazard in legal practice. Mata v. Avianca, 678 F. Supp. 3d 443 (S.D.N.Y. 2023), and the subsequent line of sanction orders against attorneys who filed AI-hallucinated cases have established that vendor-produced AI output requires independent verification before any reliance. SpotlightAICore's citation-verification step is the answer.

2 — Two-AI adversarial review

Trial-tier and Bespoke-tier deliverables go through a second AI model that audits the first model's work. The second model is given the draft report and the source documents and is instructed to find weak inferences, mischaracterized citations, overconfident claims, and arguments where the documentary support is thinner than the report's framing suggests.

How adversarial review works

1. The first model produces the draft report from the source documents, with full citations.
2. Citation verification runs on the draft (see step 1 above).
3. The verified draft is handed to a second model, which is instructed to disagree productively: where is the first model overreaching? Where is the inference too strong for the evidentiary support? Where does a cited document not actually say what the report claims it says?
4. Disagreements are categorized: factual disagreements (the underlying citation is wrong); inferential disagreements (the citation is right but the conclusion drawn from it is overstated); and stylistic disagreements (the framing of an otherwise-valid finding is too confident).
5. Disagreements are surfaced inline in the deliverable with a clear "Model-Disagreement" annotation. The reviewing attorney decides what to do about each disagreement before release: accept the first model's view, accept the second model's view, or escalate for additional source review.

Why this matters under Berkheimer

Berkheimer v. HP Inc., 881 F.3d 1360 (Fed. Cir. 2018), and Aatrix Software, Inc. v. Green Shades Software, Inc., 882 F.3d 1121 (Fed. Cir. 2018), held that the factual basis of expert testimony is itself subject to challenge on summary judgment. A vendor-produced report that surfaces its own internal disagreements — rather than burying them — is materially harder to attack than one that does not. Two-AI adversarial review is the structural answer to the "black-box AI" attack opposing counsel will otherwise mount.

3 — Founder review

No SpotlightAICore deliverable leaves the firm without personal review by the founder, M. David Hoyle — a named inventor on more than 11 U.S. patents with approximately 30 years of federal patent enforcement litigation experience. Trial-tier and Bespoke-tier deliverables are additionally subjected to Two-AI Adversarial Review, where a second AI model audits the first model's output and disagreements are surfaced in the deliverable itself. SpotlightAICore is not a law firm; the founder reviews analytical work product on behalf of the customer's engaged attorney, who retains all professional responsibility for the matter.

What sign-off covers

• The executive summary of the deliverable.
• The disposition of every model-disagreement surfaced by the adversarial review pass (Trial / Bespoke).
• The citation-verification result summary (pass / fail counts and the resolution of each fail).
• Any flagged anomalies in the source documents (suspicious authentication, OCR failures, integrity issues).

What sign-off does not do

The reviewing attorney does not provide legal advice to your firm or your client. The deliverable is analytical input to your strategic judgment, not a substitute for it. SpotlightAICore does not represent parties in litigation and is not a law firm; it is a litigation-intelligence vendor whose deliverables a licensed attorney has personally vouched for.

4 — The QC log

Every SpotlightAICore deliverable ships with a row in the matter's QC log. The QC log is a machine-readable XLSX file that records, for each deliverable, exactly how the work was produced. It is the audit trail.

What the QC log records

• Matter ID, deliverable name, deliverable version.
• Quality model used to produce the draft (Standard / Pro / Premium).
• Whether two-AI adversarial review was applied; if so, the model used for review.
• Citation verification pass count, fail count, and resolution actions.
• Model-disagreements raised by adversarial review, with the disposition decided by the reviewing attorney.
• Reviewing attorney name; co-signing attorney name (if applicable); sign-off timestamp.
• Hash of the final deliverable file at the moment of release.

Why it ships with the deliverable

If opposing counsel ever asks how a SpotlightAICore-produced deliverable was generated — in a deposition of your team, in a motion in limine, in a Daubert challenge, or in a discovery sanctions brief — the QC log answers the question. There is no black box. The methodology is documented, contemporaneous, and inspectable.

What we do not do

• We do not generate facts. Every fact in every deliverable comes from a cited source in your production. If the source does not say it, the deliverable does not say it.
• We do not draw legal conclusions. The deliverables surface contradictions, reconstruct timelines, and isolate the facts that drive the case. The strategic judgment about what to do with those facts is yours.
• We do not use any Google product or service to process client documents. No Gmail, no Google Drive, no Google Docs, no Google Cloud, no Google Workspace. This is a permanent firm rule, not a per-matter accommodation.
• We do not retain client documents after the matter closes. Default 90-day post-delivery destruction; longer retention available on request. See the Confidentiality page for the full retention policy.
• We do not subcontract analysis to third parties. The model runs on dedicated hardware we operate. Citation verification, adversarial review, and founder review all happen in-house.