Twenty posts in this series argued the paradigm. This one hands it to you as a tool.

What follows is not an essay about evaluating insurer-retained experts. It is the instrument itself — a SKILLS file, in the sense that term carries in AI practice: a self-contained instruction set that a language model can execute. Give your AI this document and a claim file, and it will know what to extract, how to code it, what to compute, and what it may and may not say about the results. Read it yourself and you will know the same things. It is written for both audiences at once, deliberately, because the document that instructs the machine and the document that withstands cross-examination about your methodology should be the same document.

Why an instrument, and why versioned

The ultimate question — is there a substantial likelihood that this expert is biased? — has no bright line. It rests within a range, and each of the four factors contributes some support toward or away from it. A determination like that is only as credible as the process that produced it. If your evaluation method changes case to case, or emerges after you have seen the expert’s report, or cannot explain why one observation is a fact and another is a judgment, then your conclusion is an adjective. The remedy is the same one forensic disciplines apply to themselves: a written instrument, fixed before the evaluation begins, versioned so that every finding traces to the exact instruction set that produced it, with every evaluative claim tethered to a verbatim quotation from the record.

The declaration is the work product. Everything in this file is designed backwards from the paragraphs you will one day sign under penalty of perjury — each module below opens with its declaration paragraph, blanks and all, and then specifies exactly how each blank gets filled.

The three-tier rule

Every field in this instrument carries a tier, and the tier governs what you may say about it in the declaration. Tier 1 — Objective: verifiable facts and arithmetic — dates, amounts, counts, ratios — stated as findings of fact. Tier 2 — Coded: fields where an explicit coding rule constrains discretion, stated as coded observations under the stated rule. Tier 3 — Evaluative: fields requiring interpretive judgment — architecturally separated, labeled as evaluative, and never entered without a verbatim excerpt from the source document in the adjacent field.

The excerpt rule is absolute for Tier 3 and for every Tier 2 bias-indicator flag: no code without a quote. An AI running this file must refuse to populate an evaluative field for which it cannot supply the supporting quotation, and must leave blank — not estimate — any metric the record does not support. Blanks are findings too; Module D explains where they go.

Module A — Factor 1: The Money

Declaration paragraph: Expert receives approximately $______ annually from Insurer, and is retained for approximately ___ claims per year. Expert derives more than ___% of Expert’s income from industry-related sources, including insurers, defense counsel, and claims vendors. In insurance matters, Expert is retained by the insurer side in ___% of engagements, with claimant-side work representing ___%. Expert’s average fee per report is $______.

Blank	Metric	Computation rule	Source records	Tier
“$__ annually from Insurer”	Annualized insurer compensation	Sum of payments in window ÷ years; window stated	Payment ledgers, invoices, 1099s	1
“__ claims per year”	Annualized engagement volume	Engagement count ÷ years	Retention records, claim-number joins	1
“__% of income from industry sources”	Industry-dependence ratio	Industry-source income ÷ total disclosed income; each payor coded insurer / defense counsel / vendor / claimant-side / non-litigation	Testimony, fee disclosures, tax records	1
“retained by the insurer side in __%”	Side-of-caption ratio	Insurer-side engagements ÷ total insurance engagements	Prior-testimony lists, disclosure statements	1
“average fee per report is $__”	Fee-per-review depth proxy	Total compensation ÷ report count	Invoices ÷ report inventory	1

Coding rules. Trace the full payment chain: insurer → vendor → expert → expert’s billing entity. Where payment runs through a vendor, the insurer’s records — not the vendor’s — are where the money data lives; log the chain itself as a field. State the window covered by the production; a window shorter than five years is logged, not accepted. The fee-per-review figure does double duty: it is Factor 1 evidence of volume incentive and Factor 3 evidence of what depth of review the price could possibly buy.

Module B — Factor 2: The Outcomes

Declaration paragraph: Of the ___ claims in which Expert rendered an opinion for Insurer during ____–____, Expert’s opinion supported outright denial in ___ (___%), partial denial in ___ (___%), and underpayment in ___ (___%) — a combined insurer-favorable rate of ___%. Expert supported full payment in only ___ of ___ matters (___%). Of the matters Insurer characterizes as paid, the payment averaged ___% of the amount sought.

The outcome codes (Tier 2)

Code	Meaning
Supported denial	Opinion supported paying nothing
Supported partial denial	Opinion supported denying identifiable components of the claim
Supported underpayment	Opinion supported payment below the record benchmark
Supported full payment	Opinion supported payment meeting or exceeding the record benchmark
Indeterminate	Payment supported, but the record contains no benchmark to test fullness

“Full payment” is a defined term. A payment is full only when measured against a benchmark the record itself supplies — the repair facility’s estimate, the treating physician’s recommended course, the contractor’s scope, the amount claimed. Fullness is computed, never assumed. Where no benchmark exists, the default code is Indeterminate, not Full — and the absence routes to Module D, because the insurer is the party that possesses the records that would test fullness, and an insurer that does not maintain them has missed its opportunity to show its expert’s neutrality in practice.

The token-payment problem. Here is an issue the bar has not yet absorbed, so it is stated plainly. There are three ways an expert can help an insurer: outright denial, partial denial, and underpayment — cost-to-repair shaved, the cheaper medication, the less costly surgery. The third way is invisible in every conventional vocabulary. Claims are “paid” or “denied”; bad-faith law organizes itself around denial; outcome statistics are compiled on denial rates. A payment of 2.5 cents on the dollar — a figure from an actual claim file, not a hypothetical — escapes all of those categories: it is not a denial in the adjuster’s log, not a denial in the insurer’s interrogatory answers, not a denial in the statistics the insurer will one day offer to prove its expert’s balance. It is a 97.5% denial wearing a payment’s clothes. The instrument’s answer is one Tier 1 field: payment ratio — amount paid ÷ amount sought — computed for every claim the insurer characterizes as paid, with amount sought and amount paid captured as separate fields with their own exhibit trails. One division converts the insurer’s best talking point into your exhibit.

Never accept the insurer’s own labels. The corollary rule, and it generalizes: outcome codes are assigned from the underlying numbers, never from the insurer’s paid/denied characterization. The same production that shows an expert found against claimants four times out of five will be presented as “our expert supported payment in one of every five claims.” Both sentences describe the same files. Code the files, not the sentences.

The judicial-criticism inventory (Tier 2/3). A companion table logging every prior judicial criticism of the expert: case, court, year, criticism quoted verbatim. Courts weighing bias have counted prior judicial criticism as part of the constellation, and the inventory feeds Module D as well: an expert’s litigation history is one of the eight record categories a diligent insurer would have collected before retention.

Module C — Factor 3: The Reports

Precondition — the hard gate. This module does not run until a standards-and-practices instrument is loaded: a written statement of what a systematic, neutral evaluation in this expert’s discipline requires, with every element pinned to a source — the expert’s own stated methodology, the vendor’s protocols, the insurer’s evaluation criteria, the field’s published standard, or the governing evidence law. Build it before reading the reports you intend to score. An evaluation without a loaded instrument is not a Factor 3 evaluation; it is an impression. (The paid Workflow gives the instrument’s construction sequence, its five source layers, and the proxy-standard method for disciplines that publish no standard.)

Declaration paragraph: Measured against the ___-element standards-and-practices instrument described above, Expert’s report deviated in ___ respects. Each deviation is identified below, stating what the applicable standard or practice requires, what the report did, and the page of the record establishing each.

Each deviation is a completed finding, not a tally. The unit of analysis is a sentence of the form: the standard requires X; the report did Y; the deviation is Z — with the excerpt pair attached. The count is merely the summary of the findings.

The ten deviation categories (Tier 2, each with mandatory Tier 3 narrative and excerpt)

Methodological absence — no stated methodology, no standard referenced, no tools or procedures described, no preservation protocol
Non-neutral assignment framing — the question as framed directs the answer; conclusion-driven reasoning; omission of the presenting facts a neutral investigation would begin from
Boilerplate and template artifacts — cookie-cutter structure, formulaic language recurring across different claimants, copy-forward errors
Omitted and overlooked evidence — witnesses not interviewed, contemporaneous records not obtained or not addressed, physical evidence unexamined, context ignored
Flawed principles — asserted indicators the evidence does not show; stated principles that misstate what the science or discipline actually requires
Experiential opinion without support — interpretive conclusions resting on the expert’s say-so, with no testing, no literature, no empirical basis
Certainty inflation — categorical conclusions with no error rate, no limitations, no acknowledgment of ambiguity; confidence disproportionate to the evidentiary strength
Verification foreclosed — selective documentation, no preservation, no replicability; an investigation conducted so that no one could check it
Authentication and preservation failures — undated or unauthenticated exhibits; no preservation instruction to custodians
Unexplained opinion shifts — a conclusion or restriction that changes across successive reports or addenda, directionally and without stated reason; before/after excerpt pair mandatory

Per-report screening fields (Tier 1/2), across the whole corpus: methodology identified (Y/N); alternative hypotheses addressed (Y/N); standards cited (Y/N); literature cited (Y/N); examination performed vs. desk review; confidence coded Unqualified / Hedged / Inconclusive. At corpus scale, the distribution of these fields is itself a finding — including the pattern of what the expert never examines.

Module D — Factor 4: The Measures

Declaration paragraph: Of the eight categories of records an insurer could maintain to evaluate the neutrality and reliability of its retained experts, Insurer produced documents reflecting ___. Insurer produced no records of its approval or retention evaluation of Expert, no compensation or volume tracking, no analysis of Expert’s outcomes, and no record of the basis for Expert’s claimed expertise. In response to Request No. ___, seeking Insurer’s own records concerning its own expert, Insurer objected that production would be unduly burdensome.

The gap log (Tier 1/2). One row per category, coded Produced / Partially produced / Objected / Silent, with the objection quoted verbatim. The eight categories:

Approval and retention records for the expert
Years of compensation and volume data
The share of the expert’s work performed for the industry
Other insurers or vendors the expert works for
Pattern-and-practice history of the expert’s outcomes
The basis for the expert’s claimed expertise
Complaints — internal and regulatory
Litigation history, including prior judicial criticism

Why the blanks score. This module is deliberately built from negative space. An insurer that has actually taken reasonable measures possesses these records, because collecting and monitoring them is what reasonable measures means. An insurer that objects that producing its own records about its own expert is too burdensome has, by that objection, described its practices. Every category coded Objected or Silent is an affirmative finding, and every blank left in Modules A and B for want of production lands here. Nothing the insurer does in discovery is neutral with respect to this instrument: produced records populate the metrics; withheld records populate the gap log.

The cumulative paragraph

Each of the foregoing factors independently supports, and together they compellingly support, a substantial likelihood that Expert’s opinions reflect bias rather than independent judgment: a financial relationship of $___ per year and ___% industry dependence (Factor 1); an insurer-favorable outcome rate of ___% with full payment supported in only ___% of matters (Factor 2); ___ deviations from the standards and practices of Expert’s own discipline (Factor 3), including ______, ______, and ______; and an insurer that maintained none of the records that would have tested — or could have vindicated — its expert’s neutrality (Factor 4).

There is no threshold sentence in this paragraph, by design. The determination rests within a range; the instrument’s job is to state, with an exhibit under every number, where in the range this expert sits. Courts have credited constellations weaker than the one this instrument is built to document — the paid Workflow maps the decided benchmarks and where your numbers fall against them.

Protocol rules

Human review is part of the instrument. Every AI-populated row is confirmed, entry by entry, by the attorney who will sign the declaration. Divergences between the AI’s extraction and your review are signals — a systematic divergence means either the coding rule needs refinement or the field captures a genuine ambiguity worth noting. The final work product is your considered judgment; the AI is an organizational and consistency aid, and the declaration says so transparently.

Version control. This file is v1.0. Any amendment — a new code, a refined rule, an added field — increments the version and is logged with its rationale. Every evaluation records the version that produced it. When opposing counsel asks whether your categories were reverse-engineered from your conclusions, the version log is the answer.

Get the operating manual. The paid companion — the Evaluation Workflow — runs this instrument stage by stage: the sequence from intake to pattern-and-practice exhibit, the five-layer standards stack with the discovery decisions that reach each layer, the construction method for the standards-and-practices instrument, the calibration table of decided benchmarks, and the cross-document audit that joins expert reports, coverage letters, and contemporaneous third-party records into a single relational record.

Get the Evaluation Workflow → Run the Master Checklist first → Read the series recap →

The sequence that produces the record this instrument runs on: The Master Checklist. The facial screen behind Module C’s categories: The Flawed-Report Checklist. The doctrine behind Module D’s gap log: Reasonable Measures. Where the burden shifts once the modules are populated: Who Has to Prove What?

Distilled from the project’s four-factor evaluation modules published in this series and the project’s extraction-system architecture. Seminal authority — Demer v. IBM Corp. LTD Plan, 835 F.3d 893 (9th Cir. 2016). Authorities named here are treated in full, with citations, in the paid module kits and the forthcoming treatise. The instrument is a working method that counsel must adapt and verify before use. Educational and informational only; not legal advice.

The SKILLS File: An Instrument Your AI Can Run Against Any Insurer-Retained Expert