Acquit.ai
← Back to Insights Services Get Your Acquit Score →
Methodology

The Disagreement
Is the Feature

Single-model AI scoring gives you one perspective on your case. One perspective has blind spots. Your opponent's attorney will find those blind spots. Multi-model adversarial scoring finds them first.

One model. One perspective.
One set of blind spots.

Every AI legal tool on the market runs a single model against your case documents. That model has architectural preferences: ways it anchors on evidence, weights risk, and frames recommendations. Those preferences create systematic blind spots.

Single-Model Scoring

One analysis architecture produces one risk ranking. It sees what it's designed to see, and misses what it's not.

  • One perspective on which risk is #1
  • Systematic bias toward one evidence dimension
  • No way to know what it's NOT seeing
  • Confidence scores without validation

Adversarial Consensus

Multiple independent analyses surface what each model sees, and what each one misses. The gaps become investigation targets.

  • Multiple perspectives, independently generated
  • Consensus findings are high-confidence
  • Divergences reveal exactly where to look harder
  • Validated intelligence, not single-source opinion

"If three independent analyses all flag the same risk, it's real. If they disagree about which risk is #1, that's where the case will be decided."

We ran the same case through
three models. They disagreed.

While validating ARPN scoring against a real appellate case (a multi-million dollar Texas LLC dispute), we ran identical case documents through three AI models with different architectures. Each scored independently, with no access to the others' results.

Where All Three Agreed

Punitive damages exposure is a top-3 risk
HIGH CONFIDENCE · consensus
Buyout mechanism failure is significant
HIGH CONFIDENCE · consensus
Expert testimony weakened the position
HIGH CONFIDENCE · consensus
VS

Where They Diverged

Different #1 risk identified by each model
INVESTIGATION TARGET · divergence
Only some models scored opponent vulnerabilities
INVESTIGATION TARGET · divergence
Behavioral amplifier weights varied significantly
INVESTIGATION TARGET · divergence

The consensus findings confirmed what we expected. The divergences revealed what no single model would have found alone, and pointed to the exact dimensions where the case outcome was most uncertain.

Different models anchor on
different evidence dimensions.

AI models aren't interchangeable. Different architectures (pattern-matching, chain-of-thought reasoning, frontier reasoning) develop different analytical preferences. Those preferences determine which risks they see first.

Dimension Process-Focused Evidence-Focused Outcome-Focused
Anchoring Prompt instructions Strongest evidence signal What actually happened
Top Risk What went wrong procedurally What a jury would react to What caused the most damage
Red Team May not flip perspective Finds the kill shots Explains why failures occurred
Strength Reliable failure identification Offensive opportunity discovery Causal analysis
Blind Spot Misses offensive opportunities May overweight dramatic facts May over-index on realized outcomes

"A process-focused model tells you what went wrong. An evidence-focused model tells you what a jury will remember. An outcome-focused model tells you why the damage was worse than expected. You need all three."

Five steps to validated
case intelligence.

Adversarial consensus scoring isn't just running more models. It's a structured methodology for extracting intelligence from their disagreements.

01

Independent Scoring

Run identical case documents through multiple AI models with the same scoring prompt. Each model scores independently with no access to the others' results. No contamination. No anchoring on another model's output.

Key principle: Independence is non-negotiable. If Model B sees Model A's output, it will anchor on those scores rather than developing its own perspective. The value comes from genuine disagreement, not echo chambers.
02

Consensus Layer

Identify failure modes that all models flagged. If three different architectures, each with different analytical preferences, independently identify the same risk, that risk is real. The minimum score across models becomes the confidence floor.

What this gives you: A confirmed risk profile. Not one model's opinion, but verified intelligence from independent analysis. These are the findings you can present with high confidence.
03

Divergence Layer

Identify where models disagree: different #1 risks, different severity scores, different perspectives on the same evidence. Flag each divergence with the specific dimension of disagreement: is it severity? Detectability? Behavioral amplification?

This is the core insight. Divergences aren't errors. They're investigation targets. Each disagreement points to a dimension of the case where the outcome is genuinely uncertain. That uncertainty is exactly where opposing counsel will probe.
04

Divergence Investigation

Each divergence becomes a specific research question. Instead of asking "what are the risks?" (which is vague), you ask targeted questions driven by the disagreement.

Example: "One analysis scored detectability at 8 (undetectable until damage done). Another scored it at 5 (partially detectable). Which assessment is more accurate for this jurisdiction and this fact pattern?"

This targeted investigation surfaces case-specific intelligence that no single model, and no generic analysis, would trigger.
05

Human Resolution

The attorney or analyst resolves each divergence based on what no AI model has: jurisdictional knowledge, relationship context, strategic intent, and courtroom experience.

Resolution inputs: How does this court weigh this type of evidence? Was this outcome genuinely surprising, or foreseeable? Are we optimizing for appeal, settlement, or trial? The human resolves. The AI surfaces what needs resolving.

The Acquit Score™.

Case intelligence, scored. One number on a 1-1000 scale that captures every dimension of your case: severity, likelihood, detectability, behavioral dynamics, and cascade effects. Validated by adversarial AI consensus. Calibrated against 201 tracked predictions at 91% accuracy.

The Acquit Score™
ARPN Risk Analysis + Behavioral Modeling + Adversarial Consensus + Human Resolution

Every Acquit Score comes with a severity tier (CRITICAL / HIGH / MODERATE / LOW) and a confidence band. Narrow bands mean high model consensus. Wide bands flag the dimensions where your case is most uncertain, and most important. The components are open methodology. The unified score is what your attorney acts on.

"Every AI legal tool runs one model and gives you a score. Acquit.ai runs multiple independent analyses, makes the disagreement the feature, and delivers one validated number: your Acquit Score."

What this means for your case.

For Litigators

Three independent analyses red-teamed your case. Here's what they agree on: your confirmed risk profile. Here's where they disagree. That's where you need to look hardest.

  • Confirmed risks vs. uncertain dimensions
  • Specific investigation questions, not generic recommendations
  • Confidence bounds on every Acquit Score™ finding

For Law Firms

Single-model scoring gives you one perspective. Adversarial consensus gives you multiple, and highlights the gaps between them. Those gaps are the blind spots opposing counsel will exploit.

  • Defensible methodology, not a black box
  • Published framework, peer-reviewable process
  • 91% Acquit Score™ calibration accuracy across 200+ predictions

Book a consultation to see adversarial
consensus scoring on your case.

Free 30-minute scope call. We'll assess whether multi-model analysis fits your case.

Read the ARPN Framework · Read Sentient Analysis · View Services

Join the waitlist for early access to the Acquit Score platform.