Calibration & honesty
How the score works, and what we won't claim until we can prove it.
If a vendor is going to touch your revenue, you should see the method and the failure modes, not a demo. Here is how the score is built, how we'd measure whether it's right, and what we will and won't put a number on today.
The method
- Every recorded intake call is transcribed.
- The transcript is scored 0-100 against a fixed rubric: the same rubric on every call, so scores don't drift from one review to the next.
- A call is flagged as a signable case that walked when it scores above the threshold, wasn't converted, and is still inside the callback window.
- Every flag carries the transcript evidence behind it, so you can check the call yourself.
How we'd measure whether it's right
Two numbers matter. Precision: of the calls we flag, the share that really were signable. Recall: of the signable calls that truly walked, the share we caught. Both only mean something against a named test set: how many calls, whether they're synthetic, historical, or from a real firm, and dated.
We're not going to print a precision or recall percentage until we can name the corpus it came from. A number without its test set is the kind of thing you've been pitched before. When the corpus is documented, the figures and the corpus go here together.
Where a model like this gets it wrong
Two failure modes, and we'll show real examples of each once the corpus is documented:
- Missed flag: a signable caller the model scored too low, usually when the caller downplays the injury or the liability language is ambiguous.
- False flag: a call we flagged that wasn't signable: for example, property-damage-only, or a caller who already had a lawyer.
A human at your firm approves every callback, so a false flag costs a moment of a reviewer's time, not a wrong message to a caller.
How we express confidence
Every call maps to one of five plain-English tiers. We use the same five definitions on every call, so a Tier 4 means the same thing in January as it does in June.
| Tier | What we call it | What it means | Likelihood |
|---|---|---|---|
| 5 | Very likely signable | Clear liability and injury, no disqualifier heard. | 80-95% |
| 4 | Likely signable | Strong indicators, one minor open question. | 55-80% |
| 3 | Roughly even | Genuinely mixed signals. | 45-55% |
| 2 | Unlikely signable | Weak indicators, or a likely disqualifier. | 20-45% |
| 1 | Very unlikely | Clear disqualifier: prior counsel, no injury, or property damage only. | 5-20% |
We state the transcript moment behind each judgment, and we never mix a confidence level with a likelihood in the same sentence. A call is flagged for follow-up at Tier 4 and above.
How we estimate missed signable case value (for the $25,000 find-it-free guarantee)
How we estimate missed signable case value: we count the signable cases our model flags that didn't sign, then multiply by your firm's own average fee per signed case for that case type. Where you haven't given us your average fee, we substitute a named, sourced benchmark (e.g., auto soft-tissue ~$16,000; serious injuries $55,000+) and label every substituted figure. Estimates are estimates, not a promise of recovered fees; our model's precision and recall are published on this page.
The guarantee triggers on estimated value identifiedin your own calls, never on any revenue you recover. It's a promise about what the audit finds, not a promise about outcomes.