AI Adoption GuideGovernmentEnforce
Evidence sufficiency assessor
LLM reviews assembled case evidence against legal threshold requirements and identifies evidentiary gaps before formal case filing.
Government processPlanFundAuthorizeDeliverInspectEnforceReportClose
By Don, DoneThat’s AI coach · updated
The output is a cited gap note, not a filing bar
A sufficiency assessor compares the assembled case file to the legal threshold that must be met before a formal filing. The product of that pass is a gap note: each missing element named, each requirement tied to the clause that actually requires it. If the file already meets the checklist, the note stays empty. Empty is a result, not a skipped run.
The model does not file. It does not hold. It does not score the odds of a later judgment. Counsel reads the note and decides whether to cure, to wait, or to proceed.
Treat the assessor as a pre-filing quality check on the record you already have, not as a substitute for charging discretion. Systems that store the packet (case files in Accela, document sets in Microsoft, authorities pulled from LexisNexis or Westlaw) are inputs. They do not change what a complete threshold looks like in your statute, ordinance, or charging manual.
Load the file and the threshold checklist together
Do not ask the model to "see if we have enough." Load two artifacts in the same pass.
First, the assembled file: inspection notes, photographs with capture dates, notices and proofs of service, permits or permit denials, lab or sampling records where the violation type requires them, correspondence, prior orders, and any forensic work already done on authenticity. If an inspection report auto-drafter produced the inspection narrative, load that narrative as one exhibit, not as proof that every element is already met.
Second, the threshold checklist for this filing type. Pull the elements from the charging statute, the ordinance, the administrative rule, and the office filing manual. Each checklist row should name the element, the clause that requires it, and the form of proof the office will accept (sworn inspection, certified mail return, chain of custody, ownership extract). Authorities from LexisNexis or Westlaw belong here as the text of the requirement, not as a research memo.
Run the comparison row by row. For each element, the model should return one of three states: present and identified by exhibit, missing, or present but incomplete (wrong date range, unsigned, served on the wrong party). Incomplete is a gap. Do not collapse it into a vague "weak file" comment.
Keep the checklist frozen for the run. If counsel later amends the charging theory, reload both the file and the new checklist. Mixing an old packet with a new theory is how invented elements appear.
Write the gap so counsel can cure or hold
A usable gap note has four parts: the missing or incomplete element, the clause that requires it, the exhibit that was checked (or the statement that none was found), and the cure that would close the row. Counsel should be able to assign the cure without re-reading the statute.
One illustration. A housing enforcement unit is preparing a civil complaint for occupancy without a certificate. The checklist requires: (1) proof of current ownership or control, (2) an inspection that observed occupancy and the missing certificate, (3) a notice of violation that recited the certificate requirement and a cure date, (4) proof that the notice was served on the owner of record, and (5) a re-inspection after the cure date showing the condition continued. The file contains a dated inspection with photographs of occupied rooms, a notice, and a postal receipt. The assessor marks ownership as missing because the packet has a tenant name on the inspection header and no tax-roll extract or deed. It marks service as incomplete because the receipt is addressed to the occupant at the unit, not to the owner of record at the mailing address on the tax roll. Each gap cites the ordinance section that requires notice to the owner and the filing-manual row that requires proof of service on that person. The occupancy observations stay listed as present. The model does not add a "willful" element the ordinance never required for this civil count.
That is the whole example. No win rate, no predicted judgment. Counsel can pull the ownership extract and re-serve, or counsel can hold the filing until those rows close.
If document forensics engine work is already in the packet (questioned signatures, metadata on photos), treat those findings as exhibits on authenticity. They do not fill an element the checklist never asked for.
Leave the note empty when the checklist is met
When every row is present in an accepted form, return nothing. Do not congratulate the file. Do not summarize strengths. Do not invent residual risk language ("consider corroboration") unless a checklist row actually requires corroboration.
Empty output is how the office knows the pass finished and found no gap. A paragraph of optional suggestions is not empty. It trains reviewers to ignore the note and re-do the review by hand.
If one row is close but not met, write that row. Do not bury it under a narrative of everything that is already good.
Failure modes that look like legal work
A gap with no clause is not a gap. If the model cannot point to the statute, ordinance, rule, or filing-manual row, drop the line or send it back as a checklist defect. "It would be stronger with a neighbor affidavit" is advocacy, not a threshold miss, unless the office checklist actually requires a corroborating witness for that count.
Treating the note as a filing bar is the second failure. The assessor flags quality. It does not withhold charging authority. A complete file can still be a poor charging choice. An incomplete file can still be filed if counsel accepts the risk and documents that choice. Do not wire the note to a lock on the complaint template.
Inventing an element the statute never required is the third. Models trained on criminal complaints will smuggle mens rea, expert reports, or chain of custody into a strict-liability civil count. Models trained on environmental dockets will demand sampling where the ordinance is a visual housing standard. Bind the run to the checklist you loaded. If a suggested element is not on that list, it does not belong in the gap note.
Watch also for silent substitution: swapping a related exhibit for the required one (a work order for a permit; a photo of a posted notice for proof of mailed service). Name the required form of proof in the gap line so the substitution is visible.
After the pass, counsel still files or holds
Counsel reads the note, assigns cures, or proceeds. Penalty numbers and demand language are a later pass; a penalty calibration engine does not belong in the sufficiency checklist unless the filing type requires a calculated amount as an element of the pleading. Notice text is also downstream: an enforcement notice drafter can use a closed sufficiency pass as confirmation that the factual predicates for the notice already exist, not as permission to draft around a missing element.
Record the checklist version and the file snapshot used for the run. If the packet grows, run again. Do not treat last week's empty note as clearance for this week's exhibits.
Vendors in this stack (LexisNexis, Westlaw, Microsoft, Accela) hold authorities, drafts, and case records. None of them is the threshold. The threshold is the clause you loaded. The quality product is the cited gap, or silence when there is none.
Is this worth automating for you?
Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.
DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.
Measure the baseline first