What goes into the score
Only what was knowable before the review week: letters, motions, deadlines. Later outcomes stay out.
Triage
Data science · Author29
The brief
Rank the matters that deserve a partner hour this week.
Sparse intake and filing signals scored so a small firm can spend three intensive slots, not search the whole docket.
8.3×
Better than guessing who to review
Compared with picking names at random from the same list.
13ms
Fast enough to score one matter after the model is trained
4
Practice areas reviewed: standing labels, not one-off noise
1
Dockets define labels
What is already on the file becomes the reference for which practice needs are visible.
2
Intake builds signal
Letters, motions, and deadlines show what a matter is becoming, before the binder is complete.
3
Scores spend three slots
The output is a ranked list: who is worth a partner hour when capacity is three.
Lift
How much better than random review?
If lift is 7×, the top-ranked matters are about seven times more concentrated with true cases than a random list.
Top 10% capture
What does a small worklist find?
If the firm can only review 10% of open matters, how many of the true cases sit in that slice?
Capacity
Why only three?
Intensive hours are scarce. We judge the ranking by who lands in those three slots.
The business translation
A rule is a switch. The model is a dial.
A simple rule can flag every file with one obvious document. The model combines many intake and filing signals into a score, so a firm can decide how broad or precise this week's review list should be.
Instead of asking partners to search the whole docket, start with the matters most likely to need a human hour.
Most stakeholder-friendly result: the top 10% of matters by employment score captures 73% of true employment cases in the held-out set.
Capacity lens
Employment
73%
true cases captured in top 10%
7.3x
random-review lift
Strong filing signature. Best fit for a capacity-limited review list.
Random review would find ~10%. Taller bars mean the ranking concentrates real work into a short list.
- Employment73% · 7.3x
- Commercial contracts51% · 5.1x
- Intellectual property54% · 5.4x
- Family47% · 4.7x
Employment-shaped example. The model (copper) pulls ahead of random (gray) as soon as the team can only review a slice.
— Model - - Random
Practice explorer
Where sparse filings are most useful.
Stronger signals appear when the paperwork is characteristic of the practice. Weaker labels still matter, but they need different expectations, and often more of the file than intake data can provide.
Matter-level explanation
NL-1042
Employment · P=0.81
Recurring signals on the docket: wage claims, EEOC correspondence, termination timeline.
Demand letters2 in 14 days
EEOC chargefiled
Termination windowinside 30 days
Counsel letteron file
Trust
Explain the score before anyone acts on it.
Each score comes with a why, so a partner can sanity-check the flag before spending the hour.
Relative contribution of intake signals. A sanity check before a partner spends the hour.
- Demand-letter cadence92
- EEOC / agency filing78
- Termination window71
- Counsel already on file54
- Days since last filing33
- We only used information available before the review week.
- Scores are a “look first” flag, not a verdict on the case.
- A person still decides. Nothing acts on its own.
How this was built
Optional. The list above is the product; this is for people who want the method.
What we compared
Guessing, biggest-file-first, a simple rule, then a model, judged by who lands in three slots, not by a fancy score.
What we show partners
A readable ranking, plus a why for each name, plus dollars at stake when two files look similar.
Honest limit
Sometimes a simple rule is close enough. We don’t pretend a fancier model always wins.
Stakeholder output
Partner worklist for this week
Start with a partner and ranking lens. Intensive top-3 slots are the hours that actually get used. Click a matter for drivers and a next action.
Tier is a percentile on the shipped intake rule (what ranks the top-3). Chance is a separate estimate for talking about the file. They will not always agree.
| # | Matter | Partner | Tier | Chance | Value | At stake |
|---|---|---|---|---|---|---|
| 1 | Rivera v. Harbor LogisticsNL-1042 · Employment | Avery | Higher | 0.81 | $180k | $146k |
| 2 | Diaz — constructive dischargeNL-1118 · Employment | Ruiz | Higher | 0.74 | $125k | $93k |
| 3 | Patel wage collectiveNL-1102 · Employment | Chen | Higher | 0.77 | $310k | $239k |
| 4 | Keystone Supply — MSA disputeNL-1088 · Commercial contracts | Avery | Higher | 0.64 | $420k | $269k |
| 5 | Atlas freight — indemnityNL-1077 · Commercial contracts | Chen | Higher | 0.55 | $260k | $143k |
| 6 | Northwind mark oppositionNL-1011 · Intellectual property | Avery | Higher | 0.58 | $95k | $55k |
| 7 | Beacon SaaS — terminationNL-1061 · Commercial contracts | Ruiz | Higher | 0.52 | $390k | $203k |
| 8 | Lumen design patentNL-1055 · Intellectual property | Chen | Higher | 0.49 | $140k | $69k |
| 9 | Harbor wordmark — 2(d)NL-1028 · Intellectual property | Ruiz | Mid | 0.44 | $70k | $31k |
Caveats
Triage, not an automated verdict.
- Fictional firm. Northline Counsel is fictional, so this can live on a public site.
- Risk ≠ rescue. A high score means “look first,” not “this matter will be won.”
- Human in the loop. Scoring is batch; action waits on a partner.
Talk about your data
Most visitors want a dashboard. This is the deeper ranking work.
Where data lives, which few numbers would change a decision, and a tracking plan you can actually run.