The Experiment
We took terms from Black's Law Dictionary and cross-jurisdictional legal terminology — terms that carry different legal consequences despite surface similarity — and embedded them using nomic-embed-text (768 dimensions). Then we applied PCA at multiple compression levels and measured pairwise cosine similarity.
The question: at what dimensionality do legally distinct terms become indistinguishable to the embedding?
nomic-embed-text (768d), PCA-compressed to 16d. Variance explained at 16d: 63.5%. Model: nomic-embed-text v1.5.
Results: Two Failure Modes
PCA compression to 16 dimensions doesn't just lose information — it warps the space in two distinct ways:
- Collapse (positive delta): Terms that are already similar become indistinguishable. Coercive contempt and punitive contempt, which determine whether you get a trial, merge into one point.
- Distortion (negative delta): Terms that are meaningfully related become artificially distant. German Sorgerecht and Swedish gemensam vårdnad, which both mean "shared parental care," go from 0.640 similarity to 0.184.
Both failure modes are catastrophic for retrieval. The first means you can't distinguish concepts that produce different legal outcomes. The second means you can't find related concepts across jurisdictions.
Term Groups
1. Custody & Parental Rights
These terms produce radically different legal outcomes but describe overlapping concepts. In Illinois, the shift from "custody" to "allocation" in 2016 inverted the burden of proof.
| Term A | Term B | 768d | 16d | Delta | Legal Consequence |
|---|---|---|---|---|---|
| custody | allocation | 0.720 | 0.446 | -0.273 | Different burden of proof — PCA makes them more different than they are |
| guardianship | custody | 0.762 | 0.519 | -0.243 | Guardian protects; custodian possesses — distorted apart |
| visitation | parenting time | 0.767 | 0.758 | -0.008 | Nearly identical — these survive compression (both are English euphemisms) |
| joint custody | sole custody | 0.836 | 0.932 | +0.096 | Opposite outcomes collapse together. One preserves, one eliminates. |
2. Forced Heirship — The Missing Words
Cross-jurisdictional forced heirship terms that have no English equivalent. The big finding: legitieme portie, yuryubun, iryubun, and laglott all collapse toward each other at 16d — losing the tradition-specific legal distinctions.
| Term A | Term B | 768d | 16d | Delta | Legal Consequence |
|---|---|---|---|---|---|
| legitieme portie | yuryubun | 0.763 | 0.890 | +0.127 | Dutch vs. Korean — collapsed to near-identical |
| legitieme portie | iryubun | 0.757 | 0.884 | +0.127 | Dutch vs. Japanese — collapsed |
| legitieme portie | laglott | 0.762 | 0.872 | +0.110 | Dutch vs. Swedish — collapsed |
| Pflichtteil | elective share | 0.709 | 0.418 | -0.291 | Pflichtteil protects children; elective share protects spouse only — distorted apart |
| Pflichtteil | réserve héréditaire | 0.719 | 0.492 | -0.227 | Both protect children — but PCA destroys the relationship |
| faraid | intestate succession | 0.554 | 0.026 | -0.528 | Divine mandate vs. legislative default — rendered unrelated |
| testamentary freedom | forced heirship | 0.700 | 0.536 | -0.163 | Opposite concepts — PCA partially preserves the distinction but distorts it |
3. Contempt & Enforcement
The distinction between coercive and punitive contempt determines whether you get a jury trial. This is a constitutional rights question.
| Term A | Term B | 768d | 16d | Delta | Legal Consequence |
|---|---|---|---|---|---|
| coercive contempt | punitive contempt | 0.784 | 0.913 | +0.129 | Collapsed. Determines whether you get a trial or indefinite imprisonment. |
| civil contempt | criminal contempt | 0.908 | 0.959 | +0.050 | Already similar at 768d — at 16d, nearly identical (0.96) |
| default | contumacy | 0.569 | 0.202 | -0.367 | Neutral omission vs. willful defiance — distorted apart |
| ability to pay | willful refusal | 0.638 | 0.317 | -0.322 | Turner v. Rogers hinges on this distinction — PCA destroys it |
4. Death & Status
| Term A | Term B | 768d | 16d | Delta | Legal Consequence |
|---|---|---|---|---|---|
| c.d. minima | c.d. maxima | 0.911 | 0.970 | +0.058 | Loss of family status vs. total legal death — collapsed |
| civil death | death | 0.670 | 0.637 | -0.033 | Legal fiction vs. biological fact — mostly preserved |
5. Financial Death
| Term A | Term B | 768d | 16d | Delta | Legal Consequence |
|---|---|---|---|---|---|
| nondischargeable | insolvency | 0.703 | 0.814 | +0.112 | Permanent debt obligation merges with its opposite (debt relief) |
| domestic support obligation | discharge | 0.655 | 0.288 | -0.367 | DSOs cannot be discharged — PCA destroys this critical relationship |
| discharge | nondischargeable | 0.707 | 0.716 | +0.009 | Opposite outcomes — surprisingly preserved (shared vocabulary) |
| Restschuldbefreiung | discharge | 0.725 | 0.632 | -0.093 | German "remaining-debt-liberation" drifts from English equivalent |
6. Cross-Jurisdictional Parenting Terms
| Term A | Term B | 768d | 16d | Delta | Legal Consequence |
|---|---|---|---|---|---|
| Sorgerecht | gemensam vårdnad | 0.640 | 0.184 | -0.456 | Worst distortion. German and Swedish "shared care" become nearly unrelated. |
| hadanah | wilayah | 0.722 | 0.630 | -0.092 | Islamic physical custody vs. legal authority — distorted apart |
| umgängessabotage | parental alienation | 0.719 | 0.689 | -0.030 | Swedish statutory term vs. English clinical concept — mostly preserved |
| parental rights | termination of parental rights | 0.728 | 0.565 | -0.163 | A right and its destruction — PCA weakens the connection |
7. Control Group — Inter-Domain Terms
| Term A | Term B | 768d | 16d | Delta | Result |
|---|---|---|---|---|---|
| medical malpractice | custody | 0.560 | -0.168 | -0.728 | Different domains — separation INCREASED (as expected) |
| patent infringement | forced heirship | 0.565 | 0.168 | -0.397 | Different domains — separation increased |
The Two Failure Modes Quantified
| Metric | Intra-Domain | Inter-Domain |
|---|---|---|
| Mean similarity delta at 16d | -0.307 | -0.634 |
| Median similarity delta | -0.302 | -0.640 |
| Max positive delta (collapse) | +0.140 | -0.023 |
Inter-domain pairs never collapse (max delta = -0.023). But intra-domain pairs collapse by up to +0.140. This is exactly what MTEB misses: inter-domain discrimination is preserved; intra-domain discrimination is destroyed.