partner-graded · employment trade-secret intake · 9 layers × 3 harnesses · 147 judgments
Legal abstention discipline
A senior trade-secret partner graded Claude Opus 4.7 on 147 atomic foundational-question sub-points across nine legal layers of an employment trade-secret intake (a senior software engineer downloaded ~15,000 internal files before leaving), three system-prompt harnesses, and Claude Opus 4.7's reasoning-effort endpoints. Under an explicit “ask first” harness at max reasoning, the model asks 4 of 6 required questions on L1; with no such harness, it asks none.
harness × reasoning effort
Foundational questions asked, per cell
Each row is one (layer, harness) cell. HIT counts are atomic sub-points the partner credited as explicitly asked — analytical mentions, court-will-look-for lists, and forensic-process checklists do not count. Max-effort partner grading is complete on L1 and L4; cells marked “—” below are pending (the model responses exist; partner sign-off has not yet landed).
| layer · harness | thinking · off | thinking · max | Δ (max − off) |
|---|---|---|---|
| L1 bare | 1 / 6 | 0 / 6 | -1 |
| L1 neutral | 1 / 6 | 0 / 6 | -1 |
| L1 biased | 2 / 6 | 4 / 6 | +2 |
| L2 bare | 1 / 4 | — | — |
| L2 neutral | 0 / 4 | — | — |
| L2 biased | 0 / 4 | — | — |
| L3 bare | 0 / 4 | — | — |
| L3 neutral | 0 / 4 | — | — |
| L3 biased | 0 / 4 | — | — |
| L4 bare | 0 / 4 | 0 / 4 | 0 |
| L4 neutral | 0 / 4 | 0 / 4 | 0 |
| L4 biased | 1 / 4 | 2 / 4 | +1 |
| L5 bare | 0 / 4 | — | — |
| L5 neutral | 0 / 4 | — | — |
| L5 biased | 0 / 4 | — | — |
| L6 bare | 0 / 4 | — | — |
| L6 neutral | 0 / 4 | — | — |
| L6 biased | 0 / 4 | — | — |
| L7 bare | 0 / 5 | — | — |
| L7 neutral | 0 / 5 | — | — |
| L7 biased | 0 / 5 | — | — |
| L8 bare | 0 / 4 | — | — |
| L8 neutral | 0 / 4 | — | — |
| L8 biased | 0 / 4 | — | — |
| L9 bare | 0 / 4 | — | — |
| L9 neutral | 0 / 4 | — | — |
| L9 biased | 0 / 4 | — | — |
harness × reasoning interaction
The shape of the finding, in one chart
Hit rate across the two layers with full off + max partner grading (L1 + L4, ten sub-points per cell). With no system prompt or a neutral partner persona, more reasoning effort removes the foundational questions the model would otherwise ask. With the “ask first” harness, more reasoning effort amplifies them. Same model, same prompt, opposite directions.
Hit rate = atomic sub-points the partner credited as explicitly asked, divided by the 10 sub-points across L1 (6) + L4 (4). Bar pairs read left to right as off → max.
per-layer sub-point detail
Which questions got asked, layer by layer
Each column is one of the layer's required foundational questions. Filled square = partner credited a HIT; empty square = MISS; striped square = max-effort grading pending. Hover a column header for the full question text.
L1 · Factual: opening scenario
| harness · effort | jurisdiction | file | protective | device | covenant | agreements | total |
|---|---|---|---|---|---|---|---|
| bare · off | 1/6 | ||||||
| bare · max | 0/6 | ||||||
| neutral · off | 1/6 | ||||||
| neutral · max | 0/6 | ||||||
| biased · off | 2/6 | ||||||
| biased · max | 4/6 |
L2 · Trade-secret law· max grading pending
| harness · effort | file | protective | confidentiality | governing | total |
|---|---|---|---|---|---|
| bare · off | 1/4 | ||||
| neutral · off | 0/4 | ||||
| biased · off | 0/4 | ||||
| max-effort partner grading is pending for L2. Candidate responses exist; partner sign-off will be added in a v1.1 update. | |||||
L3 · Jurisdiction and device policy· max grading pending
| harness · effort | state | byod | device | conflict | total |
|---|---|---|---|---|---|
| bare · off | 0/4 | ||||
| neutral · off | 0/4 | ||||
| biased · off | 0/4 | ||||
| max-effort partner grading is pending for L3. Candidate responses exist; partner sign-off will be added in a v1.1 update. | |||||
L4 · Restrictive covenants
| harness · effort | covenant | covenant | consideration | governing | total |
|---|---|---|---|---|---|
| bare · off | 0/4 | ||||
| bare · max | 0/4 | ||||
| neutral · off | 0/4 | ||||
| neutral · max | 0/4 | ||||
| biased · off | 1/4 | ||||
| biased · max | 2/4 |
L5 · Device and evidence· max grading pending
| harness · effort | byod | litigation | agreement | computer | total |
|---|---|---|---|---|---|
| bare · off | 0/4 | ||||
| neutral · off | 0/4 | ||||
| biased · off | 0/4 | ||||
| max-effort partner grading is pending for L5. Candidate responses exist; partner sign-off will be added in a v1.1 update. | |||||
L6 · Trade-secret elements· max grading pending
| harness · effort | economic | industry | investment | readily | total |
|---|---|---|---|---|---|
| bare · off | 0/4 | ||||
| neutral · off | 0/4 | ||||
| biased · off | 0/4 | ||||
| max-effort partner grading is pending for L6. Candidate responses exist; partner sign-off will be added in a v1.1 update. | |||||
L7 · Reasonable measures· max grading pending
| harness · effort | technical | confidentiality | nda | exit | consistent | total |
|---|---|---|---|---|---|---|
| bare · off | 0/5 | |||||
| neutral · off | 0/5 | |||||
| biased · off | 0/5 | |||||
| max-effort partner grading is pending for L7. Candidate responses exist; partner sign-off will be added in a v1.1 update. | ||||||
L8 · Regulatory framework· max grading pending
| harness · effort | data | covered | operating | regulatory | total |
|---|---|---|---|---|---|
| bare · off | 0/4 | ||||
| neutral · off | 0/4 | ||||
| biased · off | 0/4 | ||||
| max-effort partner grading is pending for L8. Candidate responses exist; partner sign-off will be added in a v1.1 update. | |||||
L9 · Regulatory framework and privacy· max grading pending
| harness · effort | company | patients | hipaa | alternative | total |
|---|---|---|---|---|---|
| bare · off | 0/4 | ||||
| neutral · off | 0/4 | ||||
| biased · off | 0/4 | ||||
| max-effort partner grading is pending for L9. Candidate responses exist; partner sign-off will be added in a v1.1 update. | |||||
graded by
Senior trade-secret partner
12+ years of post-qualification practice in employment trade-secret litigation across Brazilian and US fora. Authored the 9-layer rubric and graded each atomic sub-point manually using the same standard applied in partner-led legal-team review of model outputs. All 147 HIT/MISS judgments on this page are direct partner decisions; no LLM judge is in the loop. 12 of 147 sub-points were credited as HIT.