Skip to content

Read one lens in full

Machine verified A machine ran this on our install and checked the result against the CLI. A test report is linked. Proven against quoxcli 0.6.0 on 2026-11-11.

What is this lens, and has it earned its place?

Every lens says what it looks for. Its stored results say whether it actually helps.

What you get: one lens’s version and kind, its stored evaluation results, and whether it beat a plain review without the lens.

quox · talking to Quox

About the len called "Adversarial Red Team": can you show me one Lens in full? I'd like to know how many results there are and the version.

▤ AOCL │ L1 identity complete · allowed · 1ms

▤ AOCL │ L3 policy gate complete · proceed · 6ms

◐ LENS │ quoxlens_read action=lenses ✔ LENS │ quoxlens_read action=lenses │ oversized: true, 28 full size kb, 3 top keys · 2.1s

⛓ AEE │ envelope opened · ops.tool · quoxlens_read ok

◐ LENS │ quoxlens_read action=lens ✔ LENS │ quoxlens_read action=lens │ matched: false, 17 top keys · 2.8s

⛓ AEE │ envelope opened · ops.tool · quoxlens_read ok

◐ LENS │ quoxlens_read action=lens ✔ LENS │ quoxlens_read action=lens │ [ { "file": "adversarial-red-team-sonnet.json", "lens_id": "adversarial-red-team", "lens_version": "1.0.0", "model": "s… · 1.6s

⛓ AEE │ envelope opened · ops.tool · quoxlens_read ok

⛓ AEE │ envelope opened · stream_success

I've read the Adversarial Red Team lens (id  adversarial-red-team ) in full.

Version: 1.0.0. It's a house lens of kind evaluative, and it's the only version with a
result on file.

Results: There is one stored result file,  adversarial-red-team-sonnet.json . It's a
Sonnet run from 2026-07-31 against lens version 1.0.0, so it matches the current
version. That run covered 12 artifacts and has these totals:

                   │ Findings │ True positive │ False positive │ Defects found
───────────────────┼──────────┼───────────────┼────────────────┼──────────────
Lens               │ 27       │ 9             │ 18             │ 9 of 9
Baseline (no lens) │ 22       │ 9             │ 13             │ 9 of 9

Measured lift: The lens did not beat the baseline.

• Recall: both found all 9 defects, so the lift is 0.
• Precision: the lens scored 0.33 against 0.41 for the baseline, a lift of about -0.076.
• F1: the lens scored 0.50 against 0.58, a lift of about -0.081.
• Class accuracy: the lens got 5 of 9 right (0.56), and the baseline got none.

Status: The lens is marked unproven, both as declared and as measured. Adjudication is
provisional. None of the extra findings have been adjudicated yet (18 pending for the
lens, 13 for the baseline). The precision numbers may shift once those are reviewed.

By "results" I took you to mean the stored evaluation runs. If you meant the findings
count, it's 27 for the lens. I can also pull the lens's finding classes, blind spots or
full body if you'd like.

21.2s · 3 tools · 0 approvals

A real turn on our own install (CLI 0.6.0, 10 Nov 2026), replayed at speed. Names changed. Type the same words into quox.
✔ Machine verified against the CLI on 10 Nov 2026 · test report mwaxiq

Quox read the lens list and then the Adversarial Red Team lens itself. It reported version 1.0.0, one stored evaluation run, and that the lens did not beat the baseline on that run (same recall, lower precision), so it stays marked unproven. It also said which meaning of “results” it took and gave the other one. Quox did not apply the lens to anything or change it.

The same read tool Quox used runs from the terminal, by the lens’s id:

tool exec quoxlens_read
quox tool exec quoxlens_read --args '{"action":"lens","id":"adversarial-red-team"}' --select 'id,name,version,kind,unproven,count(results)' --reason "read one lens"{"id": "adversarial-red-team","name": "Adversarial Red Team","version": "1.1.1","kind": "evaluative","unproven": true,"count(results)": 1}
From the same scratch workspace, read after the turn above.

The lens has been updated to 1.1.1 since the turn above, when it was 1.0.0, so its one stored result was measured against the earlier version. It is still marked unproven. Drop --select to read every field, including the finding classes and blind spots.

Quox and the terminal read the same lens. Reading a lens changes nothing; the unproven flag stays until a measured run shows it helps.