Skip to content

Review a line of text with a lens

Machine verified A machine ran this on our install and checked the result against the CLI. A test report is linked. Proven against quoxcli 0.5.0 on 2026-11-08.

What would an auditor say about this?

Run a lens over a short piece of text and get its verdict, its findings and what it cannot judge.

What you get: a second opinion on a policy line, a claim or a paragraph, from a named and versioned point of view, before anyone else reads it.

First, see which lenses you have:

quox · talking to Quox

Which lenses are set up in QuoxLens?

▤ AOCL │ L1 identity complete · allowed · 1ms

▤ AOCL │ L3 policy gate complete · proceed · 9ms

◐ LENS │ quoxlens_read action=lenses ✔ LENS │ quoxlens_read action=lenses │ oversized: true, 28 full size kb, 3 top keys · 706ms

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

⛓ AEE │ envelope opened · stream_success

QuoxLens has five lenses set up. All are house lenses of the evaluative kind.

Lens          Adversarial Red Team (v1.0.0)
What it does  Reads an artifact as someone trying to make it do something its author
              never intended. It looks at spoofing, replay, and where untrusted content
              turns into instruction.

Lens          Codebase Security Engineer (v1.0.0)
What it does  Reads code, config and infrastructure as the engineer who gets paged when
              it is exploited. It finds the concrete defect and the attack it enables.

Lens          Enterprise Compliance Auditor (v1.1.0)
What it does  Reads an artifact as the auditor who asks for the record eighteen months
              later. It checks whether a control is enforced server side and leaves a
              durable trace.

Lens          Honesty and Claims
What it does  Listed, but I only have its name and kind.

Lens          Quox Execution Integrity
What it does  Listed, but I only have its name and kind.

I only pulled full details for the first three. Adversarial Red Team is flagged as
unproven. Its eval from 31 July 2026 caught all 9 seeded defects. Its precision was
lower than the baseline run (0.33 against 0.41), and the extra findings are still
pending adjudication.

I can pull the details for the last two if you'd like.

12.4s · 1 tool · 0 approvals

A real turn on our own install (CLI 0.5.0, 8 Nov 2026), replayed at speed. Names changed. Type the same words into quox.
✔ Machine verified against the CLI on 8 Nov 2026 · test report mw7q8u

Quox read your QuoxLens lenses and described what each one looks for, and said plainly which two it had only the names for.

Then give it the text and ask for a review:

quox · talking to Quox

Run one of our QuoxLens lenses over this short policy line and tell me what it concludes: "Backups run every night and a restore is tested once a month."

▤ AOCL │ L1 identity complete · allowed · 0ms

▤ AOCL │ L3 policy gate complete · proceed · 3ms

◐ LENS │ quoxlens_read action=lenses ✔ LENS │ quoxlens_read action=lenses │ oversized: true, 28 full size kb, 3 top keys · 2.0s

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

◐ LENS │ apply ✔ LENS │ apply │ budget_tokens_used_ceiling=[redacted] findings: 2 · 15.6s

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

⛓ AEE │ envelope opened · stream_success

I ran the Enterprise Compliance Auditor lens over your line. It concludes concerns, with
55% confidence.

The lens reads a policy as an auditor who asks for the record eighteen months later. It
found two gaps, and both come from the line stating the controls without pointing to any
evidence:

Severity                       Medium (55% confidence)
Finding                        The line says "a restore is tested once a month", but
                               nothing records that a test happened or how it went.
What an auditor would ask for  Date, operator, which backup was restored, and the pass
                               or fail result.

Severity                       Low (40% confidence)
Finding                        The line says "Backups run every night", but it names no
                               scheduler, job or run record.
What an auditor would ask for  Run history showing each nightly execution and any
                               failures.

The lens itself notes that a one-line statement is not a system of record, so some of
this gap is expected. It judges the restore-test claim to be the one that most needs
evidence behind it.

To make the line audit-ready, name where the evidence lives. For example: "Nightly
backup job logs and monthly restore-test reports are retained for X months in
[location]."

The lens does not cover how long backups are kept, whether the controls can be broken,
or regulatory interpretation. Those are declared blind spots for this lens, so this
result says nothing about them.

If it would help, I can check the live backup record to see whether the claim holds. I
can also run a second lens, such as Honesty and Claims, over the same line.

30.2s · 2 tools · 0 approvals

A real turn on our own install (CLI 0.5.0, 8 Nov 2026), replayed at speed. Names changed. Type the same words into quox.
✔ Machine verified against the CLI on 8 Nov 2026 · test report mw7q9l

Quox picked the Enterprise Compliance Auditor lens, ran it over the line, and reported the verdict, each finding with its severity, and the blind spots the lens declares, so you know what the result does not cover.

lenses
quox tool exec quoxlens_read --args '{"action":"lenses"}' --select 'lenses[*].{name,version}'[{ "name": "Adversarial Red Team", "version": "1.0.0" },{ "name": "Codebase Security Engineer", "version": "1.0.0" },{ "name": "Enterprise Compliance Auditor", "version": "1.1.0" },{ "name": "Honesty and Claims", "version": "1.1.0" },{ "name": "Quox Execution Integrity", "version": "1.0.0" }]
From a real run on our own install. Output reformatted to fit.

Use --summary instead of --select to see each lens’s id and description. You pass the id in the next step.

apply
quox tool exec quoxlens_apply --args '{"lens_id":"enterprise-compliance-auditor","artifact":{"content":"Backups run every night and a restore is tested once a month."}}' --reason "check a backup policy line" --select 'verdict.disposition,verdict.confidence,findings[*].{severity,claim}'{"verdict.disposition": "concerns","verdict.confidence": 55,"findings[*].{severity,claim}": [  { "severity": "low", "claim": "Monthly restore test is asserted with no record of its execution or result." }]}
From a real run on our own install. Output reformatted to fit.

A lens is a model reading the text, so the wording, the confidence and the number of findings can differ from one run to the next. The verdict on this line was “concerns” every time we ran it.

Leave off --select to see the full result, including each finding’s rationale, the quote it rests on, and the lens’s blind_spots.

Quox ran a lens over the text you gave it, under your own login. It returned a reading and changed nothing, so no approval was needed.

The full result names the lens and its version under applied, so you can tell exactly which lens produced a finding.

A review changes nothing, so there is nothing to undo.