Your agent works for you. Pazmo checks you.
Anyone can run a scanner. What a verdict has to answer is who ran it, how much they could actually see, and what they were allowed to conclude from it. Here is how Pazmo answers all three.
How to read this
This page follows the rule it is arguing for: every claim below carries a citation to where Pazmo actually enforces it. The claim that has nothing behind it is marked, not dressed up.
SITUATION
- no product citation — claim about your situation, not about Pazmo
Your code gets opened by someone who is not you.
Nobody schedules the day an outsider reads their code. It arrives attached to something else — a term sheet, a procurement form, an incident review. By the time it happens, the answer is already whatever it is.
- Raising a round, when the diligence checklist arrives
- Selling to an enterprise, when their security review lands
- Shipping the thing you have been demoing
- Handing off code an agent wrote for you
- The week after an incident, when someone asks whether it can happen again
All five ask one question: what is in there, and who says so.
MECHANISM
- STRATEGY §7.3 — wrapped certified OSS + translation layer
- CONCEPTS.md · False Clean
Scanners detect. The model interprets. They never swap jobs.
Detection is mechanical. Certified open-source engines sweep the target and either find a thing or do not. No language model sits in that path, and nothing is sampled: a check either ran across the whole surface, or the result says limited instead of clean.
The model’s job starts after detection and stays narrow — put a finding in plain words, rank it against the others, say what fixing it involves. It reads findings. It does not produce them.
CONSTRAINT
- CONCEPTS.md · False Accusation
- CONCEPTS.md · Checked Absence
A hunch can raise a flag. It cannot set a grade.
Heuristics, pattern smells, and the model’s own suspicion are allowed to put something in front of you. They are not allowed to decide what it is. Anything that moves a verdict traces back to a rule with a named source behind it.
That is arithmetic, not modesty. What Pazmo sells is the credibility of the verdict, so a finding you did not actually have costs more than one you were never shown — it teaches you to discount every other line in the same record.
And where a detector cannot separate a real problem from a benign case that produces the same signal, it stays quiet and records the miss on purpose, so the next person does not read the silence as an oversight and “fix” it back into an accusation.
OUTPUT
- docs/product/compact-human-flow.md
- CONCEPTS.md · Agent Recipe
One record. Two readers.
The person who has to decide and the person who has to fix are not the same person. Writing them two documents is how the two documents stop agreeing.
So it stays one record. The top of it reads without any security background: what state you are in, what moved since last time, what to do first. Underneath, the same findings carry the file, the line, and the fix — the shape an engineer or a coding agent can act on without translating anything.
PROVENANCE
- src/lib/rules/catalog.ts — readinessLenses
- rendering this on the live finding card: not built — this draft is the ask
Every verdict names its source.
A verdict with no source is an opinion with formatting. Each finding carries the rule that produced it, the engine that observed it, the scan it is bound to, and the standard it maps to.
Exposed .git directory
- Rule
- exposure/vcs-metadata
- Engine
- Pazmo passive probe — no code execution
- Bound to
- scan 8f3c… · observed 2026-08-24
- Maps to
- ISO/IEC 27001 evidence · SOC 2 Security
- Evidence
- probe ran and the artifact was present
The mapping row is drawn from lenses that already sit on the rule catalog. Showing them on the live finding card does not exist yet — that is the part this draft is asking about.
ALTERNATIVES
- CONCEPTS.md · Self-report (자기신고)
- /coverage — what is out of scope
Five ways to find out what is in your code.
| Approach | What it costs | What you end up holding |
|---|---|---|
| Do nothing | Nothing, until it costs everything at once | Whatever the other side finds, and they find it first |
| Run open-source scanners yourself | An afternoon, then an afternoon every time | A list in your own hand, graded by the person who wrote the codeself-reported |
| Book a penetration test | A budget line and a scheduling window | A deep point-in-time report from an outside party |
| Adopt a compliance platform | An annual contract and an onboarding project | Continuous control coverage, mostly self-attestedself-reported |
| Pazmo | One input, no signup, about a minute | An outside verdict bound to a commit, and the record of it over time |
These are shapes, not scores. Two of them Pazmo does not replace: a penetration test looks deeper than any passive check, and a compliance platform covers process controls Pazmo only guides. The column that matters is the last one — whoever grades the work is the one being graded.