How it works

A pipeline you can explain to your auditor.

This page is the long answer to the question every technical evaluator asks: what is actually happening between "upload" and "score"? Five stages, all of them named, none of them magic.
  1. 01Chunk
  2. 02Enumerate
  3. 03Merge
  4. 04Score
  5. 05Persist
§ 01The pipeline

Five stages, one deterministic path.

Input01 Chunk02 Enumerate03 Merge04 Score05 PersistC1C2C3C4C5ISO/IEC 27001 · 93 controlsSOC 2 · 61 criteriaGDPR · 36 obligationsNIST CSF 2.0 · 106 subcategoriesDORA · 62 requirements67.763.979.261.863.7Never averagedImmutable
  1. Stage 01

    Chunk

    Documents are split once, snapping to section boundaries, with overlap so nothing falls between the cracks.

  2. Stage 02

    Enumerate

    One pass per framework under bounded concurrency. Every control in the catalog is assessed; none can be skipped.

    • ISO/IEC 2700193 controls
    • SOC 261 criteria
    • GDPR36 obligations
    • NIST CSF 2.0106 subcategories
    • DORA62 requirements
  3. Stage 03

    Merge

    Chunk verdicts merge deterministically: best evidence wins, ties break by earliest position in the document.

  4. Stage 04

    Score

    Fixed weights produce the numbers. Full 1.0, partial 0.5, missing 0.0, totaled per framework.

  5. Stage 05

    Persist

    Results land with pinned policy versions, frozen scope, and audit events. History is immutable.

§ 02The core decision

Complete enumeration, not retrieval.

A gap analysis must examine every control, because its job is to find what is absent. Retrieval systems fetch what looks relevant, and absence is precisely what never looks relevant. So the catalogs are enumerated structured data, and gaps can never be silently dropped.
Fig. 3  Illustration · 358 controls
82358
Controls examined
2760
Never looked at
381
Gaps surfaced
VariesFixed
Coverage
CoveredPartialGapNever examined
Retrieval-based chat
  • Fetches the controls that look similar to the text it was given
  • A control with no matching text is simply never retrieved
  • Coverage depends on the question you happened to ask
  • Two runs can silently examine different controls
PolicyLenz enumeration
  • Iterates the full catalog: all 358 controls, every run
  • A control with no evidence is recorded as a finding, which is the point
  • Coverage is a structural guarantee, enforced by tests
  • One result per control, or the run does not complete
§ 03Scoring

Numbers the model cannot invent.

The model classifies; the engine calculates. Each control's verdict maps to a fixed weight, chunk verdicts merge by a documented rule, and the total is arithmetic. Ask twice and the same verdicts produce the same score, to the decimal.
Try it · NIST CSF 2.0 sampleIn scope
  • GV.OC-01
    Govern
  • GV.RM-01
    Govern
  • ID.AM-01
    Identify
  • PR.AA-01
    Protect
  • PR.DS-01
    Protect
  • DE.CM-01
    Detect
  • RS.MA-01
    Respond
  • RC.RP-01
    Recover
full    = 1.0
partial = 0.5
missing = 0.0
score = Σ weights / controls_in_scope × 100
score = 5.0 / 8 × 100
NIST CSF 2.0 · this sample
62.5

Merge across chunks: best verdict wins; ties break by earliest chunk. Scores are never averaged across frameworks. Out-of-scope controls leave the denominator instead of dragging the score down.

§ 04The LLM boundary

The model is a component, and it is treated like one.

PolicyLenz engineers around the model the way you engineer around any unreliable dependency: strict contracts, budgets, retries, and failure modes designed in advance.
Engine code · deterministicModelBounded · replaceable01 Structured output02 Budgets03 Retry ladder04 Partial, not poisonedIn: chunk + catalogOut: one verdict per control
  1. 01

    Forced structured output

    The model answers through a mandatory tool call with a fixed schema, not free text. There is no prose to parse and no room for the reply to wander.

  2. 02

    Computed token budgets

    Response budgets are calculated from the control count of the framework being assessed, not guessed. Costs stay predictable and bounded.

  3. 03

    An explicit retry ladder

    Transient failures retry with exponential backoff, then fall back to a second model, then fail cleanly. Nothing hangs and nothing retries forever.

  4. 04

    Graceful partial degradation

    If one framework pass fails, the run completes as partial with the other results intact. A model hiccup never poisons a whole assessment.

§ 05History

Assessments are records, not chat transcripts.

  1. #0001Policy versions are immutable once ingested; new uploads create new versions.
  2. #0002Every run pins the exact policy version and framework version it assessed.
  3. #0003Scope is resolved and frozen when the run is created, forever.
  4. #0004Framework content updates never disturb recorded results.
  5. #0005Hard deletes are refused while any run references the version: the audit trail wins.
  6. #0006Dispositions and evidence overlay results; raw verdicts are never rewritten.
§ 06Skeptic's FAQ

The questions we hope you ask.

Want this in writing?Download the security review packet for your team.

Q1Is this just GPT with a prompt?

No. The model performs one narrow job: reading a document chunk against a structured control catalog and answering through a fixed schema. Enumeration, merging, scoring, scoping, versioning, and the audit trail are deterministic engine code. Swap the model and the engine still stands; delete the engine and the model alone cannot produce a defensible assessment.

Q2What exactly leaves my network?

Self-hosted, Cloud AI edition: document chunks and catalog context travel to the model API over TLS, to exactly one allow-listed host, and nothing else leaves. Self-hosted, Local edition: nothing; the model runs on your own machine and the ledger shows only local engine calls. No telemetry, no embeddings shipped to third parties, no update phone-home in either case. On PolicyLenz Cloud the same rules as the Cloud AI edition apply to your dedicated instance, which talks only to the model API; your documents never touch a shared platform. Either way, every egress attempt, allowed or denied, is written to a ledger you can inspect in the app.

Q3Can it run fully air-gapped?

Yes, with the Self-Hosted Local edition: the model runs on the appliance host and no AI call leaves the machine. Two caveats: the engine needs internet once, at installation, for Docker to fetch its inference backend (or you run llama-server on the host yourself), and Docker Desktop needs a subscription for larger organizations while Docker Engine on Linux does not. The Cloud AI edition needs one outbound rule to one host.

Q4Can I reproduce a score?

The scoring is fully deterministic: given the same per-control verdicts, the same numbers come out every time, and the merge rules are documented. The model layer is where variance can enter, which is exactly why verdicts are recorded per control and per chunk, so any surprising verdict can be traced to the specific passage that produced it.

Q5What happens when the model call fails?

Retries with exponential backoff, then a fallback model, then a clean failure that marks the run partial while every other framework keeps its results. Failure states are first-class citizens, not exceptions to hide.

Q6Who reviewed the control catalogs?

NIST CSF 2.0 ships as verbatim official text. ISO/IEC 27001, SOC 2, GDPR, and DORA ship complete-coverage control objectives paraphrased by our team, disclosed as such, with a verbatim overlay path for deployers who hold the relevant licenses. Coverage is guaranteed and test-enforced; the wording is honest about what it is.

Next step

Bring the hardest question you have.

The demo is a working session with the engine, not a slideshow about it.