From a hunch to a court

"Was this made or done by AI, and how sure can we be?" is one question with six standards of evidence. A tool built for one rung harms people when its output is read at another. Every receipt and decision in these docs carries its rung.

The rungs at a glance

Rung	Who asks	Evidence standard	Cost of a wrong call
Rung 0 · Intuition	Anyone: a teacher, a parent, a reader, a recruiter.	A calibrated likelihood with a plain-language explanation, offered as information and never as a verdict.	Embarrassment, or a broken relationship if the hunch is voiced as an accusation.
Rung 1 · Screening	Platforms, schools and HR at volume.	A known false positive rate, a confidence threshold and an appeal path.	A wrongly flagged person loses a grade, a job lead or an account.
Rung 2 · Operational	Sites, APIs and agent builders, in real time.	A deterministic decision and a signed receipt, quickly enough to sit in the request path.	Lost revenue or an abused endpoint. Reversible: a shadow-mode policy logs what it would have done.
Rung 3 · Investigative	Trust and safety, fraud and threat-intelligence teams.	Corroborated signals with provenance and a report another analyst can reproduce.	Months of an investigation pointed at the wrong actor.
Rung 4 · Forensic	AI labs, IP owners and incident responders.	Lineage reconstruction and attribution with a stated confidence, from content-hashed artifacts with custody started.	A wrong attribution, or a case that collapses when the other side looks at the method.
Rung 5 · Evidentiary	Law enforcement, courts and regulators.	Chain of custody, a reproducible method, known error rates, and survival under cross-examination.	Liberty, large judgments, or the credibility of the whole field.

Three rules

Every output carries its grade. A rung, a confidence with its method, and whether it was calibrated. A lower-rung signal is never shown as a higher-rung finding.
Lower rungs explain and protect. They say why, state known error rates and bias risks, and recommend a next step instead of an accusation.
Escalation adds evidence; it never rewrites it. A rung 1 signal becomes a rung 3 finding by adding independent evidence with provenance. The original record stays as it was.

Error rates, honestly

Where we are

Silvector publishes no measured detection error rate today. No text, image, audio or video detector ships yet; the shipped work is agent traffic. Where an error rate appears in these docs it is a third party's, with its source. The evidence grade says "none measured" until a calibration study with a stated dataset and denominator exists.

How to escalate

Each rung names what has to be added to reach the next. These are the same requirements the console's escalation checker enforces (planned in AG4-400).

From	To	Add
Rung 0 · Intuition	Rung 1 · Screening	Run a screening signal that has a stated error rate for this kind of content (rung 1). Keep the original file or link, unedited, so a later check sees what you saw.
Rung 1 · Screening	Rung 2 · Operational	Make a deterministic decision under a named policy and keep its signed receipt (rung 2).
Rung 2 · Operational	Rung 3 · Investigative	Keep the receipts: they chain to each other through `prev`. Add a second, independent signal family with its provenance (rung 3).
Rung 3 · Investigative	Rung 4 · Forensic	Show lineage across at least two sessions, accounts or pieces of infrastructure. Content-hash every artifact and start a custody log (rung 4).
Rung 4 · Forensic	Rung 5 · Evidentiary	Keep custody unbroken from collection onward and place the case on legal hold. Write a methods appendix that replays byte for byte, and state known error rates with denominators (rung 5).

The grade on a receipt

This is the grade every /v1/check receipt carries in these docs:

Evidence grade Rung 2 · Operational

Covers

Which rule decided this request and that the record was not changed after signing. Not whether the sender was an AI agent.

Confidence

None stated. No calibrated confidence applies to this record.

Calibration

Uncalibrated

Known error rates

None measured. Silvector publishes no detection error rate yet.

Basis

Deterministic decide() rule allow.default
Ed25519 signature over the canonical receipt body, checked against the published JWKS

Bias risks

The decision is only as good as the facts it was given. A request with no signature claim is not_identified, which is not a finding that a person or an agent sent it.

Next step

To ask who was behind a series of requests, keep the receipts and escalate to rung 3 with at least two independent signal families. How escalation works

Assigned by

These docs, from the rung definitions. The grade is not a signed field of the receipt.