Skip to content

evidence-model.md

Latest commit

 

History

History
86 lines (68 loc) · 4.34 KB

evidence-model.md

File metadata and controls

86 lines (68 loc) · 4.34 KB
Raw

Evidence model

First rule

AntiFlock states only what its evidence supports. Evidence class, confidence, freshness, and location precision are independent dimensions. A high-confidence community report is still REPORTED; a directly measured but noisy condition is still DETECTED with lower confidence.

Evidence classes

Class Meaning Permitted assertion pattern
DETECTED The local device or an authorized gateway directly observed the condition. "The default gateway changed at 10:42."
VERIFIED Required corroboration or a trusted reviewer confirmed the claim under a documented method. "This record was verified from two independent sources."
REPORTED A public source or community contributor reported the claim; AntiFlock has not independently established it. "An ALPR installation is reported in this area."
INFERRED A deterministic rule derived the claim from named observable facts. "Traffic is inferred to bypass the approved route."
SUSPECTED Evidence is consistent with a possible active security event but is inconclusive. "Interception is suspected but not confirmed."
UNKNOWN Visibility or evidence is insufficient to classify the claim. "AntiFlock cannot verify the complete path."

VERIFIED is not permanent truth. It records that a verification policy was satisfied at a time. Expired evidence cannot remain presented as current.

Required claim fields

Every alert, finding, field marker, footprint relationship, or intelligence claim MUST carry or resolve to:

  • a stable claim/finding ID and deterministic reason code;
  • evidence class and a confidence value from 0 through 1;
  • observed or reported time and received time;
  • last-verification time when verification occurred;
  • expiry or an explicit retention/refresh rule;
  • source type and source reference, with license where applicable;
  • evidence type and integrity reference where available;
  • location precision and sensitivity where location is relevant;
  • an evidence-grounded explanation;
  • recommended response;
  • a false-positive or alternative-explanation note; and
  • all supporting evidence references, including contradictory or disputed evidence material to the decision.

Confidence MUST be calibrated within a rule or source family, never treated as a universal probability. User interfaces MUST show words before decimals and MUST NOT use confidence to erase the evidence class.

Alert contract

Before an alert template can ship, it defines:

  1. Reason code: stable, searchable, and versioned.
  2. Evidence requirements: facts, freshness, capabilities, and exclusions.
  3. Confidence rule: deterministic computation and threshold.
  4. User wording: factual condition, consequence, and uncertainty.
  5. Recommended response: reversible and scoped where possible.
  6. False-positive explanation: plausible benign or incomplete causes.

The phrase "someone is watching you" is prohibited unless evidence supports a specific actual observation or interception event and the wording has passed a security and legal review. A public network, failed tunnel, nearby camera, or reported ALPR is not that evidence.

Approved examples:

  • Protection interrupted. Your approved secure route is unavailable on an untrusted network. Protected traffic has been paused.
  • Potential interception indicators. AntiFlock detected behavior consistent with gateway impersonation or traffic interception. Active interception is suspected but not confirmed.
  • Reported monitoring infrastructure nearby. Two recently verified reports are located within the current area. This does not indicate interception of your device or traffic.
  • Visibility unknown. AntiFlock cannot verify the complete path between this device and the destination.

Corrections and aggregation

Evidence is append-only at the event boundary. Corrections reference the superseded claim or event; they do not rewrite history. Aggregators preserve source-level classifications and provenance. Multiple reports may satisfy a documented verification policy, but duplication, copied sources, shared ownership, and correlated sensors do not count as independent corroboration.

When evidence conflicts, AntiFlock lowers confidence, exposes the dispute, and uses UNKNOWN or SUSPECTED as appropriate. It does not select the most dramatic source.