[EPIC] Privacy-preserving ambient perception — ephemeral sensory input to symbolic world facts #339

New issue

Open

opened 2026-09-08 02:03:35 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-08 02:03:35 +00:00

Owner

Parent architecture: #624
Integrates with: #152/#194 Android device edge, #622 offline symbolics, existing wake/audio privacy boundaries, future typed sensor capabilities.

Goal

Let Zara perceive useful environmental events without turning always-on sensing into a permanent raw-audio/video surveillance archive.

Target pipeline:

short local circular sensor buffer
  -> bounded local detector/classifier
  -> typed symbolic observation
  -> policy/world-state reasoning
  -> raw buffer expires by default

Example observations:

observed(doorbell, at(T)).
observed(glass_break, confidence(0.91), at(T)).
observed(water_running, kitchen, duration(480)).
observed(person_present, kitchen, false, at(T)).

Privacy architecture

raw ambient audio/video is ephemeral by default with a short bounded circular buffer;
ordinary symbolic observation does not retain raw media;
retention of source media requires a separate explicit capability/policy and visible state;
local processing is preferred for detection paths capable of running locally;
no ambient transcript is silently appended to conversation or memory;
bystander/private-space policy and hardware permission state are first-class;
diagnostics record detector/version/timing/status, not raw content.

Observation contract

Each observation includes principal/device ownership where applicable, sensor source class, time interval, detector identity/version, confidence/evidence class, expiration policy and optional linkage to a retained artifact only when explicitly authorized.

Observed facts are evidence, not automatically trusted canonical facts. They feed the epistemic-memory/world-state layer under its provenance rules.

Reasoning examples

possible_problem :-
    observed(water_running, Room, duration(D)),
    D > 600,
    observed(person_present, Room, false, _).

The rule engine may propose/notify according to policy but never gain additional capability authority from a detector assertion.

Scope / sensors

Start narrow with useful, testable local signals such as doorbell/alarm/smoke/glass/water-like events and device/system state. Camera/person/object sensing requires explicit platform/privacy research and should not be implied by the audio slice.

Tests

Prove raw-buffer expiry, no transcript/history leakage, detector false/malformed output handling, bounded confidence, sensor permission denial, offline operation, two-device correlation, stale observations, retention opt-in, and no side effect solely from untrusted detector output without policy authorization.

Acceptance

Zara can derive useful typed ambient observations locally, reason over them and discard source media by default, with explicit provenance and privacy controls strong enough that ambient perception is not synonymous with ambient recording.

Parent architecture: #624 Integrates with: #152/#194 Android device edge, #622 offline symbolics, existing wake/audio privacy boundaries, future typed sensor capabilities. ## Goal Let Zara perceive useful environmental events without turning always-on sensing into a permanent raw-audio/video surveillance archive. Target pipeline: ```text short local circular sensor buffer -> bounded local detector/classifier -> typed symbolic observation -> policy/world-state reasoning -> raw buffer expires by default ``` Example observations: ```prolog observed(doorbell, at(T)). observed(glass_break, confidence(0.91), at(T)). observed(water_running, kitchen, duration(480)). observed(person_present, kitchen, false, at(T)). ``` ## Privacy architecture - raw ambient audio/video is ephemeral by default with a short bounded circular buffer; - ordinary symbolic observation does not retain raw media; - retention of source media requires a separate explicit capability/policy and visible state; - local processing is preferred for detection paths capable of running locally; - no ambient transcript is silently appended to conversation or memory; - bystander/private-space policy and hardware permission state are first-class; - diagnostics record detector/version/timing/status, not raw content. ## Observation contract Each observation includes principal/device ownership where applicable, sensor source class, time interval, detector identity/version, confidence/evidence class, expiration policy and optional linkage to a retained artifact only when explicitly authorized. Observed facts are **evidence**, not automatically trusted canonical facts. They feed the epistemic-memory/world-state layer under its provenance rules. ## Reasoning examples ```prolog possible_problem :- observed(water_running, Room, duration(D)), D > 600, observed(person_present, Room, false, _). ``` The rule engine may propose/notify according to policy but never gain additional capability authority from a detector assertion. ## Scope / sensors Start narrow with useful, testable local signals such as doorbell/alarm/smoke/glass/water-like events and device/system state. Camera/person/object sensing requires explicit platform/privacy research and should not be implied by the audio slice. ## Tests Prove raw-buffer expiry, no transcript/history leakage, detector false/malformed output handling, bounded confidence, sensor permission denial, offline operation, two-device correlation, stale observations, retention opt-in, and no side effect solely from untrusted detector output without policy authorization. ## Acceptance Zara can derive useful typed ambient observations locally, reason over them and discard source media by default, with explicit provenance and privacy controls strong enough that ambient perception is not synonymous with ambient recording.