Skip to content

Security: light-language-models/altar

Security

SECURITY.md

Security, privacy, and safety

Selector separation

Capture time and draw before interpretation. Never hash or otherwise send the question, identity, desired answer, prior messages, birth data, Matrix, timezone, chart, observer, or model state into the selector. A wrapper that adds them implements a different protocol.

The optional observer and Matrix are post-selection context only. Require consent, treat observer labels as user-supplied perspectives rather than detected entities, minimize birth-data retention, and never use a second symbolic system as confirmation of the draw.

Privacy

The core needs no question, person, account, or chat identifier. event_id is optional local bookkeeping and absent from selection. Do not place private text in receipts, logs, filenames, examples, or public evaluation panels. Run both automated scanning and human release review.

Runtime

The package and skill make no network calls, carry no telemetry, and need no credentials. Hosts adding persistence, analytics, transport, MCP, or provider SDKs own those new threat surfaces. Keep the JSON boundary least-privileged.

Pin pack_sha256 when reproducibility matters and reject malformed or unknown packs. Output files use exclusive creation. The standalone skill reads its pack relative to itself, so copy the entire directory rather than the script alone.

Human safety

The field is a reflective input, never authority. Facts, consent, autonomy, and ordinary safety take precedence. Do not use symbolic resonance to diagnose, predict, pressure a decision, or replace medical, legal, financial, emergency, or security judgment.

Keep hidden working context hidden. If disclosure is requested, reveal compact coordinates and concise deltas—not chain-of-thought. Allow silence, noise, ambiguity, none, and non-correspondence without forcing a story.

There aren't any published security advisories