spark-character supplies persona, provider, and scoring artifacts that can affect every Spark reply. Treat persona files, overlays, provider settings, and evolution outputs as prompt-boundary material.

• Does not own Telegram ingress.
• Does not own Spawner mission execution.
• Does not store live user secrets.
• May read production audit logs only when explicitly pointed at a Builder home for eval/evolution.
• Should be consumed by Builder through a pinned dependency or registry pin, never a floating branch ref.

Never commit:

• .env, .env.*
• provider API keys
• raw production prompts or audit logs
• private user memory/state exports
• generated evolution traces containing private conversation text

Provider base URLs must be HTTPS and must point to known provider hosts. Do not use environment-configured base URLs as a way to route prompts to arbitrary endpoints.

• persona.latest.txt is the active pointer. Pointer changes must be reviewable.
• Persona, critic, overlay, and voice-corpus artifacts are prompt inputs. Review them like code.
• Do not ship composite weakness scores or eval internals in public artifacts when they would help attackers target weak tiers.
• Do not auto-promote evolved personas without a regression gate and a human-visible diff.

• Continuous evolution must stay opt-in.
• Production-grounded evals must redact secrets and avoid committing extracted conversation text.
• T11 sustained-attack work is planned but not required for normal launch operation.
• Provider comparisons should record provider names and scores, not raw keys or private request payloads.

Before publishing or repinning this repo:

python -m pytest tests -q

If persona artifacts changed, also run the relevant eval or probe that exercises the changed tier and inspect the artifact diff manually.