Access → the Cast: identity, trust, audit (roadmap)¶

How asmltr's access layer grows from a capability resolver into a full cast — who everyone is, how the assistant relates to them, how trust evolves from behavior, and an auditable record of every consequential action. This is an evolution of the existing trust store, not a second system: everything hangs off principals.id.

What was already built (the base)¶

The trust store (core/src/trust/store.js, data/trust.db):

principals — one identity per person/entity (default_tier, revoked).
identifiers (surface, value) → principal — Discord id, Telegram @, email, GitHub login, api key… One principal can already hold many identifiers across channels → cross-channel identity primitive.
roles / grants — capability bundles, scoped by (surface, scope_id).
resolve(envelope) — sender → principal → unioned scoped grants → effective capabilities (default-deny).
buildAuthzPrompt() — capabilities are advisory (fed to the model); moderation is hard-gated.

The four phases¶

Phase 0 — the cast goes live locally ✅ (built)¶

Recognition + how-to-relate + cross-channel identity, injected into the prompt. Pure additive — disabling the relationship block changes nothing else.

principal_profile (1:1 on principals.id): kind (human | agent | self), who_they_are, how_to_relate, expertise, decision_authority, provenance.
relationships — pairwise, directional edges (the cast is about edges, not entities), optional scope.
engagement (principal, scope) → engage | observe | ignore — a per-scope override that retires per-connector allowed_bot_names; the connector/core asks the cast instead. ignore mutes a member in one channel; observe keeps awareness without replying.
verification_strength on identifiers (0 claimed · 1 channel-owned [default] · 2 vouched · 3 cryptographic) — the seam for later phases; a no-op today.
self principal — the assistant is itself a cast member (the anchor for self→other relationships).
buildRelationshipPrompt(resolved, envelope) at the prompt seam: who you're talking to + their cross-channel identity ("the same person you also know as @x, @y, z@…") + your relationship + peer agents present ("you share this channel with other agents — a message to one of them isn't for you").
/trust/{profiles,relationships,engagement} endpoints for management; seeded self + example peers.

Effect: peer agents are recognized by name on any channel with no per-channel config, and a person is understood as one identity across all their channels — from day one.

Phase 1 — one tamper-evident action ledger (audit) — next¶

The substrate for evolving trust (you can't evolve trust from behavior without a trustworthy record).

action_ledger — append-only, HMAC hash-chained ({seq, prev_hash, ts, actor_principal, subject_principal, action, scope, detail, result, hash}); built chain-ready so Phase 3 swaps HMAC→Ed25519.
Attribute every consequential action to a principal (not the OS/Authelia user): trust mutations (principal/grant/role/identifier CRUD), moderation decisions, control actions, cross-channel sends, draft create→approve provenance, and the manager /send + config/lifecycle changes (a current blind spot).
An Audit view in the dashboard.

Phase 2 — trust that evolves from behavior (local, advisory)¶

Tier becomes the ladder (NOVICE→COMPANION→PARTNER→GUARDIAN→SACRED): promotion slow + human-ratified, demotion instant + automatic.
A reputation view computed from the Phase-1 ledger + moderation signals: drift/anomaly → auto-demote to NOVICE + review flag; positive history → propose a promotion for a human to confirm.
verification_strength starts mattering (a grant can require a minimum strength).
All local + advisory — the local store stays the policy authority.

Phase 3 — cross-channel linking + federation / TRUST-protocol¶

Assisted linking (heuristic candidates → human confirm → raises verification_strength).
Ed25519-sign the ledger (upgrade Phase 1's chain); per-instance signed logs.
federation identifiers + resolve() delegates strength-3 verification to the TRUST-protocol ledger as an advisory oracle while the local store stays authority; gossip / cross-signed checkpoints / revocation propagation.

Non-negotiables (every phase)¶

No second trust store — principals.id is the universal join.
Behavioral/on-chain reputation is advisory — it can lower trust and raise the verification floor; it may never auto-raise past what local behavior + a local human granted.
A human closes the circle on every consequential crossing (high-tier promotion, first credential-backed action, admitting a peer, honoring a non-anchor revocation).
The membrane is first-class — warmth widens the aperture, never dissolves "your words are yours."

Still to do off Phase 0¶

A Cast tab in the dashboard (profiles + relationships + engagement editing) — extends Access.vue.
Scope the "peers present" list to actual per-channel presence (today it's surface-wide).
An importer to fold existing narrative registries (RELATIONSHIPS notes) into principal_profile.