Trust Modulation

Trust isn’t frozen at boot. Bob advises Alice to practice empty_step; her graph records the provenance (empty_step, advised_by, bob). Each time Alice subsequently performs a pose with that stance successfully, bump_trust(alice, bob, +δ) ticks the (alice, taught_by, bob) valence upward. Advice that consistently works earns trust; orthogonal advice doesn’t. Same value + δ shape as lab 12, applied to a single relationship edge.

A
alice
trust(alice → bob)
0.000
B
bob

Settings

bump δ 0.05
clamp trust ∈ [−1,+1]:
Practices: 0
Validated: 0
Trust: 0.000

Alice’s belief graph

No advised_by edges yet.
taught_by valence — mint bar = positive
(alice, taught_by, bob)
0.000
awaiting first action
How it works. When Bob instructs Alice, the reflex commits (stance, advised_by, bob) — provenance that survives in Alice’s graph. When Alice later performs a pose with that stance, a different reflex (confirm_instruction) consults the advised_by edge, confirms the practice was Bob’s idea, and calls bump_trust(alice, bob, +δ): the old (alice, taught_by, bob) edge is removed and a new one written at clamp(old_valence + δ, −1, 1). The substrate that gated the teaching loop now also receives the feedback that shapes future gating.

Tie to Eddy. This is examples/cognition_trust_modulation.pyreflex_trust_confirm_instruction watching performed signals against advised_by provenance in the agent’s own belief graph. Trust is a valenced directional edge (student, taught_by, teacher) valence ∈ [−1, +1] — positive = trust, negative = mistrust, absent = neutral. The same bump → clamp → re-commit pattern works for any role edge eddy tracks.