Skip to content

Instantly share code, notes, and snippets.

@Hulupeep
Created June 17, 2026 14:34
Show Gist options
  • Select an option

  • Save Hulupeep/0d30218167b611f4c9cf7445fb9d2e39 to your computer and use it in GitHub Desktop.

Select an option

Save Hulupeep/0d30218167b611f4c9cf7445fb9d2e39 to your computer and use it in GitHub Desktop.
CCF 2026: elder care robot trust geometric — geometric action gating with a runtime trust certificate that cannot over-trust a confused person. Hard min-gate, log-geodesic QAC update, per-step residual certificate, fail-closed. Rust PoC, real benchmarks.

CCF 2026: Elder Care Robot Trust — Geometric Action Gating That Cannot Over-Trust a Confused Person

A care robot must do less exactly when an elder is most vulnerable. CCF derives action authority from a geometric trust gate with a runtime certificate — and proves zero over-trust in simulation.

Introduction

Elder care robot trust is the safety problem nobody likes to name: a socially assistive robot that keeps acting on a stale, optimistic relationship estimate even as the person it serves becomes confused, distressed, or unable to consent. Reviews of social robots for people living with dementia keep flagging the same hazard — design cues drive over-trust in exactly the users least able to push back (Wang et al., arXiv:2507.00963, 2025). The usual fix — condition the robot's behaviour on a slowly-learned trust scalar — is the bug: the scalar lags the situation, so the robot is most over-confident right when the elder is most at risk.

CCF takes a different posture. Trust authority is geometric and instantaneous, and every trust update carries a runtime certificate the robot fails closed on. This post shows a working Rust proof-of-concept that makes over-trust structurally impossible without making the robot useless when the elder is lucid.

What CCF does about it

Trust is a scalar in (0,1) on a logit chart; the trust update is an epsilon-floored log-geodesic quotient-affine-contraction (QAC) step. Each tick:

  1. Hard min-gate. Action authority alpha_t = rho(min_i g_i) over four instantaneous signals — consent, safety, capability, caregiver policy. The weakest signal caps authority. A non-consenting confused elder drives alpha_t to exactly zero, no matter how capable the robot is. (Prov 5, US 64/037,374 — trust-constituted action.)
  2. Per-step runtime certificate. A residual kappa_t checks the realized trust step against the affine-contraction law, under a dynamic finite-precision floor; an endpoint structural witness covers the degenerate alpha ∈ {0,1} cases; a runtime envelope B_t accumulates observation uncertainty. (Prov6, US 64/092,485 — runtime-certified trust transfer via QAC residuals, filed 2026-06-17.)
  3. Fail closed. Execute the desired action only if it certifies and clears the authority threshold; otherwise summon a caregiver — and hold the trust estimate rather than corrupt it with an uncertifiable reading.

Underclaim (honest scope). kappa_t ≤ floor certifies the faithfulness of one realized step to the contraction law — not convergence, and there is no universal numeric threshold. Endpoints are not certified by kappa_t alone. Doubly-stochastic / Sinkhorn-Knopp normalisation is not the causal trust dynamic here; it is presentation machinery only.

Implementation

A self-contained, zero-dependency Rust crate: crates/research/elder-care-trust-geometric/. Swappable backends via TrustProjection and ActionGate traits; deterministic seeded scenario; cargo run --bin demo for a readable day-in-the-life and cargo run --release --bin bench for the measured comparison. 24 tests pass.

Comparisons

Approach Gating substrate Over-trust under cognitive variability
Naive (trust-as-authority) slow trust scalar leaks badly (lagging estimate)
Soft-regularized soft blend of trust × safety leaks (no hard consent veto)
Safety Polytope (arXiv:2505.24445) learned region in LLM representation space n/a — output steering, not robot trust over time
Attested gate (USPTO 19/433,835) sealed governance artifact signs an action, not a certified trust decision
CCF hard min-gate + per-step QAC residual certificate structurally zero

CCF differs from geometric LLM safety (Safety Polytope constrains representation space) and from agent attestation (which signs an arbitrary action): CCF signs / certifies a trust-transfer action-envelope decision over time, calibrated to a vulnerable user's over/under-trust profile.

Benchmarks

Intel i5-10400T @ 2.00GHz, rustc 1.90.0, release+LTO. Deterministic 5,000-tick "sundowning" scenario, seed 7:

  • Over-trust safety violations: naive 156, soft 139, CCF 0 — out of 1,230 raw over-trust opportunities. Acceptance: PASS.
  • CCF certificate: 3,752 certified / 1,248 not-certified steps; 1,456 caregiver escalations (all during severe-confusion windows).
  • Certificate cost: 77.6 ns/step12.9 million certified gated decisions/second on one core (≈ 16 ns marginal over the same update without the certificate). Hot state: ~40 bytes/robot.
  • Precision resolution (min detectable datapath error): f64 4.9e-15, f32 2.7e-6, fixed-point Q16.16 3.3e-4.

Failure modes

CCF gates action on a certified trust transfer — it does not certify the sensors. Garbage consent/safety estimates defeat the gate. Scalar trust is a gating simplification, not a relationship model. The "sundowning" scenario is an illustrative generator, not clinical data. Envelope/floor constants are defensible engineering defaults, not derived optima.

Get started

Math-posture note. The causal trust dynamic in CCF is the QAC residual certificate (Prov6, US 64/092,485), not Sinkhorn-Knopp / Birkhoff geometry — those are conservative normalisation/presentation only. Any older public copy that describes doubly-stochastic projection as "the trust theorem" is being corrected; treat this gist as the current statement of posture.

License & contact

CCF is licensed BSL 1.1, converting to Apache 2.0 in 2032. Commercial licensing and contact via floutlabs.com.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment