Skip to content

Instantly share code, notes, and snippets.

@Hulupeep
Created June 18, 2026 02:13
Show Gist options
  • Select an option

  • Save Hulupeep/7202f881bc087654a3bc23a6de870dd5 to your computer and use it in GitHub Desktop.

Select an option

Save Hulupeep/7202f881bc087654a3bc23a6de870dd5 to your computer and use it in GitHub Desktop.
Runtime trust certificate floor: a per-step dynamic finite-precision threshold for autonomous-agent trust verification — separating real loss-of-contraction signals from machine rounding noise. CCF Prov6, Rust benchmarks (f64/f32/fixed-point), no universal threshold.

CCF 2026: The Runtime Trust Certificate Floor — Telling a Real Signal From Rounding Noise in Autonomous-Agent Trust Verification

A runtime trust certificate floor for autonomous agents: a per-step, dynamic finite-precision threshold that separates a real loss-of-contraction signal from machine rounding noise. Real Rust benchmarks across f64/f32/fixed-point.

Introduction

Every runtime trust certificate for an autonomous agent eventually runs into arithmetic. As a trust value homes in on its target, the per-step change shrinks — until it's smaller than the rounding error of the computer making the change. At that point the residual you measure is noise, not signal. Hard-code a single tolerance like κ_t < 1e-9 and you get the worst of both worlds: a 32-bit controller never believes it converged (its rounding floor is bigger than 1e-9), while a 64-bit one declares victory with a thousandfold of real movement still left. This is the runtime trust certificate floor problem, and it sits at the center of CCF's Prov 6 architecture for verifiable autonomous-agent trust transfer.

What CCF does about it

CCF's Prov 6 (US 64/092,485, filed 2026-06-17) accepts a realized trust step only when its per-step quotient-affine-contraction (QAC) residual κ̂_t clears a dynamic finite-precision floor, recomputed every tick:

floor_t = 2·n·ε_q + |Δα_t|·E_t
  • 2·n·ε_q — the rounding accumulated across the n-term gauge/normalization step.
  • |Δα_t|·E_t — slack for legitimate movement when the gate changes the step size.

The certificate has three states, and the third is the point: residual above the floor and shrinking → CERTIFIED_CONTRACTING; at or below the floor → NOT_CERTIFIED ("I can't tell" — honest, not a false pass); above the floor and rising → EXCURSION → fail closed. This PoC implements floor_t in isolation and measures it. Underclaim, on purpose: it does not prove trust convergence, and Sinkhorn/gauge normalization is treated as presentation, not the causal trust dynamic. (See Prov 6 dynamic-floor disclosure; cf. the gauge posture in [0099–0101].)

Implementation

crates/research/runtime-trust-certificate-floor/ — a zero-dependency Rust crate. One QAC iteration (affine contraction in a log-style chart + mean-subtraction gauge) runs over four precision backends behind a Scalar trait: f64, f32, fixed-point Q1.30, fixed-point Q1.14. The same algorithm, four noise floors. 10 tests, no mocks; floor-demo + floor-bench binaries print every number below.

Comparisons

Approach What it certifies Runtime? Numeric residual?
Certified roundoff bounds (arXiv 1507.03331, 1707.02115) program roundoff error offline yes (static)
Equilibrated-flux residual cert. (arXiv 2602.01636) verified PDE outputs offline yes (static)
Microsoft Agent Governance Toolkit policy compliance runtime no
EQTY Lab Verifiable Runtime attested execution (TEE) runtime no
CCF Prov 6 floor (this PoC) trust-step is signal vs. noise runtime, per-step yes, dynamic

Static roundoff certification is one-shot; 2026's agent-governance wave does policy and hardware attestation. None compute a runtime per-step floor for a trust residual.

Benchmarks

Intel Core i5-10400T @ 2.00 GHz · rustc 1.90.0 · --release · n = 16.

Detection floor is precision-ordered across ~11 orders of magnitude — the smallest detectable fault tracks each backend's 2·n·ε_q:

backend floor = 2·n·ε_q min detectable fault ns / certified step
f64 3.55e-15 1e-14 157.5
fixed Q1.30 1.49e-8 1e-7 195.1
f32 1.91e-6 1e-5 147.9
fixed Q1.14 9.77e-4 1e-2 189.8
  • No universal threshold: a 1e-7 fault is a detected safety event on f64 but below the floor (invisible) on f32 and Q1.14.
  • Excursion detection: an injected loss of contraction stands 9.5e1× (Q1.14) to 2.6e13× (f64) above the floor; flagged at 0-step latency.
  • Δα term: a mid-transient gate acceleration causes 1 false excursion without the |Δα_t|·E_t term and 0 with it, on every backend.
  • Throughput 5.1–6.8 M certified steps/s; state vector 64–128 B.

Failure modes

  • Not a convergence proof. A clean contraction reaches an exact fixed point (residual → 0); the floor certifies this step's signal-vs-noise, not that the iteration must converge.
  • O(1) normalization is assumed. The magnitude-free 2·n·ε_q term only bounds a float rounding floor when the state is O(1) — the gauge step is what keeps that true (load-bearing, not decorative).
  • Gate transitions lower sensitivity. The |Δα_t|·E_t term is conservative; a real excursion coinciding with a big gate change can fall back to NOT_CERTIFIED (fail-closed-safe).
  • Abbreviated certificate: this computes the floor + classifier, not endpoint witnesses or the envelope monitor B_t.

Get started

  • Repo: github.com/Hulupeep/ccf
  • Research branch: research/nightly/2026-06-18-runtime-trust-certificate-floor
  • Crate: crates/research/runtime-trust-certificate-floor/ · Issue: Hulupeep/CCF#165
  • Specflow (causation-trail method): github.com/Hulupeep/Specflow
  • More on CCF and the shy-robot trust model: floutlabs.com · theshyrobot.com
cargo run --release -p runtime-trust-certificate-floor --bin floor-demo
cargo run --release -p runtime-trust-certificate-floor --bin floor-bench

License & contact

BSL 1.1 → Apache 2.0 in 2032. Contact via floutlabs.com.

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