Skip to content

gusitllc/dynamic-confidence-bands

Repository files navigation

Dynamic Confidence Bands per Action Type

Per-action-type confidence bands with rolling recalibration and an auto-execute / escalate / human-review routing trichotomy for AI agents.

type: defensive-publication status: published license: AGPL--3.0 date: 2026--06--25


📌 Defensive Publication / Prior Art Notice

This repository is a Technical Defensive Publication. It is published publicly and intentionally on 2026-06-25 by Gus IT LLC (Florida, USA) to establish dated, citable prior art for the technique described herein.

Its purpose is defensive: to keep this technique freely practiceable by the public and to prevent its later monopolization by others through patenting. See Appendix D of the whitepaper for the deposit and timestamp note.

  • Publisher / Copyright holder: Gus IT LLC (Florida, USA)
  • Author: Gustavo Assuncao, PhD
  • Publication date: 2026-06-25
  • Version: 1.0
  • Document type: Technical Defensive Publication (public prior art)
  • Classification: Public
  • License: AGPL-3.0-or-later (copyleft; commercial license available)

Abstract

AI agents that act on a user's behalf — sending email, posting chat replies, creating tasks, calling tools, delegating to sub-agents — need a way to decide, per action, whether to act autonomously, escalate for a quick check, or hold for full human review. The prevailing approach is a single static confidence threshold ("auto-send if confidence > 0.8"), hand-tuned by operators. A single threshold is brittle: it cannot distinguish a low-stakes chat reply from a high-stakes external email, it does not adapt as the agent demonstrates competence (or starts failing), and it forces every action type through the same gate.

This publication discloses a method in which each action type maintains its own confidence band — a (low, high) pair — rather than a single threshold. On completion of each action the system records an outcome and recomputes the band, without human intervention, as a function of a sliding window of recent outcomes for that action type: the band tightens (raises low/high) when failures rise and loosens (lowers them) as successes accumulate. A subsequent action is then routed by a trichotomy: confidence above high → auto-execute; below low → require human review; in between → escalate (a lightweight, time-boxed check). Because every action type calibrates independently, an agent can become fully autonomous on chat replies while still requiring review on external email — a graduated autonomy gradient that emerges from observed outcomes rather than from manual tuning. This document provides architecture, mechanics, data model, an enabling clean-room reference implementation, a worked example, security/failure-mode analysis, standards alignment, an evaluation methodology, and enumerated novelty claims.

Why this is published

We publish this as prior art, deliberately and in the open, so that the technique of per-action-type confidence bands with rolling, human-free recalibration and an auto/escalate/review routing trichotomy remains free for everyone to use. A defensive publication creates a dated, public, citable record that a patent examiner (or a court) can use to reject later patent claims over the same idea. We would rather the field build on this openly than see it fenced off. The accompanying AGPL-3.0-or-later license adds an express patent grant over our own contribution.

Table of contents

Document What's inside
DEFENSIVE-PUBLICATION.md The full whitepaper: architecture, mechanics, data model, enablement, worked example, security, standards mapping, evaluation, and the novelty claims (1 independent + 14 dependent).
docs/PRIOR-ART.md Fuller prior-art landscape, the delta table, and an honesty attestation.
docs/FIGURES.md All Mermaid figures collected with captions (Figure 1–6).
docs/OPEN-SOURCE-APP.md The planned open-source reference app and a generic Kubernetes/AKS deployment sketch.
src/ Original, clean-room illustrative reference implementation + a runnable self-check.

Planned open-source app

This repository is intended to seed a small open-source reference app — a self-contained "Confidence Band Router" service that any agent runtime can put in front of its action executor. The plan, minimal architecture, and a generic Kubernetes/AKS deployment sketch are in docs/OPEN-SOURCE-APP.md.

Reference implementation

src/ contains an original, dependency-free Node.js implementation of the band-router and recalibrator, with a tiny self-check you can run:

cd src
node example.js

See src/README.md for what it demonstrates and the caveat that it is illustrative/clean-room, not production code.

License & citation

Licensed under AGPL-3.0-or-later — see LICENSE and NOTICE.

To cite, use CITATION.cff or:

Assuncao, G. (2026). Dynamic Confidence Bands per Action Type: A Technical Defensive Publication (Version 1.0). Gus IT LLC. Published 2026-06-25.


Defensive publication — Copyright 2026 Gus IT LLC (Florida, USA). Published 2026-06-25. This disclosure is intentionally public to bar later patenting by others.

About

Defensive publication: per-action-type confidence bands with rolling recalibration and an auto/escalate/review routing trichotomy for AI agents.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors