Skip to content

UpCube Research Lab

Publications

Browse by type and status. Every publication carries a classification and an evidence status, so a proposal is never read as a result.

These are publications of the UpCube Research Lab, UpCube’s research organisation. They are separate from the Ethen Research workspace, a product in development, and from the Blog, which carries company and product updates.

The archive

Published items, by classification

41

Generated from the publication catalogue. Select a classification to filter the archive.

01Featured

Browse all publications

System Card

AgentTrustBench system card: autonomy boundaries on one pinned build

System card for one offline enumeration of a pinned Ethen build: 197 conditions, 194 passed. Not a paper and not a security disclosure.

Classification
Classification: Measured resultReports a measurement UpCube ran, only within the scope the publication states.
Evidence status
System Evidence
Offline enumeration · one pinned build

Also featured

02Archive

All publications

41 publications

The complete published archive.

  • Research Note12 min read

    From AI Traces to Verified Experience

    Logs, traces, trajectories, outcomes and corrections are not the same asset. A research note on what turns agent telemetry into verified experience.

    Classification: Synthesis

    Evidence status: Research Synthesis

  • Technical Report13 min read

    Work Receipts: A Verifiable Record for Autonomous AI Work

    A technical report proposing the Work Receipt: one signed record of authority, actions, effects, verification, cost and rights for every unit of autonomous AI work.

    Classification: Design

    Evidence status: Proposal / Hypothesis

  • Research Note12 min read

    Mandates: Compiling Human Intent Into Bounded Agent Authority

    A research note on mandates: one object that compiles human intent into AI agent authorization, budgets, approvals, delegation limits, expiry and revocation.

    Classification: Design

    Evidence status: Proposal / Hypothesis

  • Methods Paper12 min read

    Evaluating the Evaluators: Reward Integrity for AI Agents

    A methods paper on reward integrity for AI agents: verifier false accepts and rejects, abstention, grader drift, expert disagreement and reward hacking.

    Classification: Synthesis

    Evidence status: Research Synthesis

  • Position Paper12 min read

    Faros: Researching How Intelligence Should Choose Intelligence

    A position paper reframing AI model routing as an execution-configuration decision across model, context, tools, verification, recovery, cost and risk.

    Classification: Synthesis

    Evidence status: Research Synthesis

  • Survey12 min read

    Why Learned AI Model Routing Must Beat Good Rules

    A survey of learned LLM routing: what RouteLLM, RouterBench and LLMRouterBench show, why strong rules are the right baseline, and how to test non-inferiority.

    Classification: Synthesis

    Evidence status: Research Synthesis

  • Research Proposal12 min read

    Counterfactual Replay for AI Agents

    A research proposal for counterfactual evaluation of AI agents: replaying completed tasks under alternative models, tools, context and recovery strategies.

    Classification: Proposal

    Evidence status: Proposal / Hypothesis

  • Position Paper13 min read

    What Makes AI Data Defensible?

    A position paper on the AI data moat: why volume is not defensibility, and eight axes, from rights to outcome density and transfer, that decide what compounds.

    Classification: Synthesis

    Evidence status: Research Synthesis

  • Research Proposal13 min read

    Skill IR: Toward Model-Independent Agent Capabilities

    A research proposal for Skill IR: AI agent skills as portable, testable capability contracts with permissions, verifiers and compatibility records.

    Classification: Proposal

    Evidence status: Proposal / Hypothesis

  • Research Proposal12 min read

    A Rights-Aware Dataset Compiler for AI Training and Evaluation

    A research proposal for a rights-aware dataset compiler: dataset builds that check purpose, lineage, consent and contamination at build time and fail closed.

    Classification: Proposal

    Evidence status: Proposal / Hypothesis

  • Benchmark Design11 min read

    VerifiedWork Context: Measuring What AI Agents Must Remember

    A benchmark design measuring what AI agents must remember: obligations, constraints, permission changes, superseded facts and revoked data, across compaction.

    Classification: Benchmark method

    Evidence status: Benchmark Design

  • Research Protocol12 min read

    How to Test Whether Learned AI Routing Beats Strong Rules

    A research protocol for an LLM routing experiment: strong-rule baselines, propensity-logged data, factorial cache controls, non-inferiority tests and decision rules.

    Classification: Protocol

    Evidence status: Protocol / Planned Experiment

  • Research Protocol12 min read

    How Should We Measure the Reliability of LLM Verifiers?

    A research protocol for LLM verifier reliability: stratified gold sets, false accept and reject rates, calibration, drift and exploit audits, and grader cards.

    Classification: Protocol

    Evidence status: Protocol / Planned Experiment

  • Research Paper12 min read

    Toward a Failure Genome of Software Agents

    A research paper proposing a multi-axis AI agent failure taxonomy, the failure genome, built on recent work on failure modes, attribution and critical steps.

    Classification: Synthesis

    Evidence status: Research Synthesis

  • Research Proposal11 min read

    Process Memory: Learning How Organizations Actually Get Work Done

    A research proposal for process memory for AI agents: mining completed, verified work into per-organization process models that guide plans and flag anomalies.

    Classification: Proposal

    Evidence status: Proposal / Hypothesis

  • Research Proposal13 min read

    Tenant Replay: Private Evaluation Inside Enterprise Boundaries

    A research proposal for private AI evaluation: replaying an enterprise's own historical agent tasks inside its boundary, with only bounded aggregates leaving.

    Classification: Proposal

    Evidence status: Proposal / Hypothesis

  • Research Protocol14 min read

    A Research Protocol for Model Change Assurance

    A research protocol testing whether model change assurance predicts live outcomes: paired replay, regression certificates, randomized rollout and false reassurance.

    Classification: Protocol

    Evidence status: Protocol / Planned Experiment

  • Research Note12 min read

    Why AI Routers Should Log Propensities From Day One

    Why AI routers need propensity logging from the first decision: candidate sets, selection probabilities and verified outcomes make off-policy evaluation possible.

    Classification: Synthesis

    Evidence status: Research Synthesis

  • Research Protocol12 min read

    How to Measure Whether AI Skills Survive a Frontier-Model Upgrade

    A research protocol for AI skill portability: a pre-registered event study of whether agent skills transfer, degrade or reverse after a frontier-model upgrade.

    Classification: Protocol

    Evidence status: Protocol / Planned Experiment

  • Research Protocol13 min read

    How Should We Measure How Much Context an AI Agent Actually Needs?

    A research protocol for the AI agent context budget: dose-response curves for five context strategies, with obligation retention and revoked-data exposure.

    Classification: Protocol

    Evidence status: Protocol / Planned Experiment

  • Research Protocol13 min read

    Testing Whether Recovery Knowledge Transfers Across Tools

    A research protocol for recovery transfer: does agent recovery knowledge carry to unseen APIs, tool versions and failure mechanisms, including when to stop?

    Classification: Protocol

    Evidence status: Protocol / Planned Experiment

  • Research Protocol14 min read

    How to Test Whether Verified Experience Improves AI Agents

    A research protocol on learning from experience for AI agents: do verified, failure-rich data beat raw traces at matched volume, on held-out work?

    Classification: Protocol

    Evidence status: Protocol / Planned Experiment

  • Research Proposal13 min read

    Private AI Improvement Without Raw Data Export

    Private AI improvement without raw data export: tenant-local learning, federated evaluation, secure aggregation and differential privacy, each with its threat model.

    Classification: Proposal

    Evidence status: Proposal / Hypothesis

  • Research Proposal15 min read

    Toward a Sovereign Improvement Protocol for Enterprise AI

    A research proposal for sovereign AI: how enterprise AI could improve across private deployments without raw data leaving the customer boundary.

    Classification: Proposal

    Evidence status: Proposal / Hypothesis

03Taxonomy

What the classifications mean

The classification says what kind of work a publication is. It is derived from each publication’s own evidence note. Only a measured result reports a measurement, and only within its stated scope.

Synthesis
Analysis of existing evidence and literature. It reports no new UpCube measurements.
13
Design
An architecture or system design. It describes how something could work; it reports no measurements.
3
Proposal
A proposed direction or hypothesis that has not been tested.
12
Protocol
A defined method for an experiment that has not been run, so it reports no results.
7
Benchmark method
A benchmark specification. It has not necessarily been run, and it implies no scores.
5
Measured result
Reports a measurement UpCube ran, only within the scope the publication states.
1

04Evidence

Evidence status

Every publication declares how strong its evidence is. Format and evidence status are separate: a benchmark design is a specification, not a benchmark result.

Measured / Empirical
Ethen ran an experiment, benchmark, or measurement and reports the result.
0
System Evidence
Evidence tied to a specific implemented or pinned Ethen system or build. It does not generalize beyond that build.
1
Research Synthesis
Analysis of existing evidence and literature. No new Ethen measurements are implied.
13
Protocol / Planned Experiment
The method is defined. The experiment has not been run, so no results are reported.
7
Benchmark Design
A benchmark specification. It has not necessarily been run, and no scores are implied.
5
Proposal / Hypothesis
A proposed direction, system, or hypothesis that remains untested.
15