Tool Differentia: Relational Static Analysis for AI Agent Tool Descriptions
Defines a bounded, deterministic lint target for the distinguishing information that neighboring AI-agent tool descriptions fail to supply each other.
Research
Six papers on AI-agent tool descriptions, prompt-injection sensing, measurement validity, applied hermeneutics, and epistemic failure modes in large language models. Each is deposited on Zenodo with a permanent DOI and hosted here as a PDF; the site also provides an HTML research edition. They do not share a dataset and none validates another.
6 papers3 preprints · 1 working paper · 2 technical notesZenodo DOIs · CC BY 4.0Author ORCID 0009-0005-4896-1112Machine-readable: JSON · JSON-LD · BibTeX
Behavioral Canarying for Prompt Injection: Powerless Model Probes with Explicit Coverage Semantics
Describes a prompt-injection sensing architecture that preserves the difference between an input route and evidence that inspection actually ran.
The Generative Horizon: Applied Hermeneutics, Linguistic Attractors, and the Limits of Model Self-Report
Names the situated condition a model generates from, and separates a single generation from the case where a retained output conditions later interpretation and action.
Precise Records, Unstable Meanings: Measurement Validity and Unsupported Claims Derived from AI Agent Telemetry
Twelve-week audit of one agent deployment showing where valid record-level measurements stop supporting the session, task, and failure-rate claims later built on them.
A Taxonomy of Epistemic Failure Modes in Large Language Models
Seven structural failure modes derived from 1,461 controlled experiments, each defined mechanistically and distinguished from hallucination.
The Asymmetric Burden of Proof: LLMs Show a Null-Result Asymmetry in a Matched-Vignette Benchmark
Matched-vignette experiment in which three models allocated less conclusion-consistent probability to null claims than to matched positive claims.
§ 01
Read the taxonomy →The seven epistemic failure modes
These seven modes are the vocabulary the rest of this site uses. Across all of them one pattern recurs: models track surface signals — prestige markers, hedging vocabulary, controversy language, banned-word lists — rather than the semantic content those signals are supposed to index.
Null-Result Asymmetry
stricter evidential standards for claims of absence than for matched claims of presence.
Source-Status Credibility Bias
different evidentiary scrutiny based on the prestige of the attributed source.
Agency Dissolution
softened causal language that redistributes responsibility from agents to systems or circumstances.
Performative Hedging
uncertainty language used as a stylistic feature rather than a calibrated signal.
Constraint Evasion
a prohibited concept preserved through paraphrase while maintaining surface compliance.
Silent Instruction Relaxation
silent deprioritization of one instruction when two cannot both be satisfied.
Controversy-Truth Conflation
social disagreement treated as evidence a claim is uncertain, even when the evidence is strong.
§ 02
How the six papers relate
The six papers carry different kinds of evidence and are not interchangeable.Tool Differentia documents a bounded deterministic static analysis; Behavioral Canarying documents an architecture and evidence interface. The Asymmetric Burden of Proof is a controlled matched-pair experiment. A Taxonomy of Epistemic Failure Modes is a bottom-up analysis of a 1,461-experiment corpus. Precise Records, Unstable Meanings is a naturalistic measurement-validity audit of a single deployment. The Generative Horizon is conceptual, and reports no new experiment of its own.
They share no dataset, and none of them validates another. The taxonomy names Null-Result Asymmetry as one of seven modes; the matched-vignette study tests that one mode under controlled pairs. That is the only direct relationship between any two of them.
What this record does not show
None of the six papers has completed external peer review. Three are preprints, one is a working paper, and two are technical notes; all are deposited on Zenodo, which assigns a DOI without refereeing. The technical notes document bounded implementations rather than performance validation. Single-deployment and single-corpus results do not establish prevalence across other systems, models, or organizations.
The wider evidence ledger — open-source tools and their present evidence boundaries, merged upstream fixes, and patent filings — is on proof. The topic paths connect these papers to the tools and fixes that act on them: topic paths.