Applied hermeneutics
Language is not only what an agent reasons about. It is part of what the agent reasons through — and that has engineering consequences.
1 paper · 2 tools · 3 terms
Claim 01
For a language-model agent, the represented situation is not only an input; it is part of the medium the next output is produced through.
Instructions, prior turns, retrieved records, tool results, summaries, and corrections form a situation the model interprets and generates from at once. The Generative Horizon paper names that situated condition and treats it as an engineering object rather than a metaphor.
The horizon exists within a single generation and does not require persistent memory. That qualifier is load-bearing: without it the idea collapses into ordinary context accumulation, which is a different thing.
What this does not show
The paper is a preprint published 30 July 2026 and reports no experiment of its own. Nothing in this catalog tests the framework against a deployment, and hermeneutics is a long-standing discipline — no coinage is claimed for the term.
Claim 02
Translating that condition into engineering yields four requirements: provenance, status, revision, and authority.
Where did this representation come from. What is it — a measurement, an inference, a summary, an authorization. Has it been corrected or superseded, and is that visible. And how much may it govern, given the evidence behind it.
The paper's rule for the fourth is the strict one: bind a representation's authority to evidence proportionate to what it is allowed to decide. A summary may inform. It should not authorize.
What this does not show
These are requirements argued for in a preprint, not a specification anyone has implemented end to end. No system in this catalog satisfies all four.
Claim 03
One of the four requirements is operationalized: hermeneutic distinguishes a correction from a supersession and gates the next response on it.
hermeneutic mines correction triples from chat logs — what was said, what the correction was, what should hold now — and checks a proposed response against them before it ships. A correction that was made once and then forgotten is the ordinary case it exists to catch.
This covers revision, and only revision. Provenance, status, and authority remain unaddressed by any tool here.
What this does not show
hermeneutic is a drift check, not proof that the proposed response is correct or complete. Running it tests one requirement in one narrow form; it does not test the paper's broader argument.
Claim 04
A second requirement — keeping a retrieval distinguishable from its source — is what a zero-LLM store buys you.
If recall paraphrases, provenance is already lost at the moment of retrieval: you cannot tell the record from a reading of the record. fidelis returns the stored passage itself, which keeps that distinction available downstream.
The connection is the paper's own; fidelis is cross-linked from it rather than assembled for this page.
What this does not show
fidelis preserves provenance for stored retrieval only. It says nothing about provenance or status for inferences, summaries, or corrections, and its install-to-recall behavior has not been independently verified.
Every artifact on this path
| Artifact | Kind | What it establishes | Status |
|---|---|---|---|
| The Generative Horizon | Paper | Applied hermeneutics as requirements for agent systems | preprint |
| hermeneutic | Tool | Correction-versus-supersession gating on the next response | PyPI 0.1.7 |
| fidelis | Tool | Retrieval that stays distinguishable from its source | tagged source only |
| Recursive interpretive conditioning | Term | Retained output changing a later horizon | illustrative vector |
If this is the failure you are seeing, tell us the system and the symptom. Book a 30-minute call, or send a note and get a written read instead.