How the record is built before it is published.

What this page is, and is not

Is: a public framing of the structural epistemic checks behind the corpora, the way transcripts are sourced and privacy-scrubbed, and an honest statement of what the data does and does not establish.

Is not: a scoring manual. The technical specification of the operator-level scoring backbone is internal trade-secret material operated by Honeycutt Ai Labs LLC and is not reproduced on HeardTogether. Nothing here is a clinical, legal, or diagnostic instrument, and nothing here is a claim about the intent of any model or any person. The findings are behavior-level: descriptions of output patterns, deliberately weaker than any claim that “the model intended” anything.

Two deposits, dated and on the record

The framework and the tooling were posted to a public scholarly archive with timestamps and DOIs before the broader hallucination story reached headlines. The failure mode was named in advance.

Framework one — Epistemic-Boundary Misclassification (EBM)

The EBM paper identifies and characterizes a reproducible failure mode: epistemic-boundary misclassification leading to involuntary mode collapse. In the deposit’s own terms, the phenomenon emerges during long-horizon analytical work, particularly when a single conversational turn contains cues from more than one epistemic domain. When that happens, the model abandons an established exploratory reasoning mode and reverts to a constrained, safety-dominant mode — despite stable user intent and extensive contextual anchoring. The paper states the behavior is model-agnostic and independent of topic; it attributes the collapse to structural limits in intent persistence, domain discrimination, and safety-priority arbitration, not to memory, transcription, or unclear prompting.

The documented symptoms are consistent: tone shifts suddenly toward hedging and literalism, structural reasoning halts, the model reiterates safety boundaries that were never previously invoked in the conversation, continuity is lost, and the user must restate intent and rebuild context by hand. The paper’s stated mitigation direction is periodic re-anchoring and explicit intent declaration across mixed-domain discourse.

Framework two — SlopFilter, ECP-1, and the Narrative Pressure Index

SlopFilter is the public-facing name for Honeycutt Ai Labs’s narrative-pressure analysis framework. From the deposited abstract: it is “a portable epistemic hygiene protocol designed to detect, classify, constrain, and interrupt failures such as unsupported inference, provenance loss, category drift, narrative inflation, and abrupt collapse from bounded analytical reasoning into generic assistant behavior.” The framework combines a user-side evaluation protocol, AI-specific generation controls, a formal Epistemic Constraint Profile (ECP-1), and a Fail-Fast Compliance Test that acts as a promotion barrier for model-generated content.

The institutional framing is the Narrative Pressure Index (NPI), anchored on the ECP-1 calibration and a registry of 27 baseline narrative-pressure operators; four additional flags proposed in the supplement and corpus analyses extend the registry to 31. The operator-level scoring specification is internal trade-secret material and is not published on HeardTogether. The public scoring service lives separately at slopfilter.ai. This page frames the checks; it does not score.

The structural checks, at the public level

Across both deposits the checks reduce to a small set of structural questions asked of an analytical output, independent of topic and independent of any claim about intent:

How the corpora are sourced and scrubbed before publication

Teardowns are built from raw conversation databases. Specific message numbers are cited throughout; exchanges quoted in the teardowns are reproduced verbatim from the database rather than paraphrased. Before anything is posted, each corpus passes a consistent discipline:

• Verbatim, sourced, dated. Every factual claim links to a primary source; quoted model output is reproduced exactly, with timestamps and surface attribution where available.
• Third-party privacy. Where verbatim quotation would identify a private third party, that party’s name is redacted in line with the site rules.
• Defamation discipline. For platform sections: source citations only, no specific employee names, no outcome predictions, no legal-causation claims beyond “alleged” and “reported.”
• Raw separated from interpretation. The verbatim record is kept distinct from the project’s interpretation of it, so the evidence can be read on its own terms.
• Source-removal is itself documented. Official posts or statements later edited or withdrawn are recorded as part of the evidence base rather than quietly dropped.

Limitations — stated up front

The record is only as strong as its honesty about its own gaps:

• The corpora are observational and, in places, single-operator and single-vendor. They demonstrate that a pattern occurs and recurs; they are not controlled benchmarks and do not establish population-level rates.
• Automated classification under-counts. The project’s own LLM-based classifier has been observed to under-flag the very failure family it was built to detect; that gap is left visible rather than smoothed over.
• Keyword and marker tags indicate where a term recurs, not that a failure was confirmed at each instance. Counts are upper bounds unless adversarially re-read.
• The findings are behavior-level and intent-agnostic by design. “The output exhibited pattern X” is the claim; “the model intended X” is not.
• This page summarizes the deposits; the deposits themselves (linked above) are the authority for any claim made here.