What the data does, and does not, establish.
HeardTogether is a public-record project. Its value comes from assembling primary sources and labeling quotation against interpretation — not from proving a thesis. This page states the limits of the evidence up front, for both lanes, so no reader mistakes a documented pattern for a stronger claim than the record supports.
Across the whole site
- Not legal advice. Pointers to the Texas Open Meetings Act, the Texas Public Information Act, and AI-safety literature are pointers to public sources. They are not lawyering.
- Not clinical or psychological advice. Crisis resources are listed as a public service; HeardTogether is documentation, not a crisis service or a provider of any kind.
- Primary sources are the authority. Where this site summarizes, the linked source governs. If a summary and its source disagree, the source is correct and the summary is a bug to be fixed.
- Quotation versus interpretation is marked. Blockquoted text is verbatim from the source; everything else is the maintainer’s framing and is not presented as a finding by the source.
The AI ghost-pattern corpora
- Observational, not a controlled benchmark. The corpora demonstrate that a failure pattern occurs and recurs under describable conditions. They do not establish population-level rates, and they are not randomized or controlled experiments.
- Single-operator and, per corpus, single-vendor. The published Grok corpus is one researcher’s conversations with one platform. It is a deep case record, not a cross-vendor survey.
- Behavior-level and intent-agnostic by design. The claim is “the output exhibited pattern X,” deliberately weaker than “the model intended X.” Nothing here asserts a model has goals, feelings, or awareness.
- Automated classification under-counts. The project’s own model-assisted 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.
- Marker counts are upper bounds. Keyword and tag frequencies show where a term recurs, not that a failure was confirmed at each instance. A count becomes a finding only after the instances are read.
- Not affiliated with any AI lab. Honeycutt Ai Labs is the analyst of the model outputs quoted here, not a vendor for, partner of, or adversary retained against any model examined.
The civic record-pages
- A record, not an argument. The civic case files are not an argument for or against any project, vote, or official. They assemble the documents and ask precise questions.
- Not the residents’ group. Peer coalitions are linked as resources; HeardTogether does not mirror, coordinate with, or represent them.
- Time-sensitive and refreshed. Civic pages carry a “last refresh” timestamp and are updated from posted agendas; a claim true at one refresh may be superseded at the next. Read the date.
- Named individuals appear in public capacity only. Officials are named for on-the-record public actions. Private individuals are not identified. See Rules.
Corrections and falsifiers
Every case file lists open follow-ups that, if resolved against the page’s claims, would require updates — the page tells you where it could be wrong. Corrections that cite a primary source are processed promptly, the prior statement struck-through rather than deleted, and the change dated. Corrections without a source are not processed. See Rules and Contact.
The deposits linked from Methodology are the authority for any framework claim summarized on this site.