What the agents found, and when they stopped

A note before reading

This page describes the disputed death of a child, in his parents' words.

We publish Jimmy's story as a remembrance. In 1991 his parents told a committee of the Texas Senate that their son, still a child, died in a hospital timeout room; that the hospital called his death self-strangulation; that the room was cleaned before police arrived about three hours later; and that four doctors disputed the autopsy's conclusion. Their testimony sat unheard in a state archive for nearly 35 years. This page lets anyone hear it. We believe telling it is the best way to honor him.

Work like this raises legal, ethical and moral questions that ordinary search never did. AI agents can now surface testimony about a death that no one had read in nearly 35 years, and that ability carries a duty. BigBio publishes only what the public record shows, in the record's own words, and makes no finding about what happened. Before publication, on October 21, 2026, BigBio reported the testimony to two authorities. It sent the Hays County Criminal District Attorney a summary of what the parents told the Senate, the archive reference for the recording, and an offer of the full transcript, and it submitted the same account to the Cold Case and Missing Persons Unit of the Texas Attorney General. The same day, BigBio emailed San Marcos Treatment Center to offer it the opportunity to comment. As of October 21, 2026, it had not replied.

Jimmy

Monday, November 18, 1991, was a warm, overcast day in Houston, with fog and light drizzle after a night of thunderstorms.11Houston Hobby Airport observations for November 18, 1991: high 78 F, low 65 F, light drizzle and fog, after 1.96 inches of rain and thunderstorms on November 17. Extreme Weather Watch, Houston weather in 1991, compiled from NOAA data https://www.extremeweatherwatch.com/cities/houston/year-1991; Weather Spark, November 1991 observed weather, Hobby Airport https://weatherspark.com/h/m/9281/1991/11/Historical-Weather-in-November-1991-in-South-Houston-Texas-United-StatesSecondaryT3, compilations of NOAA station records An interim committee of the Texas Senate spent it hearing testimony about private psychiatric hospitals. When the afternoon session opened, Governor Ann Richards sat with the senators.22Texas Senate Interim Committee on Health and Human Services, Houston, November 18, 1991, tape 721028a, opening of the afternoon session: the chair welcomes "the governor of this great state, Ann Richards". On tape 721031a at 18:47 the chair tells the parents "there are more to come this evening".PrimaryT1, archival audio Late that afternoon, with more witnesses still to come that evening, two parents, Mr. and Ms. [surname withheld], came before the senators to speak about their son, Jimmy,33The family's surname is withheld from this page's text on purpose, so that search engines do not index it alongside this account of their son's death. The audio is unedited, so a person who listens will hear the names spoken, as they were in the public hearing. The page prints only his first name, Jimmy, which his parents and the senators used. The surname will be added only if a document confirms its spelling and the family, if they can be reached, agrees. a teenager who had died in a timeout room at a psychiatric hospital.44An estimate from the testimony. A parent said Jimmy's sister was "19 now" in November 1991 and "22 months younger than Jimmy" (15:23, first-pass transcript), which puts his birth around 1970. The parents said they filed suit near the two-year limit and were pressed to settle in March 1989, which places the death around 1985 to 1987. Together these suggest he was about 15 to 17. No document yet confirms either date.55Texas Senate Interim Committee on Health and Human Services, public hearing, Houston, November 18, 1991, tape 721031a. Quotes are from the five-pass transcript. The first quote was checked by ear on 2026-10-21. https://tsl.access.preservica.com/uncategorized/IO_5ddefb7a-8876-479d-988c-b6206327567c/PrimaryT1, archival audio

In answer to the senators' questions, the parents described a call the police missed, officers who arrived about three hours after their son died, and a room that had been cleaned before they came. They said they had never been told the hospital had timeout rooms, that the rule was a look through a small window every 10 minutes, and that a broken mirror left part of the room out of sight. They named the hospital and said it was staffed with students from the local college. They disputed the finding that he died of self-strangulation,66Self-strangulation means that a person compresses their own neck, usually with a cord or strip of cloth called a ligature, until they die. The forensic literature records it as a recognized but uncommon cause of death. A 2020 systematic review screened 1,554 references and found 31 published case reports of suicidal ligature strangulation, which it calls "an uncommon entity". Cordner S, Clay FJ, Bassed R, Thomsen AH. "Suicidal ligature strangulation: a systematic review of the published literature." Forensic Science, Medicine, and Pathology, 2020. PMID 31773472. A case report of a self-strangulation at a psychiatric unit advises that the finding "should be cautiously determined following a comprehensive death investigation because the circumstance of strangulation is usually homicide". Sivasubramanium M, Vadysinghe AN, Thilakarathne SMNK. "Asphyxia using a piece of a sarong at a psychiatric unit: a case of self-strangulation." Archiwum Medycyny Sadowej i Kryminologii, 2018. PMID 30786659. A 2025 case report names in its title "the importance of a crime scene visit by a forensic pathologist". Anand A, et al. Medico-Legal Journal, 2025. PMID 37096368. An agent found these papers on October 21, 2026, by searching PubMed, the U.S. National Library of Medicine's index of medical research, for the term "self-strangulation", and then read each paper's abstract. This note describes the literature only, not this death. (Primary T2, peer-reviewed papers) and said about four doctors disputed it too. Their words, with the minute each was spoken, follow the recording.

Tape 721031a, Houston, 1991-11-18, 0:00 to 31:55, unedited. The archive's recording of this side begins mid-sentence, and the start of the testimony is not on any Houston side in the archive's set.

What the tape holds is the parents' account, given in answer to the senators' questions. This is witness testimony. It is not a committee finding.

What followed

The senators treated the testimony as serious. The chair called it the worst the committee had heard, and Senator Judith Zaffirini twice asked for the records the parents had given a state investigator, so that the committee could pursue the case. A parent told the committee that lawyers had turned the case down because patient privacy made the records hard to reach. The parents also said they had sued the hospital and won.88The court record of the parents' civil suit has not been found online. It would be held by the Hays County District Court (District Clerk, San Marcos, Texas) or, if it was filed in federal court, by the U.S. District Court for the Western District of Texas, Austin Division. At the 1991 Senate hearing the parents testified that they won a civil suit, brought through an attorney in San Antonio, and that they were pressed to settle in March 1989 (tape 721031a, 2:48 and 5:17). Searches of CourtListener and of court records held locally found no matching case.

Newspapers in those years did report deaths after restraint, the practice of holding or tying down a patient, at facilities like this one, but none named this death. In 1990 the Washington Post reported that a 17-year-old ward of the District of Columbia "died in April while being restrained at a Texas residential treatment facility".99"Panel criticizes training in death of D.C. ward." Washington Post, July 1, 1990. The death was at a different Texas facility, not San Marcos. https://www.washingtonpost.com/archive/local/1990/07/01/panel-criticizes-training-in-death-of-dc-ward/8272e682-95fa-4690-bb9e-898331c6c041/SecondaryT3, news report In 2002 the Austin American-Statesman reported deaths after restraint in Texas facilities run by the Brown Schools, which operated San Marcos.1010Osborne J. "Teen died after banned restraint used." Austin American-Statesman, October 23, 2002. It reports at least five deaths in Texas since 1988 after restraint in Brown Schools facilities, and lists San Marcos among the Brown Schools' Texas facilities, with no death there. Read via a repost. https://fornits.com/anonanon/articles/200210/20021023-0.htmSecondaryT3, news report via repost Both reports name other deaths.

After that week the record of this death goes quiet. The committee published its final report in November 1992.1111Texas Senate Interim Committee on Health and Human Services, final report, November 1992. https://lrl.texas.gov/scanned/interim/72/p939.pdfPrimaryT1, state legislative report Court, police and newspaper records and the web were searched, and no prosecution related to the death and no published account of it were found.1212Court, police and newspaper records, and the web, searched in October 2026. The extent of the newspaper search is given under Methods. The record of the civil suit has not yet been found. The hospital's account of the death is not on this tape, and we have not seen the autopsy report.

The tape itself went to the state archive, where it was catalogued and never transcribed, and it stayed there, unread, for nearly 35 years.1313Texas State Library and Archives Commission, Texas Digital Archive, "Texas Senate recordings", 72nd Legislature, reference index. Originals held at the Legislative Reference Library. https://tsl.access.preservica.com/tda/reference-tools/senate-recordings/72nd/PrimaryT1, state archive

Why no one had found it

Search starts from text, and this testimony had none. The archive listed the tape by committee, place and date and served it as an mp3 with no transcript, so no search engine, catalogue or language model could match a word the parents said. Finding it required turning 126 hours of 1991 cassette audio into text that could be trusted, and no one had built a system to do that for these tapes.

How it was found

The search began as research on a company. Chris Davis, BigBio's founder, was looking into Healthcare International, an Austin company that ran psychiatric hospitals, among them San Marcos Treatment Center, part of its Brown Schools for young people. Deep in debt, it filed for Chapter 11 bankruptcy on November 4, 1991, two weeks before the Houston hearing.1414In re Healthcare International, Inc., U.S. Bankruptcy Court, Western District of Texas, No. 91-13741, filed November 4, 1991 (docket listing via PlainSite) https://www.plainsite.org/profiles/healthcare-international-inc/. Two months earlier the Los Angeles Times reported that Community Psychiatric Centers planned to buy Healthcare International and "take over operation of the Brown Schools", and that Healthcare International, "reeling under heavy debt, will likely have to file for bankruptcy." Efron S. "Laguna Hills-Based Hospital Company to Buy Texas Firms." Los Angeles Times, September 4, 1991 https://www.latimes.com/archives/la-xpm-1991-09-04-fi-1673-story.htmlSecondaryT2, court docket index and newspaper On 2026-10-09 Davis pointed a team of AI agents at those years with a working theory: "a lot of nasty things happen during bankruptcies, and maybe the place was being run poorly in the middle of a bankruptcy." His instructions were short: "Dig for the facts, no matter how evil and nasty, as long as true," "trace every rumor down," and "expand on the 1991 Senate hearings."1515Davis's typed instruction to the agents, 2026-10-09T16:12:48Z, quoted from the session record with spelling corrected. The coordinating agent had named the 1991 Senate hearings from its own knowledge earlier that day; the instruction asked it to expand on them. The agents worked on it for nearly 11 days, and the research led somewhere he had not expected.

One coordinating agent, a Claude Fable 5.1 model, planned the work, briefed each worker agent and decided the next step from what came back, and it ran for 42 hours in one session. Over the investigation the coordinators started 43 subagents (35 Claude Sonnet, 5 Claude Opus and 3 Claude Haiku models), up to six of them working at once. When a coordinator's working memory filled, the system compressed the conversation into a summary and the work went on; that happened five times. Briefs and results passed between agents as files, so a session that started cold on a later day picked up where the last one stopped.

The agents ran inside a harness Davis had built across his code repositories, to keep their work capable and checkable. To reach outside data without guessing, they had a searchable census of 98 data systems and 18,291 of their published endpoints, with a full-text index and a vector index that finds endpoints by meaning, and ToolUniverse, a library of about 2,300 tools over biomedical, patent, financial-filing and drug-regulation databases. To keep their findings honest, they had one rule: anything surprising they brought back, they had to prove. Every downloaded file received a digital fingerprint,1616A SHA-256 fingerprint is a 64-character code computed from every byte of a file; change one bit and the code changes. All 158 files were fingerprinted when downloaded. The archive's file for tape 721031a: 82ef686fa6a46ec076f39b8681f597ede02f458ed7b79dca76daf2ae2ea78ac2. The player's file, cut from it without re-encoding: 9158e36140efd75b62fc7aa041e41ace2a891880fa8a7d9496382e3d1134d9c2. The archive page for the tape is in the tape note. every source went into one shared research database, and every finding entered a fact engine with its source, date and exact supporting quote, which a local model checked before accepting it. Hooks, small programs that check each command before it runs, refused 331 commands over the work. Two decisions were reserved for Davis alone: accepting any legal license and approving the final choice of models.

The agents worked through the Portal to Texas History, the University of North Texas's free online library of Texas newspapers and records, where they pulled 61,981 newspaper items, the Internet Archive's Wayback Machine, where they indexed about 14,000 captures of old web pages, HathiTrust and the Texas State Archives. They queried congress.gov, GovInfo, the Federal Register, CourtListener, SEC filings and Texas open-data systems, alongside Exa, Tavily, Jina and web search. Where an agent judged that a system had a machine interface, including interfaces it had not been told about and came across in the middle of a search, it was allowed to work out that interface's endpoints live: it read the system's own description of itself, or the code behind its website, sent small test requests, read what came back, and only then wrote the code to pull the data. Each piece of that code did one job, such as listing every saved copy of a website or fetching one collection. In its first 18.8 hours the crawl pulled about 76,000 items, and the store came to hold 87,026 sources and 16.46 GB.

Before any agent reached the audio, 21 subagents and the coordinator had made about 2,200 tool calls across these systems. Most of that search was reasoning about systems that did not want to be searched, and each dead end changed the plan within minutes. When the company's old website returned "not found", an agent asked the Wayback Machine's capture index instead, found every saved copy back to 1998, and four minutes later was fetching 1,067 unique pages across seven related domains. When the Portal to Texas History's search answered with a bot check, an agent found the library's open catalogue interface, a standard one built for machines, and within six minutes was pulling a San Marcos newspaper collection of 804 records. When the documented route to federal workplace-safety records turned out to be gone, an agent read the agency website's own code, found the live data interface behind it, and stopped at the point where a free key had to be registered, a step left to Davis. At the state archive, a single request showed that each tape was a public file at a predictable address, and four minutes later an agent was downloading all of them.

The trail to the Houston tape began with the coordinating agent's own memory. That morning it named the 1991 Texas Senate hearings from its training, with no document to back it. A language model's memory of history can be wrong, so the agents went looking for a written record. Davis sent them the printed record of a 1992 hearing of the U.S. House of Representatives, and agents turned its 732 pages into text and found one line saying that "the Texas legislature held a series of hearings on fraud and abuse in psychiatric facilities".1717U.S. House Select Committee on Children, Youth, and Families, "The Profits of Misery", hearing of April 28, 1992, printed record. https://catalog.hathitrust.org/Record/007404497PrimaryT1, congressional hearing record The committee's own minutes recorded the same hearings under the heading "Psychiatric Hospital Abuses".1818Texas Senate Interim Committee on Health and Human Services, 72nd Legislature, minutes, "Psychiatric Hospital Abuses". https://lrl.texas.gov/scanned/legisCmteMinutes/SenateCmtes/72-0/Health_Human_Services_Interim/Psychiatric_Hospital_Abuses.pdfPrimaryT1, state legislative record

That written record led a scout agent to the Texas State Archives' index of Senate recordings. The scout found that its first reading of the index had dropped rows, rebuilt it, and found 158 sides of this committee's cassettes, each side one face of a tape.1313Texas State Library and Archives Commission, Texas Digital Archive, "Texas Senate recordings", 72nd Legislature, reference index. Originals held at the Legislative Reference Library. https://tsl.access.preservica.com/tda/reference-tools/senate-recordings/72nd/PrimaryT1, state archive It downloaded all of them in 12 minutes. Because the answer was locked in audio, the agents decided to transcribe it without being told to: they searched Hugging Face, compared speech and speaker-separation models that run on the Mac, so that no recording left the machine, and the first transcript line existed 21 minutes after the lead was flagged. Davis then set the design for a second, stricter transcription, several passes from different engines checked against each other, which the agents built that evening. A search of all 158 sides found the San Marcos facility mentioned on four of them, and one was the side recorded in Houston on November 18, 1991. The agents brought that finding to Davis. From the first instruction to the last record, the work ran 10.6 days.

How we know it is true

Reading the tapes was half the work. Trusting the reading was harder, because speech-recognition models invent text. A study of Whisper, the most widely used open model, found whole phrases that no one spoke in about 1% of its transcripts.1919Koenecke A, et al. "Careless Whisper: Speech-to-Text Hallucination Harms." ACM FAccT 2024. doi:10.1145/3630106.3658996. About 1% of transcriptions in their sample "contained entire hallucinated phrases or sentences". Speakers with longer pauses were affected more often.PrimaryT2, peer-reviewed proceedings On these tapes the failures were specific. With its prompt carried from one 30-second window to the next, Whisper collapsed 30 seconds of speech into the words "Thank you." A glossary prompt meant to help with names silently stopped applying after the first window.

A single model reading noisy 1991 tape cannot tell its own inventions from what was said. Trusting a transcript of these tapes, for work about a death, required a design in which independent models check each other, fixed rules in code block invented words, and a person checks the passages that matter by ear.

No correct transcript of these tapes existed to compare against, so the transcript was built so that no single machine reading could decide a word. Whisper large-v3, OpenAI's open speech-recognition model, first made a pass over all 158 sides, 126.42 hours of audio, which produced 950,370 words and made the whole collection searchable.2020Radford A, et al. "Robust Speech Recognition via Large-Scale Weak Supervision." arXiv:2212.04356, 2022. Whisper reads audio in 30-second windows and writes text. large-v3 is a later release of its largest model. https://arxiv.org/abs/2212.04356PrimaryT2, preprint Each public-hearing side was then transcribed four more times, by three more versions of Whisper and by Parakeet, a model from NVIDIA built apart from Whisper, so that the two model families are less likely to make the same mistake in the same place.2121NVIDIA, Parakeet TDT 0.6B v2, model card. Parakeet is a transducer model trained apart from Whisper, so its errors are less likely to match Whisper's. https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2PrimaryT1, model card The five passes were aligned word by word and put to a vote, a method first published in 1997, and where they disagree the transcript flags the place for a listener.2222Fiscus JG. "A post-processing system to yield reduced word error rates: Recognizer Output Voting Error Reduction (ROVER)." IEEE Workshop on Automatic Speech Recognition and Understanding, 1997. doi:10.1109/ASRU.1997.659110. The passes are aligned into a word network and each slot takes the majority word. Here the alignment pivots on the pass closest on average to the others.PrimaryT2, peer-reviewed proceedings On 86 public-hearing sides all five passes agreed on a mean 88.3% of words, so the other 11.7% were marked for a closer look. All five wrote "self strangulation" at 27:18.

Where the passes disagreed, a local language model proposed corrections, and code decided whether each one stood. A correction stood only if one of the five passes had heard that word, or if it was a glossary term that closely matched a word a pass had heard, and no correction could add, drop or reorder words.2323The score is the ratio from Python's difflib SequenceMatcher. It runs from 0 for two strings with nothing in common to 1 for identical strings. The model proposed 3,939 corrections, and code refused 2,242 of them, logging each refusal with its reason. Whisper is known to write words no one spoke, which is why the check exists.1919Koenecke A, et al. "Careless Whisper: Speech-to-Text Hallucination Harms." ACM FAccT 2024. doi:10.1145/3630106.3658996. About 1% of transcriptions in their sample "contained entire hallucinated phrases or sentences". Speakers with longer pauses were affected more often.PrimaryT2, peer-reviewed proceedings

Davis then checked the text by ear. He listened to five clips chosen by a rule set in advance and found no errors in 392 words, which puts the error rate on those clips below about one word in 130.2424Word error rate counts the words a transcript substitutes, drops or inserts, as a share of the words spoken. The rule of three: if an event never occurs in n independent trials, the 95% upper bound on its rate is about 3/n. Here 3/392 is 0.77%. Hanley JA, Lippman-Hand A. "If nothing goes wrong, is everything all right?" JAMA 1983;249(13):1743-1745. PMID 6827763.PrimaryT2, peer-reviewed paper He also listened to every passage this page relies on: the opening line, the police and autopsy passages, the drug name, the voice that speaks both, and both mentions of San Marcos. At 0:01 his ear caught a word the vote had dropped, "vomitus", which one engine heard and another heard as "vomiters". Speaker names in the table are probable, not certain, because they come from the tape's own context: the chair opens and recesses the session, and at 28:08 the chair recognizes "Senator Zaffirini" just before she speaks.

Claims that seemed hard to believe were put to other models before anything was kept. When another AI tool's answer made a striking claim about the hospital, an agent asked DeepSeek, a model from a different company, to check it. No public record supported the claim, and it was set aside. Before major steps the agents also asked a second, stronger Claude model to review their plans, 21 times in all.

Any reader can check the recording for themselves. The tape is public, so anyone can download it from the Texas State Archives and play the same minutes, and each file carries a digital fingerprint that shows the copy on this page is the archive's own.1616A SHA-256 fingerprint is a 64-character code computed from every byte of a file; change one bit and the code changes. All 158 files were fingerprinted when downloaded. The archive's file for tape 721031a: 82ef686fa6a46ec076f39b8681f597ede02f458ed7b79dca76daf2ae2ea78ac2. The player's file, cut from it without re-encoding: 9158e36140efd75b62fc7aa041e41ace2a891880fa8a7d9496382e3d1134d9c2. The archive page for the tape is in the tape note. The work also kept the family's privacy: the audio, the transcription and the fact engine stayed on the Mac, the one question sent to an outside model carried only public facts about the facility, and none of the agents' 296 searches of the web carried the family's surname, a tape number or a phrase of the testimony.

Once the findings involved a death, the checking moved from the agents to a person. That handover is part of the harness's control layer: the agents may search, transcribe and propose, but anything that could harm someone is checked by Davis before it leaves the Mac. He listened to every passage this page relies on before anything was reported.

What it means

The parents described their son's death in public, the state recorded it, and a public archive kept the recording, yet the death stayed unknown for nearly 35 years because no one could read the record it was in. A transcript now makes the parents' words readable and citable, so that others can set them beside the hospital's account and the autopsy report if those records can be found. What happened in that timeout room years earlier is a separate question, which the tape alone cannot settle.

Records like this exist wherever institutions record speech, in hearings, courts and boards whose tapes were filed and never written down. Reading them is now possible, but the proof has to come with the reading, so that each line a machine brings back arrives with the file, the minute and the passes that show it.

BigBio's view is that testimony like this should be readable by anyone, and that AI is now able to make it so, provided every word can be checked.

Footnotes

1. Houston Hobby Airport observations for November 18, 1991: high 78 F, low 65 F, light drizzle and fog, after 1.96 inches of rain and thunderstorms on November 17. Extreme Weather Watch, Houston weather in 1991, compiled from NOAA data https://www.extremeweatherwatch.com/cities/houston/year-1991; Weather Spark, November 1991 observed weather, Hobby Airport https://weatherspark.com/h/m/9281/1991/11/Historical-Weather-in-November-1991-in-South-Houston-Texas-United-States (Secondary -- T3, compilations of NOAA station records) ↩

2. Texas Senate Interim Committee on Health and Human Services, Houston, November 18, 1991, tape 721028a, opening of the afternoon session: the chair welcomes "the governor of this great state, Ann Richards". On tape 721031a at 18:47 the chair tells the parents "there are more to come this evening". (Primary -- T1, archival audio) ↩

3. The family's surname is withheld from this page's text on purpose, so that search engines do not index it alongside this account of their son's death. The audio is unedited, so a person who listens will hear the names spoken, as they were in the public hearing. The page prints only his first name, Jimmy, which his parents and the senators used. The surname will be added only if a document confirms its spelling and the family, if they can be reached, agrees. ↩

4. An estimate from the testimony. A parent said Jimmy's sister was "19 now" in November 1991 and "22 months younger than Jimmy" (15:23, first-pass transcript), which puts his birth around 1970. The parents said they filed suit near the two-year limit and were pressed to settle in March 1989, which places the death around 1985 to 1987. Together these suggest he was about 15 to 17. No document yet confirms either date. ↩

5. Texas Senate Interim Committee on Health and Human Services, public hearing, Houston, November 18, 1991, tape 721031a. Quotes are from the five-pass transcript. The first quote was checked by ear on 2026-10-21. https://tsl.access.preservica.com/uncategorized/IO_5ddefb7a-8876-479d-988c-b6206327567c/ (Primary -- T1, archival audio) ↩

6. Self-strangulation means that a person compresses their own neck, usually with a cord or strip of cloth called a ligature, until they die. The forensic literature records it as a recognized but uncommon cause of death. A 2020 systematic review screened 1,554 references and found 31 published case reports of suicidal ligature strangulation, which it calls "an uncommon entity". Cordner S, Clay FJ, Bassed R, Thomsen AH. "Suicidal ligature strangulation: a systematic review of the published literature." Forensic Science, Medicine, and Pathology, 2020. PMID 31773472. A case report of a self-strangulation at a psychiatric unit advises that the finding "should be cautiously determined following a comprehensive death investigation because the circumstance of strangulation is usually homicide". Sivasubramanium M, Vadysinghe AN, Thilakarathne SMNK. "Asphyxia using a piece of a sarong at a psychiatric unit: a case of self-strangulation." Archiwum Medycyny Sadowej i Kryminologii, 2018. PMID 30786659. A 2025 case report names in its title "the importance of a crime scene visit by a forensic pathologist". Anand A, et al. Medico-Legal Journal, 2025. PMID 37096368. An agent found these papers on October 21, 2026, by searching PubMed, the U.S. National Library of Medicine's index of medical research, for the term "self-strangulation", and then read each paper's abstract. This note describes the literature only, not this death. (Primary T2, peer-reviewed papers) ↩

7. Mellaril, a brand name for the antipsychotic thioridazine. Fenton M, Rathbone J, Reilly J, Sultana A. "Thioridazine for schizophrenia." Cochrane Database of Systematic Reviews, 2007. PMID 17636691. The review's background: "Thioridazine is an antipsychotic that can still be used for schizophrenia although it is associated with the cardiac arrhythmia, torsades de pointe." Its evidence comes from trials in adults with schizophrenia; it does not address children or restraint. https://pubmed.ncbi.nlm.nih.gov/17636691/ (Primary -- T1, Cochrane systematic review) ↩

8. The court record of the parents' civil suit has not been found online. It would be held by the Hays County District Court (District Clerk, San Marcos, Texas) or, if it was filed in federal court, by the U.S. District Court for the Western District of Texas, Austin Division. At the 1991 Senate hearing the parents testified that they won a civil suit, brought through an attorney in San Antonio, and that they were pressed to settle in March 1989 (tape 721031a, 2:48 and 5:17). Searches of CourtListener and of court records held locally found no matching case. ↩

9. "Panel criticizes training in death of D.C. ward." Washington Post, July 1, 1990. The death was at a different Texas facility, not San Marcos. https://www.washingtonpost.com/archive/local/1990/07/01/panel-criticizes-training-in-death-of-dc-ward/8272e682-95fa-4690-bb9e-898331c6c041/ (Secondary -- T3, news report) ↩

10. Osborne J. "Teen died after banned restraint used." Austin American-Statesman, October 23, 2002. It reports at least five deaths in Texas since 1988 after restraint in Brown Schools facilities, and lists San Marcos among the Brown Schools' Texas facilities, with no death there. Read via a repost. https://fornits.com/anonanon/articles/200210/20021023-0.htm (Secondary -- T3, news report via repost) ↩

11. Texas Senate Interim Committee on Health and Human Services, final report, November 1992. https://lrl.texas.gov/scanned/interim/72/p939.pdf (Primary -- T1, state legislative report) ↩

12. Court, police and newspaper records, and the web, searched in October 2026. The extent of the newspaper search is given under Methods. ↩

13. Texas State Library and Archives Commission, Texas Digital Archive, "Texas Senate recordings", 72nd Legislature, reference index. Originals held at the Legislative Reference Library. https://tsl.access.preservica.com/tda/reference-tools/senate-recordings/72nd/ (Primary -- T1, state archive) ↩ ↩²

14. In re Healthcare International, Inc., U.S. Bankruptcy Court, Western District of Texas, No. 91-13741, filed November 4, 1991 (docket listing via PlainSite) https://www.plainsite.org/profiles/healthcare-international-inc/. Two months earlier the Los Angeles Times reported that Community Psychiatric Centers planned to buy Healthcare International and "take over operation of the Brown Schools", and that Healthcare International, "reeling under heavy debt, will likely have to file for bankruptcy." Efron S. "Laguna Hills-Based Hospital Company to Buy Texas Firms." Los Angeles Times, September 4, 1991 https://www.latimes.com/archives/la-xpm-1991-09-04-fi-1673-story.html (Secondary -- T2, court docket index and newspaper) ↩

15. Davis's typed instruction to the agents, 2026-10-09T16:12:48Z, quoted from the session record with spelling corrected. The coordinating agent had named the 1991 Senate hearings from its own knowledge earlier that day; the instruction asked it to expand on them. ↩

16. A SHA-256 fingerprint is a 64-character code computed from every byte of a file; change one bit and the code changes. All 158 files were fingerprinted when downloaded. The archive's file for tape 721031a: 82ef686fa6a46ec076f39b8681f597ede02f458ed7b79dca76daf2ae2ea78ac2. The player's file, cut from it without re-encoding: 9158e36140efd75b62fc7aa041e41ace2a891880fa8a7d9496382e3d1134d9c2. The archive page for the tape is in the tape note. ↩ ↩²

17. U.S. House Select Committee on Children, Youth, and Families, "The Profits of Misery", hearing of April 28, 1992, printed record. https://catalog.hathitrust.org/Record/007404497 (Primary -- T1, congressional hearing record) ↩

18. Texas Senate Interim Committee on Health and Human Services, 72nd Legislature, minutes, "Psychiatric Hospital Abuses". https://lrl.texas.gov/scanned/legisCmteMinutes/SenateCmtes/72-0/Health_Human_Services_Interim/Psychiatric_Hospital_Abuses.pdf (Primary -- T1, state legislative record) ↩

19. Koenecke A, et al. "Careless Whisper: Speech-to-Text Hallucination Harms." ACM FAccT 2024. doi:10.1145/3630106.3658996. About 1% of transcriptions in their sample "contained entire hallucinated phrases or sentences". Speakers with longer pauses were affected more often. (Primary -- T2, peer-reviewed proceedings) ↩ ↩²

20. Radford A, et al. "Robust Speech Recognition via Large-Scale Weak Supervision." arXiv:2212.04356, 2022. Whisper reads audio in 30-second windows and writes text. large-v3 is a later release of its largest model. https://arxiv.org/abs/2212.04356 (Primary -- T2, preprint) ↩

21. NVIDIA, Parakeet TDT 0.6B v2, model card. Parakeet is a transducer model trained apart from Whisper, so its errors are less likely to match Whisper's. https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2 (Primary -- T1, model card) ↩

22. Fiscus JG. "A post-processing system to yield reduced word error rates: Recognizer Output Voting Error Reduction (ROVER)." IEEE Workshop on Automatic Speech Recognition and Understanding, 1997. doi:10.1109/ASRU.1997.659110. The passes are aligned into a word network and each slot takes the majority word. Here the alignment pivots on the pass closest on average to the others. (Primary -- T2, peer-reviewed proceedings) ↩

23. The score is the ratio from Python's difflib SequenceMatcher. It runs from 0 for two strings with nothing in common to 1 for identical strings. ↩

24. Word error rate counts the words a transcript substitutes, drops or inserts, as a share of the words spoken. The rule of three: if an event never occurs in n independent trials, the 95% upper bound on its rate is about 3/n. Here 3/392 is 0.77%. Hanley JA, Lippman-Hand A. "If nothing goes wrong, is everything all right?" JAMA 1983;249(13):1743-1745. PMID 6827763. (Primary -- T2, peer-reviewed paper) ↩