What the 1973 Acali Raft Experiment Reveals About Software That Holds Intimate Data

Ten strangers, 101 days, and a researcher who collected everything: the constraints behind a booking and messaging stack for a consent-led practice that keeps its clients unnameable

A 1973 raft experiment put ten strangers into an intimate container to test whether forced closeness produces violence. The group bonded, plotted a mutiny, and left one researcher holding the most sensitive record of eleven people's lives that anyone aboard would ever produce. The booking and messaging system I built for a consent-led intimacy practice had to solve the same problem backwards, starting with a data model that cannot join a session to a person.

An anthropologist put ten strangers on a raft in 1973 to see whether forced intimacy produces violence. Eleven were aboard the Acali when it left Las Palmas that May, six women and five men including the researcher, Santiago Genovés Tarazaga, who was 49. The crossing took 101 days and ended in Cozumel, Mexico. The participants filled in 46 questionnaires, many repeated weekly, and Genovés alone generated 1,042 pages of notes from the answers. The questionnaires asked who they felt closest to, who annoyed them most, whom they wanted to sleep with, and which of the other people they would remove from the raft if they could.

On day 51 he lost patience with the pace of his own experiment and began reading the confidential answers aloud to the group that gave them. Those answers had been collected under an expectation of confidentiality, which made the act a breach rather than a provocation.

Although the voyage was designed to produce violence, nobody on the raft attacked anybody else. The crew discussed killing him, weighing a knife, a shove over the rail and a lethal injection that one participant described as "with all of us on the syringe" (BBC Mundo). They decided against it, and Genovés later wrote that he had realised the only person aboard who had shown aggression was himself, and he withdrew from the experiment he had designed, stopped the questionnaires and fell ill in a way his subjects did not believe.

I read that story in a video, and I am writing about it because of what I built next: the booking system, website chat and WhatsApp reply agent for a consent-led tantra massage and somatic practice that books by request. The practice is built for people who do not want to be identifiable as clients. Genovés's apparatus was the opposite kind of machine, and that contrast turned out to be the most useful specification document I had.

Where the popular retelling goes wrong

The video is The Acali Raft Experiment Might Restore Your Faith in Humanity, uploaded by the channel Horses on 4 January 2023: 13:25 of first-person narration over the expedition's 16mm footage, and loose with facts in ways that would matter if you repeated them.

Its auto-generated captions render the researcher's name as "Santiago terrazaga" and the raft as "the Alkali raft." The raft was the Acali, Nahuatl for "house on the water," and the researcher's maternal surname is Tarazaga.

The video says the university that sponsored the work severed ties and that Genovés was essentially fired, and no source I could find supports that. Long-time colleagues said the publicity was damaging the university, and critics questioned his double role as both scientist and participant, but he kept his career at UNAM and later helped conceive the 1986 Seville Statement on Violence. He died in 2013 at 89.

The video also quotes a consent document requiring participants to "give their souls bodies and Minds" to Genovés. That wording appears in the video and in articles derived from it, and in no reputable source. What the document actually did, per the encyclopedic account, was record voluntary participation and release the sponsors from liability in the event of death. Nine people signed it and the captain, Maria Björnstam, did not. Her boyfriend called it a slave contract and gave her an ultimatum, and she went anyway.

The crew did not become lifelong friends who stayed in each other's lives, and the sources disagree on how to describe what they did become. BBC Mundo reports a bond that still holds, and most participants described the voyage as a success, while the Guardian notes the survivors had not met in 43 years when the 2018 documentary reunited them. Both accounts agree on the part that matters here: the group's hostility went to the person running the thing.

The instrument was the risk

Genovés instrumented the group rather than merely observing it: weekly questionnaires on sexual contact and masturbation, logged menstrual cycles, psychological testing before departure and for a week after arrival. Then he converted the private answers into a tool, reading them aloud to set people against each other, throwing a bucket of water at one participant and carrying reports of what others had said from person to person. POV Magazine's review of the documentary calls this what it is: consistent breaches of confidentiality, alongside a lack of informed consent. The same review notes, on his own account, that he had ongoing sexual interactions with at least one participant while running the study, a claim that rests on a single secondary source.

The published result was tepid: Genovés wrote that intimate relationships aboard did not significantly interfere with the group, and that friction "could neither be linked specifically to any postulated aggression instinct, nor to any particular biological stock." The sexual and privational engineering he designed produced none of what he wanted. What the raft had was a shared, finite, consequential goal, continuous contact, a hostile authority, and no exit. Muzafer Sherif's work on group formation and Gordon Allport's contact hypothesis describe conditions of this kind, including shared goals and cooperation, and a meta-analysis of 713 independent samples by Pettigrew and Tropp found that intergroup contact reliably reduces hostility. Applying that literature to eleven people on a raft is an inference, since the voyage had no control group, no baseline, and a sample of ten. Those conditions were met without any help from the questionnaires, which mainly produced a record of eleven people's inner lives and handed it to one man who then showed what an operator with that record can do.

What that means for software that holds intimate data

For a studio where the appointment is the sensitive fact, the data model is the safety surface, and the phrase people reach for in this space is already taken. Servury's Privacy is Marketing. Anonymity is Architecture. (December 2025) makes the point with the Mullvad warrant raid: a provider that cannot hand over what it does not have cannot be compelled to. This extends it into a domain where the constraint touches every screen.

A booking system for this business has a list of things it must be unable to do. It cannot join a session to a person. It cannot produce a searchable client table from the public side. It cannot let an analytics vendor infer that someone was looking at a practitioner page at 2am. It cannot have the operator's own name appear in a machine-readable graph about the brand, for reasons that have nothing to do with the clients.

That last constraint is where the two identities in this story nearly touch, and the whole system is built to keep them apart. The practice is anonymous by choice. The people who run it do not want to be the public face of a tantra business, and the software had to be built so that no machine-readable surface, no structured data, no analytics profile, no page template quietly invents the link that everyone else forgot to make. That constraint is enforced rather than promised: a structured-data rule asserts that a specific list of strings and hosts, including the operator's own name and personal domain, must never appear in the site's JSON-LD graph, in the working directory or live.

What actually shipped

The system is two deployments. A static marketing site of 141 pages sits on Cloudflare Pages, and an application on a separate host holds 75 API route files, with 17 Postgres tables underneath it. Two courses sell through Stripe entitlements and play as member-gated pre-signed URLs whose storage keys never reach the browser. Every count comes from commit 26dff6a of that repository, which is private, so a reader cannot re-run them; treat the numbers as claims about that commit rather than audited facts.

A deterministic gate runs before the model does. The WhatsApp reply agent's tool loop passes through a policy check that lives in code and is evaluated before any model call, so its result does not depend on what the model returns.

Approval runs through endpoints. Drafts are created for everything inbound. POST /api/drafts/:id/approve accepts an optional replacement body, so the reviewer edits and approves in one action, POST /api/drafts/:id/reject closes it, and GET /api/approvals is the queue. The rotation counter that spreads practitioner referrals increments on approval rather than on draft creation, so an unsent draft cannot consume someone's turn.

Nothing is deleted, and sending has four brakes. Inbound passes three mark-and-gate layers: a 15-messages-per-60-seconds limiter per phone number, an intent gate on length, link ratio and booking keywords, and soft spam flags. A message that fails the intent gate still produces a pending draft with auto-reply disabled, so nothing is dropped. On the way out there are four independent anti-send checks, including a double-send guard on the approve route and a check for whether the owner has replied in the thread since, and the send function makes no network call unless the send mode is explicitly set to live, with dry-run as the default in the code. The live deployment reports its own send mode as live, which was a human decision rather than a default.

Booking runs on a per-client token. /book requires an id and a token, so the flow is an offer link rather than a public availability page. Slots render in the practice's own timezone, and choosing one is a request with a fallback to the practitioner's calendar.

The first approach, and why it lost

The original plan for practitioner profiles was the obvious one: an editor writes, an admin clicks publish. What shipped is a revision-based lifecycle, draft to in_review to published, hidden or held, where an administrative hold blocks approval with an HTTP 409 and each transition has its own route. A single mutable field is not an audit trail, and in a business where a practitioner's public page is a promise about consent and limits, the transitions are the safety property.

The repository also carries its own failures as comments and gate test names. There is a parity test between the website concierge and the WhatsApp agent's copy of the same client-facing facts, and it exists because those copies drifted once and shipped; the test's own comment says so. A pipeline that piped test output through tail reported success on failing suites, because the pipe returned the exit code of tail and never the tests. And 285 Playwright runs in 12 days ate 88 percent of the repository's CI minutes before browser tests moved to a nightly schedule.

Where identity still re-enters

A claim of total anonymity in this system would be false, and one screenshot of the approval queue would end it. Three places require a person to be identifiable, and all three are operational rather than public: the phone number that a WhatsApp conversation carries, which a human then reads; the payment processor that takes a name and a card, with the entitlement granted by a signature-verified webhook on the raw request body; and the support, abuse-handling and refund work that normally requires knowing who you are talking to. Those three cannot be designed away, and a product description that omits them is inaccurate.

What is architectural rather than promised is narrower and testable: the public marketing surfaces accumulate no client identity, analytics run cookieless with person profiles off, the website chat keeps session-only memory, the booking path is a tokenised offer rather than a browsable client list, and the structured-data graph is gated against the operator's identity.

What the evidence says about the cost

Anonymity in a service like this is usually framed as a tax on engagement, so the relevant study is the one that tested it. Arnold and colleagues, in 404: Personal data not found (Internet Interventions, 2026), randomised 452 adults with heightened social anxiety to an eight-week self-help programme delivered either platform-anonymously or through accounts. Both versions improved substantially, and the account-based version was not superior: between-version effect size at post-treatment was 0.09, with a 95 percent confidence interval of minus 0.17 to 0.35. The authors' own caveats matter: the trial was powered for superiority and had no non-active control, so formal equivalence remains unestablished. And their most awkward finding deserves the same airtime: most participants preferred the account-based version.

The reason to design for anonymity anyway is the record of what happens when wellness data meets advertising. The FTC's 2021 action against Flo Health concerned fertility-tracking data shared with Facebook and Google after privacy promises. In July 2026 the FTC, Utah and California sued Hims & Hers over sharing sensitive condition data with advertising platforms while claiming privacy. And the regulatory net under most of this is thinner than clients assume, in the FTC's own words: many companies collecting health information, including fitness trackers and diet apps, "aren't covered by HIPAA. Not at all." The framework that does apply is the Health Breach Notification Rule and the FTC Act, after the FTC withdrew its 2021 health-app policy statement in September 2026.

No survey or paper I could find studies client expectations for tantra, bodywork or intimacy practices specifically. The 1,000-patient AMA and Savvy Cooperative survey is the closest proxy: more than 92 percent said health data should not be available for purchase, and only 20 percent knew which companies had access to their data. Reasoning from adjacent verticals to this one is an analogy, and I am labelling it as one.

What this build cannot claim

There is no published booking SLA in the repository or on the site, so no response-time promise appears in this article. The slot-first marketplace exists as a design document and was never shipped; the tokenised offer is what customers get. The zero-trace rule reads structured data, and the internal link checker skips absolute URLs, so the claim that ties are enforced holds for the machine-readable graph and should not be stretched to cover every anchor a future contributor adds. The live status endpoint currently reports its own release sha as unknown. The reviews file is empty, and the generator will not let a placeholder into it, so nothing here rests on collected ratings.

The raft's own limit is the honest end of the comparison. Genovés wanted a container that would manufacture conflict and it produced cohesion against him, which is a finding about how badly his instrument was built rather than a lesson about engineered intimacy. The people in the film were the ones who survived and chose to be filmed, which is the only outcome data that exists. And the raft's active ingredient was inescapability: no motor, no way back, no books, a deliberately public toilet. A visitor can close the tab at any point, and the raft gave nobody an equivalent exit.

The remaining claim for the software is smaller than the story that opened this article, and it is the one I would defend. An app can decide, in code, what its operator is able to know, and make that decision visible enough that a client can check it. Genovés held 1,042 pages about ten people and used them as a weapon inside a week of losing his temper. The equivalent risk here is three named surfaces, and most of the work was keeping every other surface off that list.

Read next: AI Agent Loops Fail When They Grade Their Own Work

Max Petrusenko writes about AI products, agent safety, and the data models that decide what a company is able to know. Read more on Medium, X, and LinkedIn.

If this changed how you think about what your own software is able to know about the people who use it, pass it on to someone building in a sensitive industry.