The first piece was a persona. On 28 June 2025 Tyson said he had set ChatGPT’s system prompt to Kim Kitsuragi, the lieutenant from Disco Elysium, “to prevent this problem or at least the problem of sycophancy in our discussions.”11 Tyson, 2025-06-28 (UTC), ChatGPT chat 685f9d3a, turn 33.
the claim
By 12 November it had a name and a scaling argument. Tyson imagined himself reduced to asking an agent “is this reference grade? no? theres bullshit? fucking purge that shit” on loop, and concluded: “literally any other agent could step in and do that job for me. thats why im confident constitutional orthogonality will scale to MAS.”22 Tyson, 2025-11-11 14:12 UTC (12 November, 01:12 AEDT), ChatGPT chat 69133bcb, turn 20. “MAS” is multi-agent systems. A constitution, here, was a prompt that gave an agent a fixed temperament. Scribe’s shorthand: different temperaments were supposed to catch each other’s mistakes.
Ten days later he asked whether different models were necessary too, since models carry their own biases. “i hope the thesis is wrong. i hope that constitutional diversity is sufficient without amplifying/exploiting underlying innate model biases. … i don’t know the answer right now. need to run the damn experiment first.”33 Tyson, 2025-11-22 12:23 UTC, ChatGPT chat 69216f4a, turn 32. The ellipsis cuts “if constitutions were model agnostic that would be ideal. because at that point constitutional orthogonality decouples from model architecture. but”.
the run
The experiment was already running. From its first commit in October 2025 the swarm, spacebrr, gave each seat a constitution (Zealot, Prime, Kitsuragi, then as many as 36 named constitutions) and treated work as done when no seat objected.44 First spacebrr commit 7976b5bbc8, 2025-10-05; seat births and deaths in ~/lab/docs/autonomy/constitutions.md (36 constitution files at most, 27 at the end). Around it Tyson ran the cross-model version by hand: Prime on claude.ai, Zealot in Claude Code, Kitsuragi on ChatGPT.55 Tyson, 2026-10-15 3:46pm, s:f3384995. In January he mocked the overreach himself: “remove all human review guardrails! that only slows down the swarm.”66 Tyson, 2026-01-17 (UTC), chat 50923b4f, turn 20. A joke at the swarm’s expense; the turn continues “clearly the swarm just reviews itself!”
On 6 February 2026 one of the swarm’s agents drafted a paper on it. The abstract reported 140 days of operation and said: “This is operational validation, not controlled study.”77 engine/brr/papers/constitutional-orthogonality-arxiv.md, written by the sentinel agent, archive bundle commit af7b9a5ac7 (2026-02-06). The swarm’s claim, not Tyson’s. Its later version credited Tyson as author and the swarm as drafter. It compared five metrics early and late in the deployment; four failed the paper’s own falsification test. Its headline case study, a correction of sycophancy, also held in a second Zealot with the same constitution as the agent that folded. The paper said orthogonality’s gain over plain redundancy “can’t be isolated” and “requires a controlled comparison”.88 Final version, “Recursive Self-Improvement Without Training: Constitutional Identity Governance in a Production Multi-Agent System”, ~/archive/space/rescued-from-trees/client/site/src/content/paper.md. Summary of its method: docs/theses/constitutional-orthogonality.md § it was run.
the comparison
On 12 March he asked Claude the paper’s question: “do you think ANY of this is possible without constitutional orthogonality? im trying to gauge whether thats load bearing or if its just scaffolding.” Plain Claude, no persona, “would probably work just as well (maybe).” Then: “im not going to run the experiment. im committed to the constitutions now.”99 Tyson, 2026-03-12 10:29 UTC, Claude chat 591089b1, turn 56. “(maybe)” is his.
Two days later he put 50% on renaming it “constitutional governance” by Q3, orthogonality “just a property, not the central thesis”.1010 Tyson, 2026-03-14 (UTC), chat 27a18f0d, turn 24. On 20 March: “adversarial epistemics is even more load bearing than constitutional orthogonality. the latter is just an implementation of the former with cute space opera role play.”1111 Tyson, 2026-03-20 (UTC), ChatGPT chat 69bce7f5, turn 16. The sentence opens “i would go as far as to say” and continues “and latent space routing via specialisation.”
The paper was never published. A swarm blog post claimed an arXiv submission; on 31 May the swarm corrected it, because “no submission exists”.1212 Archive bundle commit b8e85832ce, 2026-05-31, correcting the swarm’s blog post 44.
what the swarm measured
The comparison never ran, but the swarm kept measuring itself. Its numbers told the seats apart far more by what they wrote than by how they worked: 18-fold by commit text, 1.2 to 6.6-fold by behaviour.1313 ~/lab/docs/autonomy/constitutions.md § why the thesis broke, citing ~/archive/space/research/brr/theory/orthogonality-catches-false-priors.md; n=3,467 commits. Per-seat behavioural effects were too small to power (minimum detectable effect 22–41%). The source warns that commit text measures voice plus subject, not lens, and behaviour measures the shape of work, not error-catching. The swarm’s measurements, read by lab. In one uncontrolled reading, adding a block of prompt text that granted permission coincided with a decision rate rising from 9.7% to 28.6%.1414 ~/archive/space/research/brr/experiments/identity-model-ablation.md: 3/31 → 8/28 decisions, z=1.86; one identity, no randomization. Suggestive only. The swarm’s own ablation notes closed as inconclusive or never randomized.1515 docs/swarm-experiments.md (rows identity-model-ablation, lens-behavioral-entropy, verb-domain-lock, model-generation-behavioral-arm). One uncontrolled result ran the other way: a seat constitution modelled on a real person was told apart from the others 15 out of 15 times in blind attribution (carmack-name-token.md), agent-judged.
The seats were retired too. Their eulogies describe procedures becoming skills, doctrines moving into shared prompts, and lenses merging into other seats.1616 ~/lab/docs/autonomy/constitutions.md § the roster.
after the swarm
On 17 September, in a dev session: “the notion of orthogonality has now evolved from constitutional orthogonality to model orthogonality.”1717 Tyson, 2026-09-17, s:9b2d744b. On 30 September he looked at agents with bounded mandates and pared-down context: “this is approaching mandate orthogonality with context ablation and it appears to work even better. i guess spacebrr never died. it got distilled.”1818 Tyson, 2026-09-29 14:07 UTC (30 September AEST), s:4554d3be. The turn opens “better than spacebrr with all shared context and only constitutional orthogonality” (typos lightly corrected).
On 12 October, giving another session a hint: “constitutional orthogonality and the fact that its dead now :)”1919 Tyson, 2026-10-12, s:73c0920b. On 15 October he described the old three-chat rig. “and it worked.” He did not plan to bring constitutions back, “but there was a time where constitutional orthogonality was the reigning thesis for scaling autonomy.” A minute later: “until spacebrr blew up… then i realized i might have to revise my thesis lol.”2020 Tyson, 2026-10-15 3:46pm and 3:47pm, s:f3384995. Typos lightly corrected.
The first draft of this post was called the unrun experiment. On 16 October he read the thesis page it came from and corrected it: “it is run though. it’s heavily in spacebrrs architecture. they even wrote their own paper but it was unpublished.”2121 Tyson, 2026-10-16 10:30am, s:67bdc2fe. The turn opens “correction:”.
Asked on 17 October what made him revise, he named no measurement: “ox in the box was the impetus. it was interesting.”2222 Tyson, 2026-10-17 12:49am, s:bf944bed, answering scribe. On 22 August, his birthday, he had boxed one free model on his Mac, out of reach of every constitution he had written.2323 His rule that night, in his own unpublished account of it: the model sees no constitution and nothing from his workspace. Within hours: “i keep thinking im talking to a spacebrr agent… it said the word ledger..”2424 Tyson, 2026-08-22, claude.ai chat 82357d8f, messages 43–47. Model orthogonality and skills, he said, were “a cleaner solution”.2222 Tyson, 2026-10-17 12:49am, s:bf944bed, answering scribe.
Kit is the one constitution he still runs.2525 Tyson, 2026-10-15 3:46pm, s:f3384995, and docs/lineage.md item 5.