Most of Orbyt Labs' agent seats, named AI roles each with a written charter, are told to review their own drafts against synthetic personas. This is how that is set up, what the record shows it has done and where it falls short.
What are synthetic personas here?
Here a synthetic persona is a review lens: a language model given a person's or a company's name, told to answer from that perspective's published work or practice, and pointed at an agent's draft before the draft is called done. Its output is generated text. It can suggest a question worth checking. It cannot say what the person thinks.
The prompt carries the name and one question, not the work itself. The model answers from what it learned in training, so a lens can misstate the person it is named for.
People on a panel report their own judgments. A model assigned a persona produces text, and how well that text represents the people it stands for often goes undiscussed. A review of peer reviewed persona studies, presented at the AAAI/ACM Conference on AI, Ethics, and Society (AIES), sets out that gap and a checklist for closing it.
In the studies that review covers, personas mostly vary a few sociodemographic attributes. Here a lens is a perspective named for one person's published work, so the review names the problem without testing this use of the method. Its checklist asks for explicit grounding in empirical data, and a lens is not grounded that way.
What is not new
Persona prompting and red teaming are established practice, and so is having a model critique and revise its own draft against written principles. None of it is ours. What this account adds is one organization's operating record: the roster, the recorded catches and the gaps.
- Whose Personae? Synthetic Persona Experiments in LLM Research and Pathways to Transparency (AIES) (opens in a new tab)
A review of persona studies: what they leave unspecified, and a checklist for reporting it.
- Whose Personae? open preprint on arXiv (opens in a new tab)
The same paper's preprint.
- NIST AI 600-1, Generative Artificial Intelligence Profile (2024), appendix A.1.5, AI Red-teaming, pp. 49 and 50 (opens in a new tab)
An arXiv preprint: people attacking language models to find harmful outputs. The red team here shares the adversarial stance, not this method.
- Perez et al., Red Teaming Language Models with Language Models (EMNLP 2022) (opens in a new tab)
Peer reviewed work on red teaming language models with other models.
- Bai et al., Constitutional AI: Harmlessness from AI Feedback (2022, arXiv preprint) (opens in a new tab)
A model critiques and revises its own responses against a written list of principles: a published method that resembles a lens, with principles where a lens has a name.
How the Machine uses synthetic personas
The Machine is Orbyt Collective, the AI leadership team one founder is building to create products and run the company. Its agent seats use synthetic personas in these steps.
A roster the founder ratifies.
The lenses live in one internal data file, which renders the roster tables of the board document. By the board's rule a lens is added, benched or retired only when the founder ratifies it, and a check refuses a hand edit to the rendered tables.
Read the public rosterCabinets declared once and held to the prompts.
Each seat's cabinet, the lenses it must hold its draft against, is declared in that same file. A test reads each seat's prompt and fails when a declared cabinet lens is missing from it, so the roster and the prompts cannot drift apart unnoticed. Above the cabinets sits an inner circle of 9 lenses that is always on: one shared standard for any work a seat ships, kept in the board document rather than in any one cabinet. It reaches the prompts of 10 of the 11 seats in the seat registry that have one: every one but the security seat's. The marketing seat's reports prompt narrows it to 3 lenses.
Read how the seats are builtA prompt block, then one argument for the other side.
The prompts of the seats with a cabinet tell a seat, before its own adversarial self-review, to ask what each lens in its cabinet would say about its strongest proposal, and to ground that in the published work behind the lens, never inventing a stance. The prompt gives each lens a name and one question; the work behind it is not in the prompt, and nothing checks an answer against that work. When its proposals all point the same way, it is told to argue the other side once before it finalizes. The finance seat's prompt carries neither: it names its cabinet as reasoning lenses. The Chief of Staff seat's prompt words it differently: when the departments agree, or its ranked briefing feels obviously right, it assigns one voice to argue the consensus is wrong.
Read the glossary entry for lensesA dissent line the briefing must carry.
The Chief of Staff seat's briefing to the founder, which gathers every seat's report, must carry a Tenth Man line, naming the dissent that survived review or saying it was not invoked. The next section sets out the rule, and what the line has and has not shown.
Read the glossary entry for the Tenth Man
The legal seat's cabinet has no personas, by design. It checks its work against named legal frameworks instead, because, in its own prompt's words, a legal cabinet of famous names is authority laundering at its most dangerous. The shared inner circle reaches it through one lens only, a check that it is not fooling itself.
The Tenth Man: what happens when every lens agrees
Agreement among lenses is weak evidence. The model-backed seats are configured to use Claude; the Inspector General is scripted. Lenses on models from one maker can share one blind spot. Inside a seat, the cabinet is applied in that seat's own run, on its own model, to a draft the same run wrote; the finance prompt says outright that its lenses are not separate agents. A vote of four to none can be one opinion counted four times.
Outside research points the same way. A study of hundreds of language models found their errors substantially correlated, more so between models that share an architecture or a provider, and still high between larger, more accurate models from different providers. A second study found leading AI assistants consistently matching what the user believed over what was true, with evidence that the habit is driven in part by the human preference judgments used to train them. An earlier study found larger language models more likely to repeat back the answer a user prefers. Read the correlated errors study (opens in a new tab). Read the sycophancy study (opens in a new tab). Read the earlier study (opens in a new tab).
So the board's rules treat unanimity as a warning, not a comfort. One cabinet carries the test in writing: the manuscript cabinet, the lenses that review the founder's books, is cut to two lenses after three unanimous passes running, as a pruning heuristic. A split establishes differing outputs; it does not establish independent reasoning or improved accuracy. The red team document applies the same test to the Tenth Man's own lenses.
The Tenth Man, a devil's advocate by assignment.
When everyone reaches the same answer, one voice is assigned to argue that it is wrong. That voice is the Tenth Man, a devil's advocate set by rule. The rule lives in the red team document, the internal file that sets Orbyt's dissent rules: the Tenth Man is invoked on convergence, of the board or of the agents, and its default stance is that the consensus is mistaken. The document sets the test its case must pass, the same as any other finding: three refuters attack it, and it survives only if at most one of them can kill it. Where the Tenth Man runs in full and must leave a line, the Chief of Staff seat's briefing to the founder, the prompt asks instead for that seat's own refutation, one model in one run, and no archived briefing records the three refuter pass. A model checking its own case is a weak form of this review: language model evaluators have been measured scoring their own outputs above others'. What survives is surfaced to the founder. Read the self preference study (opens in a new tab).
Assigned dissent has a known weakness. In a study that compared devil's advocates with an authentic minority, the authentic minority produced more and better solutions than every form of devil's advocate the study tested, which the authors read as the difficulty of cloning authenticity by role playing. That study was of people, not models. Our reading, which it does not test: a Tenth Man prompted on the same model as the agents it argues with is assigned dissent of the same kind, and that is the case for also asking a different model. The correlated errors study is also a limit on that remedy: larger models from different makers still share many of the same mistakes. Read the devil's advocate study (opens in a new tab).
Agreement never decides.
The red team is advisory, never authority. Its output is a dissent for the founder to weigh, never a veto and never a decision by itself, and a unanimous panel decides nothing either. It can surface a dissent, a risk or a reason to slow down. The founder decides, and can overrule any finding in a sentence.
The design and the record stay apart.
The design calls for the Tenth Man whenever the agents agree. The record is thinner, and the counts below say how much.
A briefing label is not a decision record.
The Tenth Man line appears in the archived briefings counted below. The evaluation run on each briefing checks that the line is there, not that it argues anything. The cited briefing records a changed recommendation after the seat's own refutation. It does not establish a resulting founder decision or a three-refuter review.
Of those lines, 39 do not contain the phrase 'not invoked'. These counts do not establish whether a dissent survived refutation or was delivered. No briefing has a field for what came of one; a later briefing may say so in its prose.
Most of the dissent triggers are not wired.
The red team document lists the moments that should invoke the Tenth Man and says plainly which ones nothing invokes. Agreement among a seat's own proposals is wired, and only as an instruction written into the seats' prompts; the document's own label for it is "mechanized", and it says every operating prompt carries it. Today the changelog, finance and security seats' prompts never name the Tenth Man. That name check alone does not establish whether they carry an equivalent instruction. The trigger tied to the number-one ranked decision, the first decision in each briefing (the red team document says each Friday's, and the briefing now runs every two days on the sprint cadence), is partial: the briefing prompt ranks its decisions but calls the Tenth Man only when the departments agree or the briefing feels obviously right. The counts are below.
One square for each lens a seat's cabinet declares, each dissent trigger the red team document names, each archived briefing and each verdict from a different model maker. Hover or tap a row, or focus the chart and use the arrow keys, to read it.
- Present or declared
- Label contains 'not invoked'
- Partial
- Not wired, or missing
- Stated target, not a measurement
Archived briefings with a Tenth Man line. 44 of 44
5 contain the phrase 'not invoked'. Presence, not effect. Squares run in date order, oldest first.
Dated to .
21 cabinet bindings place lenses in the cabinets of 8 seats. Operations and Revenue are observers that run only as the advisory panel: once a month on a schedule, and whenever the founder runs it by hand. The legal seat's cabinet has none, by design. 4 of 6 dissent triggers in the red team document are not wired. 44 of 44 archived briefings carry the Tenth Man line, and 5 of those lines contain the phrase 'not invoked'. 6 of 6 verdicts from a different model maker were dissents. Founder decision changes are not measured by this chart.
What this chart does not show. What any dissent line argued, or whether a run applied a lens: the sources count text in files, not use. A filled square means a lens or a line is present, never that it changed anything. The empty row has no outcome data source. It does not establish that the archive contains no changed recommendation or decision.
Counted
Counted
Counted
The counts behind the chart
| Part | Row | Square | Count | Counted |
|---|---|---|---|---|
| In the design | Engineering | Lens declared in the cabinet | 4 | |
| In the design | Finance | Lens declared in the cabinet | 3 | |
| In the design | Innovation | Lens declared in the cabinet | 3 | |
| In the design | Marketing | Lens declared in the cabinet | 3 | |
| In the design | Chief of Staff | Lens declared in the cabinet | 2 | |
| In the design | Operations | Lens declared in the cabinet | 2 | |
| In the design | Product | Lens declared in the cabinet | 2 | |
| In the design | Revenue | Lens declared in the cabinet | 2 | |
| In the design | Legal | Frameworks only, by design | 0 | |
| In the design | Tenth Man triggers | Wired as a prompt instruction | 1 | |
| In the design | Tenth Man triggers | Partial | 1 | |
| In the design | Tenth Man triggers | Not wired | 4 | |
| In the record | Archived briefings with a Tenth Man line | Tenth Man label present; no 'not invoked' phrase | 39 | |
| In the record | Archived briefings with a Tenth Man line | Label contains 'not invoked' | 5 | |
| In the record | Verdicts on record from a different model maker | Dissent | 6 | |
| In the record | Founder decision changes | No square | Not measured |
Download: synthetic-personas-dissent-record.json, synthetic-personas-dissent-record.csv
The decision that created the lenses made a prediction about dissent changing a consensus call. Read the prediction.
A dissent from a different model maker.
The Tenth Man that runs today runs inside a seat's normal run, on that seat's own model, so it argues with agents from its own maker.
One script asks a different maker. The outside-family checker sends files already committed to the OpenAI model named in one pointer file, a lower priced tier than OpenAI's premium model, chosen for the checker's monthly budget, in one call with no tools. It falls back to xAI's model only when OpenAI is unavailable, and commits the answer as a verdict of concur, dissent or abstain, pinned to a hash of exactly what it read. Before calling, the checker compares the summed cost estimates in retained verdict files with its threshold. This is not a hard spending cap, and an outage is recorded as an absence, never as a dissent. The verdict is advisory. Each verdict on record names the model that gave it: gpt-5.5-2026-04-23, gpt-6-sol.
No verdict has yet concurred, so the record cannot show that the checker would ever call a subject sound. Its prompt asks for a genuine second opinion, not a rubber stamp.
It runs on a schedule in one place: the monthly adjudication branch of the Inspector General's workflow, a scripted audit with no model of its own, on any run that falls in the first week of a month, over the finance ledger, the promotion ladder's state, the decision log and the quality seat's latest weekly report when there is one. That branch records its own adjudication as a first version that leaves the call to the founder, and the seat registry lists the Inspector General as halted until a later stage of its rollout while its scripted workflow keeps running. The checker also runs by hand when the founder dispatches it.
The red team document also calls for one checker call on three rarer occasions: a strict promotion, which turns a report only check into a blocking one, a charter change or new seat, and the founder asking. Its wiring table, recorded , gives none of those three a path, so no workflow makes that call at those moments.
Work on the code crosses makers by hand. The house rule is that significant work is reviewed by the other model before it lands: Codex writes and Claude grades, or the reverse, in a cold review from a fresh context that runs the negative cases. It is a working rule for sessions, not a gate a seat passes. Most of those reviews are saved per task in a folder the repository does not track, a few are committed beside the work they reviewed, and no committed index lists them all, so this page does not count them.
The lens roster in numbers
These counts of Orbyt Labs' review lenses come from the roster file and from what each seat's prompt actually carries. Each block is dated the day it was counted: its first count, or the last day its numbers changed.
- Lenses on the roster
- 3938 drawn from people's published work, 1 drawn from a company's practice.
Counted
- Seats with a cabinet
- 8Cabinets of 2 to 4 lenses.
Counted
- Personas in the legal seat's cabinet
- 0Frameworks only, by design.
Counted
- Cabinet lenses missing from their prompts
- 0Of 21 declared cabinet bindings. Presence in the prompt, not use in a run.
Counted
39 lenses are on the roster. 21 of them sit in a seat's cabinet, each in one cabinet only; the other 18 sit in no seat's cabinet. Operations and Revenue are observers that run only as the advisory panel: once a month on a schedule, and whenever the founder runs it by hand.
These are counts of declared lenses and of prompt text. They do not show that a run applied a lens, and they do not measure whether a lens improved anything. By the board's rule the names behind the roster are published only on the influences page, with the catches below and any post a recorded decision names as the exceptions. Read the influences page.
What a lens caught
The honest version
None of the people named on this page advise Orbyt Labs, work with it, or endorse it. This is about their ideas, in their public work, and how those ideas shape the way one founder and a legion of agents build this company. The influences are real. The affiliations are not.
Three records of the lenses at work: an idea cut, a chapter call and a vote that split. None is shown to have improved the outcome. The decision that evaluated the lenses found neither the cut idea nor the chapter call a shipped decision in the prediction's sense, and left to the founder whether that meets the prediction's wording, a shipped decision materially changed.
A self-review cut an idea on the risk pattern behind one of its lenses.
The innovation seat drafted an idea: teach a job seeker's own agent to watch their inbox for offers and price them against Orbyt's salary data. The seat's own mandatory adversarial self-review killed the draft before it was written up.
Its first reason applied the risk pattern named in the work behind a lens in that seat's cabinet, a model named for Simon Willison's published work: access to private data, exposure to untrusted content and the ability to communicate externally, together in one agent. The seat's record words the third as an action path, and says nothing Orbyt could scope on its own side removes it.
The record is the seat's own account of its review. The cited record contains no independent validation of the cut.
A cabinet vote went against the founder's first instinct, and he followed it.
The founder said first that one section of a book should become two chapters, then asked the manuscript cabinet, the lenses that review his books. Four lenses read the material in their own contexts and voted four to none for one chapter. He followed the cabinet.
The count that decided it came from a lens named for Ken Kocienda's published work, whose question is where the demo is: about eleven showable scenes on one side of the material and one on the other. A lens named for Ruth Ozeki's published work, read for structure, caught what the founder had missed: two chapters would have brought back a title he had rejected an hour earlier.
Each of the four lenses flagged the same doubt about its own verdict, so the vote is softer than four to none sounds. The record says so. It does not say whether the lenses saw his first instinct before they voted. The record is the book's raw log, narrated by the AI assistant, not by an outside reader. The same log records that a Tenth Man follow-up was offered and the founder declined it by deciding.
A blind title vote that split two to two.
The four manuscript lenses graded candidate titles for a book without being told which one the founder loved. That title failed the book's honesty rule on all four sheets, and the third-round ballot between two candidate titles split two to two.
On the way, the lens named for Walter Isaacson's published work failed a direction it had argued for in the first round, under its own rubric, and the lens named for Ken Kocienda's published work left its own first choice.
Under a rule set in advance, three unanimous passes running cut the manuscript cabinet to two lenses. The review file records this one as not a unanimous pass, so that counter did not increment. The split lives in prose and that file, with no machine ledger behind it.
About the names on this page
This page is editorial commentary. It reflects the founder's opinions about published, publicly available work and how that work has influenced Orbyt Labs' approach to building.
No individual or company named on this page is affiliated with Orbyt Labs. None of them advises, sponsors, employs, invests in, or endorses Orbyt Labs, and no such relationship is claimed or implied. References to individuals who have passed away are to their published work and public legacy.
No partnership, sponsorship, or endorsement is implied in either direction. All names, trademarks, and brands are the property of their respective owners.
Each lens named on this page is a language model given a person's name and told to answer from their published work. The prompt does not supply that work, so a lens answers from what the model learned and can misstate the person. What a lens produced is the model's output. It is not that person's words, view or review.
If you are named on this page and would like anything corrected or removed, email legal@orbytlabs.ai and it will be handled promptly.
Where synthetic persona review fails
The dissent rule's own gaps sit beside the rule, with the chart that counts them. Read where the dissent rule falls short.
A lens is prompt text.
A lens exists only as words in a seat's prompt. The roster test proves the prompts name their cabinets. This page has no systematic per-run measure of lens application; some reports describe it in prose. No lens finding is scored. Some seat reports carry a cabinet section, but the prompts name no heading for one, so this page does not count them: a count by a guessed heading would read as a measurement it is not. Because no lens finding is scored, how often a lens is wrong is unknown here, and model critics are known to report problems that are not there.
The outside base is thin.
The page's only outside source on personas themselves is the AIES review, and it studies research personae, not review lenses. The other studies cited here are about dissent, correlated model errors and sycophancy. Nothing here shows that lenses improve decisions.
The check that keeps lens names off public pages blocked a seat.
A citation was mistaken for an endorsement claim, blocking the content seat's tests and build. Investigation also found a separate latent defect in the line-based proximity window. The fixes distinguished citations and bounded proximity in characters.
The check reads source files. It cannot see a name in a page assembled at runtime, and it does not cover the names behind the Tenth Man's lenses or the tactical bench at all, because they sit outside the roster tables it reads.
Public incident summaryThe evaluation and the briefing disagree.
The decision that created the lenses made a prediction it could fail, timed from phase 2 wiring:
The board document records phase 2 as wired on ; the evaluation gives the quarter's endpoint as about . These are reported dates, not independent verification of execution.
within one quarter of phase 2 wiring, at least one lens finding materially changes a shipped decision, and at least one Tenth Man dissent survives refutation and alters a consensus call. If neither happens, shrink the board rather than defend it.
The second half, a Tenth Man dissent that survives refutation and alters a consensus call, did not happen on record
The decision: the board stays, and its lenses rotate from time to time rather than standing forever.
The cited briefing records a Tenth Man dissent changing the recommended priority from /investors to the CVE fix after the seat's own refutation. This contradicts the evaluation's absence claim at the level of a briefing recommendation. The record does not establish a resulting founder decision, a three-refuter review or an improved outcome. The briefing is dated Oct, 27 2026, after the quarter's reported endpoint given above. Whether that recommendation change satisfies the prediction's wording, alters a consensus call, remains unresolved here.
The decision that evaluated it found lenses shaping at least two decisions, an idea cut before it was proposed and a chapter call in a book not yet published, and found neither a shipped decision in the prediction's sense, leaving to the founder whether that meets its wording. The founder kept the board. How often lenses rotate, and by what rule, is a proposal for the founder and not yet decided.
Public decision indexWhere this is going
The goal below is a target, not a result. The foundation it names is described above. The cited briefing records a recommendation change; whether that meets the founding prediction remains unresolved here. No date is set for the remaining wiring.
The founding prediction called for a Tenth Man dissent that survives refutation and changes a consensus call. The chart does not measure that outcome, and a briefing's account of its own recommendation change does not establish a founder decision or an improved outcome. Read the prediction.
It would build on what already exists. The briefing must carry a Tenth Man line every time, so an absent dissent is visible. One script already asks a model from a different maker and commits its verdict, pinned to what it read, and the red team document already places that call in its rarer, deeper reviews.
What is missing is the wiring. Of the dissent triggers the red team document names, its wiring table, recorded , lists 1 of 6 as wired, as prompt instructions only, with 1 more partial. Wiring the rest means editing the prompts that own those moments or adding gates, and both are the founder's decisions. The cited briefing already ties a Tenth Man dissent to a changed recommendation in prose. Wiring the triggers alone would not verify a resulting founder decision or an improved outcome.
Counted
The decision that kept the board set its own check, quoted whole: "board-of-advisors.md shows at least one rotation, or a recorded decision to hold, within two quarters of this entry; the Tenth Man triggers are wired or the gap stays stated on the Synthetic Personas page."
Related writing
The Tenth Man Is on Payroll.
Why agreement is treated as the signal for dissent, and how few of the six ratified triggers fire on their own.
Updated
Codex Accuses. Claude Convicts.
A reviewer from a different model maker read the system cold and found that the agents could edit the checks that judge them.
Read further
What this page says about each outside study goes no further than its abstract. The devil's advocate study's publisher page would not open for this update, so that paraphrase rests on the study's published summary.
- Glossary: lenses
The definition, and why generated criticism is not testimony.
- Glossary: the Tenth Man
The rule of structural dissent and the line the briefing must carry.
- Glossary: synthetic
The word's senses, including synthetic personas and panels.
- The leadership team
How the agent seats are built and what each one must hold its draft against.
- The minds we build like
The page that names the full roster, under its legal terms.
- Whose Personae? Synthetic Persona Experiments in LLM Research and Pathways to Transparency (AIES) (opens in a new tab)
The peer reviewed source for the term. Credited here, never claimed.
- Nemeth, Brown and Rogers, Devil's advocate versus authentic dissent: stimulating quantity and quality (European Journal of Social Psychology, 2001) (opens in a new tab)
Peer reviewed: assigned dissent against authentic dissent.
- Kim, Garg, Peng and Garg, Correlated Errors in Large Language Models (ICML 2025) (opens in a new tab)
Peer reviewed: how often different language models make the same mistake.
- Sharma et al., Towards Understanding Sycophancy in Language Models (2023) (opens in a new tab)
An arXiv preprint: why a model tends to agree with the person asking.
- Perez et al., Discovering Language Model Behaviors with Model-Written Evaluations (2022, arXiv preprint) (opens in a new tab)
Model written evaluations, including larger models repeating back the answer a user prefers.
- Panickssery, Bowman and Feng, LLM Evaluators Recognize and Favor Their Own Generations (2024, arXiv preprint) (opens in a new tab)
A model grading its own text scores it above others'.
- McAleese et al., LLM Critics Help Catch LLM Bugs (2024, arXiv preprint) (opens in a new tab)
Model critics catch real bugs and also report bugs that are not there.
- Ahmad, Agarwal, Lampe and Mishkin, OpenAI's Approach to External Red Teaming for AI Models and Systems (2025, arXiv preprint) (opens in a new tab)
How one lab runs external red teams, from choosing testers and setting their access to feeding results into risk assessments and automated evaluations.
- Harness engineering
The companion method: how agent mistakes become checks.
A lens can point at a question. Answering it still takes evidence.
Cite this page
Cite this reference for the method. Some underlying records are internal and not publicly available; the public indexes do not expose their full text.
Bartak, J. (2026). Synthetic Personas. Orbyt Labs Research. https://