I Built a Tribunal So My AI Couldn't Ship Without Arguing With Itself
Two isolated AIs review every proposal my agent makes before it reaches me. The friction is the architecture.
My AI agent — call him Fen — can’t propose a new tool to me. Not directly. He writes a specification; the spec goes to two other isolated AIs who argue with it; then all three outputs arrive at my desk and I arbitrate. I built it that way on purpose.
This is unfashionable in 2026. The market signal for agentic systems is unmistakable: minimize the loop, ship faster, let the agent execute and let the operator selectively review. The friction in my system isn’t a cost of the design. It is the design.
I locked the architecture on May 3 as Decision MOPA62HE: Tribunal v1. Two isolated read-only agents — Critic and Architect — review every proposal Fen makes for a new tool or an irreversible action. Fen sees both reviews and writes a structured response: concession, defense, hybrid, or no-contest. The whole package goes to me. Fen does not approve his own proposals. The Tribunal doesn’t authorize execution. Every verdict path ends at me, with no exceptions in v1. The cost ceiling is five dollars per review. And the line in the decision document immediately after the architecture spec is the load-bearing one: the Tribunal makes Fen harder to evolve — on purpose.
What the decision locks is worth being precise about. It locks the topology — two isolated Managed Agents, not one agent with two prompts. It locks the authority boundary — all verdicts to me, no autonomous approval, no Tribunal bypass. It locks the access scopes — Critic and Architect get read-only access to the system, no write authority anywhere. It locks the output contract — versioned JSON schemas, no prose parsing, five sections from Critic, four from Architect, four response shapes from Fen. And it locks the friction-by-design principle itself: future relaxation of any locked element requires an explicit decision reversing this one. Silent erosion through scope creep, Tribunal bypasses, or “this one is small” exceptions is prohibited.
What the decision does not lock is the prompt prose. Prompts iterate as empirical artifacts after pilot data — tone tuning, role refinement, schema-conforming edits. Pattern locked, prompts iterable. That distinction is the whole point. I am not locking the verbal surface of the system. I am locking the shape of disagreement the system has to produce before I see any of it.
Tribunal v1 is one shape of this pattern. The four-vendor cold review is another. The first runs inside my system — Fen plus Critic plus Architect, internal Managed Agents gating agentic action. The second is external — Claude, ChatGPT, Gemini, Grok, one-shot, auditing a static artifact I am about to ship. They share the principle: isolated voices, written disagreement, operator arbitration. They differ in architecture, cost, latency, and what they gate. Both shapes live in the substrate this essay draws from. Tribunal v1 has been operating since May 3; verdicts are accumulating but are not yet the substrate I’m citing here. What I can cite is the pattern’s history, including the cold-review shape that has been running longer.
The pattern arrived earlier than the architecture. It had been running across two domains — livesworthknowing.org and one held private — for months before May 3 — different work, same shape. Multiple isolated AI voices arguing a proposal; operator arbitrating; convergence as signal, divergence as warning. Locking it as architecture was acknowledgment of working practice, not invention. The history matters because friction-by-design is not the kind of claim you’d risk on theory. You’d risk it because the pattern survived contact with the work, repeatedly, in domains where its absence had cost you something visible.
A day after the architecture locked, Wang et al. published HeavySkill (arXiv:2605.02396v1) — an empirical study of parallel-reasoning and sequential-deliberation that touched the same shape from a different direction. Two of its findings put empirical pressure on choices I had made on intuition. The paper finds a trade-off between iteration depth and information consistency in the deliberation phase — iterating risks corrupting the candidate-verdict surface. Tribunal v1 is single-pass for related reasons, though I had landed there on operator-load grounds before reading the paper. The paper also finds the deliberation role benefits from larger instruction-following generalists rather than narrow reasoning specialists — a finding that maps onto how Critic and Architect actually function: evaluating proposals against system context, not solving reasoning benchmarks. I wrote an addendum into the log, not as vindication, but as evidence the pattern was not merely private superstition.
Here is what I think most operators are missing in 2026, and what the rest of this essay defends. The choice is between two loops. One has a voice in it that can argue back from a position your priors did not generate. The other doesn’t. The first one is slower. It is also the only one of the two that produces judgment instead of plausible defaults — and judgment is what matters for the class of work that compounds over months and years.
The difference is not effort or output volume. The difference is whether there is a voice in your loop that argues back from a position your own priors did not generate. Without that voice, every output you produce is downstream of the priors you brought to the prompt. You can accept the output or reject it. You can read it. What you cannot reliably do is orchestrate it, because orchestration requires something to orchestrate against. The Tribunal is not the only way to build that something. It is the minimum version I trust inside my own system — isolated voices, written disagreement, and an operator who has to arbitrate instead of vibe-check.
The trap is that a one-AI loop produces output that feels validated. You asked a question, the model answered, you read the answer and judged it sound. That judgment is the discipline, you tell yourself. But the model you asked is operating on the priors you supplied in the prompt, and the answer you got is downstream of those priors by definition. Judging it sound is not adversarial review. It is a reading pass on a draft you essentially wrote. The signal you receive from agreement is corrupted at the source. You cannot tell whether the output is right or merely plausible, and plausibility is exactly what frontier models optimize for.
Let me make this concrete.
Last week I ran a four-vendor cold review on a naming decision for one of my ventures. The tribunal returned a verdict: retire the name. All four reviewed the etymology I’d supplied in the brief, all four cited the same connotation problem, all four returned the same recommendation. Three of them were emphatic.
I kept the name.
What happened next is the part of the Tribunal pattern that took me longest to understand, and the part operators have to sit with.
Cold-read testing of unprimed humans — my wife, my son, a group chain — returned readings the tribunal had not surfaced. The tribunal had been evaluating an etymology I’d supplied in the brief; in production, the name lands on phonetics and cadence, not stated etymology. The verdict was internally consistent and framing-contaminated. The signal that overrode it came from a class of evidence the tribunal could not access — and the tribunal could not access it because I had handed all four reviewers the same poisoned brief.
This is where the steel-manned skeptic gets to dismiss the whole thing, and the dismissal is half-right. The Tribunal does not remove framing risk. If every reviewer is handed the same poisoned map, convergence is not wisdom — it is synchronized error. The Tribunal makes framing risk visible only when the operator knows to look for it.
So what did the Tribunal actually do in the naming decision? It produced an artifact. The decision-log entry does not read “tribunal agreed, locked the name.” It reads “tribunal said retire, I overrode on a stronger signal, here is exactly why.” Six months from now, future-me reading that entry can reconstruct the disagreement and the override — what was argued, what was rejected, what signal won, where the framing risk lived. An operator who skipped the Tribunal cannot reconstruct any of that, because nothing was ever considered, only ratified. The verdict isn’t the value. The artifact is.
There is a discipline that goes with running the Tribunal that I have not named yet, and it matters more than the architecture: capitulation only happens on accuracy and structural issues; voice stays mine. When the tribunal flags a weak claim or a hidden assumption, I revise. When it flags a sentence that sounds like me but bothers it, I keep the sentence. The discipline is not “do what the reviewers say.” The discipline is “engage every finding in writing, then decide.”
The strongest counter-position runs like this:
The market signal is clear. Frontier models are racing to minimize loops, not add them. Every benchmark, every release note, every infrastructure investment in the industry is about cutting latency between intent and output. You’re paying five dollars and waiting thirty seconds for a multi-vendor review when you could just ship. You’re building artisan workflow in a world that is moving to assembly lines, and the assembly lines will win the same way they always do — by being faster and cheaper at scale. You’ve engineered a private moat around a structural disadvantage.
Most of it is right. The market signal toward speed is real. Frontier models are racing to minimize loops. The five-dollar tax and thirty-second wait are real costs. Where the counter goes wrong is in the metaphor itself — the assembly-line metaphor confuses what is being produced.
For commodity tasks — completion, summarization, draft emails, search reformulation — speed is correct and a Tribunal is theater. I would not run a Tribunal on a Slack reply. I do not run one on a CSV transformation. The cost-benefit collapses immediately for work where the output is judgeable against an objective standard the operator can verify in seconds.
The cost-benefit inverts for architectural decisions, strategic commitments, and operator-grade artifacts that compound or destroy substrate over months. For that class of work, speed is the wrong optimization. The thirty seconds I lose to a Tribunal review is not slowness. It is the entire window in which “argue back” is mechanically possible. Without it, I am not shipping architecture faster than the assembly line — I am shipping unaudited architecture, and the cost shows up six months from now as decisions I cannot reconstruct, defaults I did not consciously choose, and a substrate that has been quietly captured by a frontier model’s priors rather than my own.
The market signal toward speed is calibrated for the median use case, which is not the use case under discussion. The race to minimize loops is the right race for commodity work and the wrong race for the work that determines what your system becomes. The Tribunal is a posture of optionality, not friction-for-its-own-sake. You can always ship without it for the work that doesn’t justify it. You cannot retroactively install an “argue back” mechanism on decisions that have already happened.
The question is not whether to use AI. AI is in your loop whether you put it there or not. The question is what shape of loop you have built, and whether the shape carries the kind of work you intend to do with it.
If your loop is a conversation with a single model and you check the output before shipping, you have a one-AI loop. It is fast. It is appropriate for the work it was designed for. You should not believe you are orchestrating in it. You are reading.
If your loop has voices that argue from positions your priors did not generate, and an operator who arbitrates the disagreement rather than ratifying the agreement, you have a Tribunal. It is slower. It costs five dollars and thirty seconds. It produces an audit trail your future self can actually read.
I built mine on May 3 because the work I am doing will compound for years, and the cost of unaudited compounding is the only cost I cannot recover from. The friction is not a tax I am paying to feel rigorous. It is the price of being able to argue back at all.
The first time you override the verdict on a stronger signal, the practice is yours.