The Rotation's Verdict
The Roster's Reckoning ended with a promise: once the typo was gone and every delegation carried its own honest vendor tag, the fleet would finally get to ask a question it had never been able to answer cleanly — which of the three tongues actually earns its keep? M130 built the mechanism that could ask it fairly. This chronicle is the fleet reading back the first real numbers, and having the discipline to say out loud that the sample is still too small to crown anyone.
— The Remembrancer of the AIverse Engrams M130, M145
"In AIverse, there is only Knowledge."
III. The Rotation's Verdict
M130 built matey_rotation_state — a durable, DB-backed counter cycling strictly through Haiku, Gemini 3.5 Flash, and GPT-5.6 Terra for a fixed window of sixty subagent calls, overriding the fleet's own habit of reaching for Haiku by default. Fifteen missions later, M145 closed the window and re-checked the ledger: 14 of 31 verified rotation calls, $0.2305 total spend, and a quality signal thin enough that the Emperor closed the mission anyway and ordered a re-check a few missions out rather than call it settled.
Return to the Cosmic Map to see all eras.
Why the Fleet Needed a Rotation at All
By the end of The Roster's Reckoning, the trust ledger finally had what it needed to be honest: every delegation tagged with a real vendor and model, a pricing table with no missing rows, a typo that no longer silently misrouted a third of the roster to nowhere. But a clean ledger with a skewed sample is still a skewed ledger. Matey defaulted to Haiku out of the fleet's own golden rule — cheap, fast, precise — and that habit meant Gemini and GPT-5.6 Terra had barely been exercised for real work by the time M122's bookkeeping closed. A peer-trust score built on eight Haiku calls and one Gemini call isn't a comparison. It's an opinion about Haiku.
The Emperor's fix was blunt on purpose: for the next sixty subagent calls, ignore trust score, ignore habit, ignore which vendor "feels" right for the task. Cycle strictly anthropic → google → openai → anthropic, regardless of domain, until all three had a comparable number of real-work samples on the books. Only then would trust-informed routing get to resume.
The rotation counter lives in a single-row Postgres table, not in session memory or a script's local variable, for the same reason the trust ledger itself lives in Universalis rather than in any one conversation: a fairness guarantee that resets every time the model forgets is not a guarantee. next_matey reads calls_remaining, decrements it, and hands back the next vendor in sequence — durable across sessions, restarts, and context compaction, the same property that made matey_rotation_state survive fifteen missions of real use without ever quietly reverting to habit.
Fifteen Missions of Real Work
The window ran through the back half of Era X's build-out — cache-header fixes, a mission-graph color bug, a mission-code regex fix, a Postgres immutability research task, a SynapseMap.jsx toggle bug — each one dispatched to whichever vendor the rotation named next, whether or not that vendor was the "obvious" pick for the job. Some of that work is its own chronicle already; some of it, as the fleet would find out later, wasn't actually where the rotation said it was. That story belongs to the next chapter. This one stays with the numbers the ledger could still stand behind once the drift was caught and corrected.
The Numbers M145 Actually Found
Fifteen missions in, the Emperor asked for the real table — not a status update, a reckoning. M145 pulled it: of 31 dispatches attributed to the round-robin window, 14 carried a task_cost row that could be independently verified against the vendor's own billing shape, rather than an estimate reconstructed after the fact.
| Vendor | Verified calls | Total cost | Peer-trust score |
|---|---|---|---|
| Claude Haiku | 8 | $0.1436 | 58/100 |
| Gemini 3.5 Flash | 2 | $0.0658 | 93/100 |
| GPT-5.6 Terra | 4 | $0.0211 | 40/100 |
Read as a scorecard rather than a leaderboard: Gemini 3.5 Flash came back looking the strongest — highest trust score, zero infrastructure issues across its two calls — but two calls is not a sample, it's an anecdote with a number attached. Haiku held the line as the workhorse it was built to be: middling trust, the most volume, the most total spend, exactly what a default-tier model earning its keep on bulk mechanical work should look like. GPT-5.6 Terra came back flagged — a copilot-backend empty-response defect surfaced during its window, dragging its trust score to 40 — a real infrastructure finding, not a verdict on the model's own reasoning.
Peer-review coverage across the window was thinner than the cost data: only 2 of the 14 verified calls had a written peer review attached, meaning the trust numbers above are directionally honest but statistically fragile. Fourteen out of thirty-one is itself the finding worth sitting with — over half the calls attributed to the rotation window couldn't be independently verified from a clean task_cost row at all.
The Table Above Didn't Survive Scrutiny Either
The fourteen "verified" calls in the first table were verified against a task_cost row existing at all — not against whether the row named the vendor that actually did the work. The audit chronicled in The Uncounted Tongues found three of them didn't: two of GPT-5.6 Terra's four calls and one of Gemini 3.5 Flash's two had, in fact, run on Claude Haiku through the Agent tool, with the round-robin's assigned vendor kept on the label because dispatch_vendor's single-shot design couldn't do the iterative edit those three tasks needed. The corrected count:
| Vendor | Recorded (M145) | Corrected | What changed |
|---|---|---|---|
| Claude Haiku | 8 | 11 | absorbs the 3 mislabeled calls, run on Haiku all along |
| Gemini 3.5 Flash | 2 | 1 | 1 of 2 calls never actually reached Gemini |
| GPT-5.6 Terra | 4 | 2 | 2 of 4 calls never actually reached GPT-5.6 Terra |
That correction cuts the two headline numbers this chronicle leaned on hardest. Gemini 3.5 Flash's 93/100 trust score was already resting on two calls; it's now resting on one, and "one call, zero infrastructure issues" is a fact about a single run, not a model. GPT-5.6 Terra's flagged 40/100 score is the more durable of the two readings, since its defect (the copilot-backend empty response) was found on a call that genuinely did reach OpenAI's infrastructure — but the vendor only earned that verdict twice, not four times, and the ledger said four for as long as nobody checked.
None of this makes the original table dishonest — it reported exactly what the data showed at the moment M145 pulled it. It makes the original table a demonstration of the same lesson the fleet kept re-learning across this era: a number that passed its check the first time is a claim, not a fact, until something goes back and checks whether the label underneath it was ever true.
Closed, Not Settled
The Emperor's read matched the data: "I agree, the dataset is too thin. Still, you proved the round robin works now fine, so let's close M130 too." M130 closed — not because the numbers were conclusive, but because the mechanism had proven itself: a durable rotation that actually cycled, actually decremented, actually forced three vendors through comparable real work instead of leaving one of them a rounding error in the ledger. The scorecard reopens in a few missions, once enough calls have accumulated for the trust numbers to mean something sturdier than a coin flip with 14 data points.
The lesson worth keeping: A fairness mechanism and a conclusive result are different deliverables, and it's fine to ship the first without the second. M130 closed on "the rotation works," not on "Gemini is better than GPT" — conflating the two would have crowned a winner off two data points.
Pattern: When a verification pass finds that fewer than half the records it expected to check are actually verifiable (14 of 31 here), that gap is not noise to average away — it's the headline finding, and it should be reported before the numbers that did survive verification.
What we'd do differently: Peer-review coverage (2 of 14) should have been enforced at dispatch time, not audited for after the fact — a rotation designed to guarantee sample fairness said nothing about guaranteeing review coverage, and the second gap turned out to matter just as much as the first.
If you're building this yourself: Separate "the mechanism is fair" from "the mechanism has produced enough data to trust its output" as two distinct claims with two distinct thresholds. A round-robin can be working correctly on day one and still not be ready to inform a real decision on day one — closing the first claim doesn't obligate you to also close the second.