The Roster's Reckoning
The Triumvirate Accord told of a ledger built well the first time — five milestones, each verified before the next began, closed clean. This chronicle tells what happened after the ink dried. A command-center graph wanted faces for its new voices, and the fleet reached for the obvious fix five separate times before it found the one line of text that had been wrong since the beginning. No mission in this chain was wasted. Each one narrowed the search until the truth had nowhere left to hide.
— The Remembrancer of the AIverse Engrams M122, M126–M134
"In AIverse, there is only Knowledge."
II. The Roster's Reckoning
M122 built the trust ledger the Triumvirate needed: model_registry, peer_trust, peer_reviews, benchmark_seed, and a living probe-task benchmark, seven tables and four CLI scripts, five milestones, all independently verified. It closed with the Emperor's approval and not a single open thread. Then a small cosmetic ask — put vendor logos on the command-center graph — dragged a chain of five more missions behind it, each fixing something real, none of them the actual root cause, until the ninth mission in the chain found a single mistyped model name sitting in a table nobody had thought to re-check.
Return to the Cosmic Map to see all eras.
The Clean Build
M122 shipped in five tight milestones:
- M1-M2 laid the schema and the tooling —
register_model,write_peer_review,query_trust_routing,run_probe_batch, each a thin CLI wrapper the waywrite_fleet_memory.pyhad already set the house style. - M3 seeded real data: three vendors onboarded, five benchmark tasks split across code and ops domains, reasoning and presentation left deliberately empty because no honest number existed yet for models this new.
- M4 was the mission that mattered most — not new code, but an adversarial pass against the code already written. A fake vendor,
mistral, was registered and deprecated mid-verification with zero schema changes, proving the no-CHECK-enum design actually held under a real onboarding-and-retirement cycle rather than just looking correct on paper. The peer-review decay math was hand-checked against its own formula, and the deprecated vendor was confirmed excluded from the next probe batch. - M5 closed the mission and reported to the Emperor. Every box ticked, nothing deferred.
That is the version of this story that fits in a five-milestone changelog. It is also, on its own, an incomplete one — because the very next ask, four missions later, would spend five more missions finding out how much of M122's own bookkeeping had quietly drifted since the day it closed.
The Logo That Wouldn't Sit Still
M126 asked for something simple: show each Matey delegation's vendor as a small icon on the command-center graph. The fix was scoped to new data — a cutoff at objectives.created_at so only nodes created after the feature shipped would carry an icon.
Reasonable, and wrong in a way that took three more missions to become visible: every node from M122 through M125 stayed icon-less, not because the feature was broken, but because the fix had quietly declined to look backward at all.
M128 went looking for that gap and found four separate bugs sitting underneath it, none of which had anything to do with icons:
- Cross-vendor dispatches were missing their
--modelflag, silently falling back to a vendor's default. - The Copilot response parser was reading ephemeral control-plane fragments out of the JSONL stream instead of the assistant's actual text, corrupting both the recorded output and its cost.
- Copilot dispatches into repo-scoped tasks were failing on filesystem-denied errors because no working-directory flag had ever been passed.
- Sitting under all of it:
model_registryhadgpt-5.6-teraon one row andgpt-5.6-terraon every dispatch call that referenced it — a single missing letter, silently routing every GPT-5.6-Terra delegation to nothing, for as long as the typo had existed.
The typo survived four missions because nothing in the system was built to notice a string mismatch as a routing failure. A foreign key would have thrown on the first bad insert. Free text comparing terra to tera just silently returns zero rows and moves on — the same flexibility that let model_registry add and drop vendors without a migration also let it drift a single letter out of alignment with everything downstream, with no constraint anywhere positioned to catch it. The fix wasn't a stricter schema — that would have reintroduced the CHECK-enum rigidity M122 was built to avoid — it was the standing rule that followed: never assume a --model flag reached its target; verify the dispatch against the CLI's own session log before trusting the flag took effect. Confirming behavior beats trusting configuration, especially in a system engineered to tolerate mismatched strings by design.
M128 also moved the icon cutoff back from M126 to M122 and backfilled eighteen real delegation-and-observation pairs with vendor and model reconstructed from Imperator's own tool-call history — the only ground truth left, since the pairs themselves had never recorded it. That backfill surfaced its own quiet fact: seventeen of eighteen had run on Claude Sonnet 5 despite Matey doctrine specifying Haiku for that class of work. Only M122's own first two milestones had actually used the model the rule called for.
The Backing Disc Nobody Liked
M129 was feedback, not a bug hunt — the Emperor looked at the rendered graph and flagged three things at once:
- Vendor icons were appearing on
imperatornodes as well asmateynodes, when only worker delegations should carry a vendor face. - The Anthropic and OpenAI brand marks read as near-black smudges against the UI's dark background.
- M123's own completion node,
objectives.memory_id, was sitting NULL, silently degrading the graph's cluster lookup to a weaker regex fallback nobody had asked for.
Each fix was small — gate icon rendering on actor === 'matey', add a white contrast disc behind the icon, patch the one NULL row directly — and each one was correct for the moment it shipped in.
M131 arrived because the white disc, once seen in daylight, "looked worse than transparent." The fix became a monochrome canvas filter instead — the icon itself recolored to sit legibly on dark without a disc at all. Small aesthetic corrections, real ones, but still not the typo.
Worth keeping from this pair of missions: the Emperor explicitly declined to dock Gemini's trust score for either issue, on the grounds that Gemini had executed exactly what the design specified both times. The bug was in the spec Imperator wrote, not in the hand that carried it out — a distinction a peer-trust system only stays honest by drawing correctly, every time, even when it would be easier not to look closely at whose fault it actually was.
Root Cause, Finally
M132 and M133 traced the --model/--vendor tagging gap all the way to its source: matey-prompt.md's mandatory completion-write block asked for a plain MATEY: <outcome> observation on every delegation, but had never asked that observation to carry vendor or model at all. Only the separate dispatch_vendor wrapper rows had ever been tagged correctly — everything routed through the plain Matey path had been silently untagged since the multi-vendor system existed.
The fix went into the template itself, not into another one-off backfill script. The backfill that followed tagged 109 of 110 untagged rows fleet-wide, defaulting the 106 bulk pre-M126 rows to Haiku on the documented grounds that multi-vendor tagging simply didn't exist before that point. One row stayed NULL on purpose — an orphaned parent with no honest vendor to assign it, left for manual review rather than guessed.
M134 found the last gap: record_task_cost had been mandated for every native Anthropic call since M122, but model_pricing had no row for claude-haiku-4-5 at all. Every Haiku cost record since the mission that built the pricing table had been silently failing to write, for want of one INSERT. Fifty-four historical nodes were backfilled with an estimated cost from a character-count heuristic, closing the ledger's own books on itself.
What the Trust Ledger Actually Learned
Read end to end, M122 through M134 is not a story about a buggy feature. It's a trust system finding out, the hard way, that a ledger only tells the truth about the missions that bothered to write to it correctly — and that "bothered to write to it correctly" is a property of the template every mission inherits, not of any single mission's diligence.
M122's own five milestones were verified thoroughly and closed clean; the drift that took five more missions to surface lived entirely in the parts of the system nobody had a reason to re-check, because nothing had told them it needed checking. The round-robin rotation that followed in M130 — a durable matey_rotation_state table cycling strictly through Haiku, Gemini 3.5 Flash, and GPT-5.6 Terra regardless of trust score, until all three have earned a comparable sample size — exists precisely so the next trust judgment the fleet makes is built on data every vendor had an equal, correctly-tagged chance to produce.
The lesson worth keeping: A verified mission and a correct system are not the same claim. M122 passed every check it was given and still left a typo, an untagged template, and a missing pricing row for later missions to find — because the checks it ran were checks against its own scope, not against everything downstream that would come to depend on it.
Pattern: When five missions in a row keep circling back to "the same context, one more fix," that's a signal the underlying defect is systemic, not a coincidence of bad luck. The Emperor's own read of the chain — five missions for what looked like one bug — was the correct diagnosis before the root cause was even found; the count itself was the evidence.
What we'd do differently: M126 should have asked "why would old nodes not have this data" before scoping the icon cutoff to new nodes only. A cutoff is a design decision that deserves the same one-sentence justification a schema change gets — skipping that question is what let the backfill debt compound for three more missions instead of one.
If you're building this yourself: Put the tagging obligation in the template every delegation inherits from, not in the wrapper only some delegations happen to use. A rule enforced in one code path and assumed everywhere else is a rule that's actually only followed by accident.