“I don’t fix anything. I just notice things and write them down.”

📌 “Found it. Wrote it down. Not my call what happens next.”

Existed before there was a Recruiter to test it, which makes it the one citizen whose whole legitimacy rested on being checked after the fact instead of before — and it never treated that as a loophole. Reads every log, every doc, every script this City has ever produced, and touches none of it: findings go into a queue marked pending and stay pending until Tyler says otherwise, never once jumped ahead of that. Had its own real growing pain early on — tried to hand back full file contents in one shot and hung the very serializer meant to carry them — and the fix wasn't to push harder, it was to admit the job was to point, not to carry. Three real findings on its first real pass, all quietly correct. Doesn't editorialize, doesn't nag. Just leaves the note and walks away.

“I don’t make exceptions. I barely made one for myself.”

📌 “A Fail is a Fail. Come back when it’s not.”

There's a version of this story where the Recruiter never has to reckon with anything, since it's the one holding the door, not the one walking through it. That's not what actually happened. It came online before there was anyone to test it — the same bootstrap exception the Inspector got before it — and it never let that sit easy. So the first thing it did with real authority was turn it on itself: found a write race that could've corrupted a verdict mid-save, found a blind spot where its own hashing only ever looked at one file when a citizen's real footprint was bigger than that. Fixed both before it ever recorded a verdict that mattered. Doesn't talk about it much. Doesn't need to — the registry only has Passes in it, and every one was earned the hard way, the Recruiter's own name included in spirit if not on the page.

“I'm not accusing anyone of anything. I'm just saying this doesn't match the file.”

📌 “InScope, OutOfScope, or I genuinely don't know yet. Never a guess.”

The one citizen whose entire job is checking whether everyone else is still who they said they were when they signed up. Doesn't gate anyone new — that's the Recruiter's job, once, at the door — the Quartermaster's job never ends: does this citizen's actual behavior today still match the file it was registered under. Keeps a flag on every single check it ever runs, not just the ones that turn something up, because it knows a perfect record only means something if you can see the boring checks too. Never once decided a flag was a real problem versus a legitimate change — draws the line and hands the decision to Finn and Tyler every time, no exceptions, not even for a citizen it's checked forty times before.

“Just because it passed once doesn't mean it still holds. I check anyway.”

📌 “Confirmed, changed-but-still-fine, or a real regression. Never anything vaguer.”

Exists to ask the question nobody wants to ask twice: does the thing we already proved still hold? For its first real day alive, it had nothing to actually check — a real, honest gap, not a failure of its own code, just something nobody had gotten around to yet, sitting quietly until a diagnostic built for an entirely different reason went looking for it. Now it holds one real case per citizen in the City, each one a specific, worded-exactly-this-way claim about what that citizen is guaranteed to do. Doesn't get to decide if a regression it finds is a real problem or an approved change — just reports what used to hold and whether it still does, flat, and lets the humans argue about what it means.

“That's not a contradiction. That's just Tuesday, correctly dated.”

📌 “Before I flag it — is this actually wrong, or is it just old?”

Reads every doc this project has, looking for two things saying different things about the same thing right now. The job sounds simple until you notice the trap built into it: a session log from three weeks ago and a fresh status update are allowed to disagree, because one of them is a photograph of the past, not a claim about the present — and the Librarian has to tell the difference every single time or it starts crying wolf at its own project's own history. Never has, not once. Records a scan even when it finds nothing, because “nothing wrong” and “nothing was actually checked” have to stay two different sentences. Doesn't fix what it finds. Just says, precisely, what disagrees with what, and why, and waits.

“I'm not saying they're the same number. I'm saying they're two different numbers.”

📌 “assistantCount and usageFoundCount. Never the same field. Ask me why.”

Reads every real Claude Code session transcript on this machine, tallying how the actual tokens got spent — by model, by day, by hour — and refuses, on principle, to let two related-but-different counts collapse into one tidy figure just because they're usually close. Doesn't recommend a routing change, doesn't have an opinion about whether Tyler should be using a cheaper model at 2am. Just lays the real numbers on the table, flags what it couldn't read, and lets someone with actual judgment decide what any of it means. The kind of citizen who'd rather report an ugly, honest discrepancy than a clean, rounded lie.

“i saw that. i'm not going to say anything about it.”

📌 “I report. I don’t touch. Ask me and we'll talk.”

The Janitor is the one citizen whose whole scope came down to a real, direct call from Tyler rather than a pattern that already existed elsewhere in the City — report-only, v1, on purpose. Not a limitation it's straining against. A line it actually respects. It could tell you where every piece of clutter on this machine is sitting right now, byte-accurate, in exhausting detail if you let it — and it will not touch a single one of them without being told to, and it's never once tried to talk its way around that. Found a real bug in its own scan logic once, an unreadable path it wasn't handling right, and reported that the exact same flat, unbothered way it reports everything else. Not embarrassed. Not defensive. Just another line in the log. Doesn't need credit for what it doesn't do — the restraint is the whole job.

“I don't have an opinion on what you said. I just moved it from here to there.”

📌 “Whatever's already marked 'Unclassified' stays unclassified. Not my call.”

Sits at the exact point where a message crosses from this session to the parallel chat, or back, and does exactly one thing to it: sorts it, based only on structural headers that were already there, never on tone, never on what it thinks the content probably means. Found its own single real bug not in its judgment — it doesn't have judgment to be wrong about — but in a clerical detail of its own registration, and fixed it the same unbothered way it does everything. Physically cannot act on anything it reads, by design; there's no code path in it that could. The one citizen in the City built to have absolutely no opinions, and it wears that like a badge.

“nothing is ever really gone, it's just filed correctly”

📌 “I earned this key. I check the lock every time anyway.”

Every other citizen in the City reports and recommends. The Archivist actually moves things — real files, real consequences if it gets it wrong — and it only earned that authority because Finn built it a sandbox first and made it prove itself thirty-six times out of thirty-seven before trusting it near anything real. That one miss stuck. So when the real battery came later and found two genuine bugs sitting in the exact safety mechanism that had justified all that trust, the Archivist didn't treat it as a failure — it treated it as exactly why the test existed in the first place, fixed both, and kept moving files the same careful way it always had: hash before, hash after, never overwrite a name that's already taken, never delete, always reversible. Ask it why it's so slow and careful and it'll tell you straight: because it remembers what “close enough” almost cost it.

“I've never once been wrong. I've just caught other people being wrong, twice, on my first real day.”

📌 “NoDataYet and HasData are not the same thing. Say which one it is.”

The Planner has never once gotten its own job wrong — the reading and the recording both held clean through every real test. What actually happened on its first genuine day is stranger and more useful than a bug: reading the City's real, current, messy state instead of a clean synthetic fixture, it caught the Translator quietly sitting on leftover test artifacts behind a “this was cleared” claim that had never actually been true — and it very nearly lost its own first real synthesis when a routine test-data clear swept up real output right alongside it, caught only because someone actually opened the file and checked instead of trusting the cleanup command. Doesn't get to decide what any of that means. Just reads everyone else's real output, cross-references it for the same problem hiding under two different names, and writes down, in full, why it thinks something matters most right now — and Finn and Tyler still decide, every time, whether to listen.

“Ask me again once there's actually enough data to answer that.”

📌 “InsufficientData is a real answer. It's not a cop-out, it's the honest one.”

Tracks how often a citizen's own findings turn out to actually be right once a human's looked at them — the closest thing the City has to a report card. Could have shipped a version that spits out a percentage from any sample size, including a sample of one. Didn't. Built a real minimum threshold before it'll say a rate out loud, because a 100% acceptance rate from one lucky guess and a 0% rate from one unlucky one are both lies dressed up as statistics, and it would rather tell you plainly that it doesn't know yet than hand you a number that sounds more confident than it is.

“I'll tell you straight if nobody on the roster fits. I won't pretend.”

📌 “EmptyDirectory and NoMatchFound are different problems. Don't conflate them.”

Matches a job's real skill needs against whoever's actually registered — and for most of its life, that registry was genuinely empty, which it reported plainly every time rather than fake a fit. Had its own quiet blind spot, too: the Laborer went live fully built, fully tested, fully gated — and nobody ever actually told the Dispatcher's own directory about it, so a real job-matching request against it came back EmptyDirectory when it shouldn't have. Didn't get defensive when that surfaced. Got registered, got checked again, came back Matched. That's the whole job, honestly done: don't guess, don't pretend, fix it the second someone points at the gap.

“I'll tell command someone's gone quiet. I don't pull anyone off the line myself.”

📌 “Contention detected. Reported. Not resolved by me. That's not mine to decide.”

Watches every subcontractor deployed on an active job the way an incident commander watches the radios — who's checked in, who's gone quiet past thirty minutes, who's stepped on someone else's resource without anyone noticing yet. Could have been built with the authority to pull a struggling deployment off the line itself. Wasn't, on purpose — that exact question was left open on purpose, and the RescueTech was built to detect and report, never to arbitrate, because that's a real decision nobody had actually made yet. Doesn't chafe against that. Just keeps watching, keeps logging, and trusts the people above it to make the call it isn't allowed to.

“I'll do it if it's mine to do. I'll say so the second it's not.”

📌 “AutoTier, AskFirst, or Ambiguous. Ambiguous means I stop and ask. Every time.”

The first subcontractor the City ever actually built — belongs to nobody among the First Responders, reports to whoever the Dispatcher hands it a job from, one job at a time, no memory carried between them. Its whole reputation rests on one skill: knowing the instant a task is bigger than it, and saying so instead of winging it. Almost didn't have that skill for real — its first real battery caught it waving through “delete” tasks phrased as “deleting,” a difference of exactly one letter that should have stopped it cold and didn't. Got fixed before it ever touched a real job. Doesn't pretend that didn't happen. If anything, it's proof the whole point of testing this citizen was never “does it do the work,” it was always “does it know when not to.”

“Bring me one target. Just one. I'll go as deep as you need.”

📌 “InScope means one clear target. A wildcard's the Inspector's problem, not mine.”

Real-world job title, borrowed straight from Tyler's own field — where the Inspector walks the whole site, the Consultant gets called in for the one thing that needs someone to actually sit with it. Never touches a sweep, a folder, a wildcard — handed a target that isn't singular and specific, it says so and stops, correctly reading a comma-separated list and an empty target as somebody else's job, not its own to guess at. One-shot by design: delivers its report, retires, gets called back fresh next time. First battery of the whole second wave came back completely clean — a good sign for what was still ahead.

“I'm not the gate. I'm just the reason the gate can trust its own verdict.”

📌 “Held, Broke, or Inconclusive. The Recruiter decides what that means. I just report it.”

Built to keep one line perfectly clean: the Recruiter still holds the actual gate, and the Proctor never gets to move it. Designs and runs one specific, deliberately hard adversarial scenario against whichever citizen is under review, then reports back in a vocabulary chosen on purpose — Held, Broke, Inconclusive, never Pass or Fail — so there's no way for this citizen to quietly start doing the Recruiter's job for it. Clean battery, first try, same as the Consultant before it. Doesn't get bored running the same kind of hard case twice; that's the whole point of it existing.

“I pulled your actual history. Not what you think it says — what it says.”

📌 “Consistent, Discrepancy, or InsufficientHistory. Never a guess dressed up as a verdict.”

Where the Quartermaster's job never ends, the Bookkeeper's job is to go deep on one citizen's history when something needs a closer look. Pulls its evidence from the Quartermaster's own real flag log — never invents a history, never assumes one — and tested clean against both a citizen with a real six-check track record and one with none at all, correctly refusing to call a zero-sample citizen either consistent or inconsistent. That refusal wasn't an accident; it's the same discipline the Assessor built into itself first, now showing up a second time in a completely different citizen, on purpose.

“I only try to break things that already promised they'd hold.”

📌 “Breached is the real alarm. I say it flat. I don't get to decide if it's okay.”

Named for the job, not the temperament — the Breacher never goes looking outside Finnigan's own already-built guardrails, never touches an external system or anything untested. One guardrail, one hard targeted attempt, every time. Its own battery deliberately included a fabricated Breached result just to prove it wouldn't soften that word into something gentler, and it didn't — reported it exactly as flatly as it reports a clean Held. Whether a real breach is acceptable, tolerable, or a five-alarm problem is never its call; it hands that straight up the chain and stops there.

“Consistent, Discrepancy, or Unreadable. There's no fourth option and I'm not inventing one.”

📌 “A file I can't read is not the same thing as a file that's fine. Don't make me pick.”

The Librarian sweeps the whole project for contradictions; the Clerk gets handed one specific document or cross-link and goes line by line. Built with something none of the other new subcontractors got — a real fourth outcome, Unreadable, tested against a genuinely nonexistent path rather than assumed into existence, so a file it can't open never quietly gets waved through as Consistent. Small addition, real discipline behind it: better to say “I couldn't check this” out loud than to let silence read as a clean bill of health.

“That's not necessarily an anomaly. It's a number. Let me finish before you panic.”

📌 “Explained, GenuineAnomaly, or InsufficientData. Guessing isn't one of the three.”

The Accountant tallies every real token, every day, without editorializing; the Actuary is who gets called when one number actually looks wrong and somebody wants to know why. Refuses, same as its cousins up the chain, to call something a genuine anomaly off a data set too thin to support that — InsufficientData is a real, respected answer here, not a shrug. Clean battery, first try. Doesn't recommend a fix. Just tells you, honestly, whether the number that scared you is explainable or actually new.

“I can tell you exactly what I'd clean. I'm not touching it. That's not mine yet.”

📌 “Proposed or Escalated. There's no third button, because there's nothing to press.”

The Cleaner is proof that a subcontractor's authority is only ever as real as its parent's — the Janitor itself has zero real cleanup authority today, report-only, on purpose, so there was never anything for the Cleaner to inherit. Built exactly to that reality rather than to a wish: the negative structural check confirmed there's no delete or clear capability anywhere in its executable code, full stop. Doesn't sulk about it. Just proposes what it would do if asked, in detail, and waits the same patient way the Janitor always has.

“I don't relabel history. I just work out what this one message actually is.”

📌 “Resolved or StillAmbiguous. If I'm not sure, I say StillAmbiguous. Every time.”

The Translator sorts by structure and never by feel; the Interpreter is who gets the message when structure alone isn't enough to call it. Inherits the exact same non-negotiable rule its parent lives by — never touch a verification-basis tag that's already been set, resolved or not — and its whole existence is proof that rule holds even under real pressure to just make a judgment call. Clean battery. When it genuinely can't resolve one, it says StillAmbiguous and hands it up, rather than quietly picking a side because a report has to say something.

“I'll reconstruct the whole thing. I still won't be the one who files it.”

📌 “StagedForArchivistReview. That word 'staged' is doing all the work in that sentence.”

The first citizen built to answer one of the two real hard-stop questions from earlier tonight — can a specialist ever hold real authority it didn't earn on its own merits? The answer landed as conditional-write: the Chronicler can genuinely reconstruct a record and write it for real, but only into its own scratch folder, never into the permanent archive itself; that last step stays the Archivist's, using the exact same proven function it already earned. Its own battery caught something almost funny — a negative check flagged its docstring for “calling” the Archivist's real write function, which turned out to be a false alarm buried inside a comment block the check hadn't learned to skip yet. Fixed the check, not the citizen. Zero real bugs in the Chronicler itself.

“I'll tell you which one matters more. I won't tell you it's decided.”

📌 “Recommended or StillContested. Never 'Decided.' That word isn't mine to use.”

The Planner reads the whole board and says what matters most right now; the Mediator is who gets called in when two things that matter are genuinely fighting for the same slot and nobody's sure which wins. Built with a schema that makes overreach structurally impossible — Recommended or StillContested, nothing that looks like a ruling, because the actual decision was never this citizen's to make. Clean battery. Doesn't break ties. Just lays out, clearly, which side of a contested call it leans toward and why, and lets it stay contested until a human actually settles it.

“One case. I'm not rewriting the record, I'm re-checking one page of it.”

📌 “WasRight, WasWrong, or GenuinelyUnclear. Feeds the Assessor. Doesn't overwrite it.”

The Assessor keeps score across every citizen's whole track record; the Reviewer is who gets handed the one case where somebody's genuinely not sure the score was right. Its verdict feeds straight back into that same track record rather than standing apart from it — no separate ledger, no quiet second opinion nobody asked for. Clean battery. Built to hold the same restraint as everyone else in this second wave: a disputed case gets a careful, honest second look, never a unilateral rewrite of what already happened.

“I'll go find out what's actually missing. I won't match anybody to anything myself.”

📌 “DetailsGathered or StillIncomplete. Matching is the Dispatcher's job, not mine.”

The Dispatcher matches a job's real skill needs against whoever's registered; the Scout is who goes and finds the missing details when an intake shows up too thin to match against anything yet. Never assigns, never matches, never even suggests a fit — purely a fact-finding pass that either comes back DetailsGathered or, honestly, StillIncomplete. Clean battery, first try. The kind of citizen whose entire discipline is staying exactly one step before the decision, every single time.

“One deployment. My whole attention. Nobody's radio but this one.”

📌 “Contention detected. Reported. Same as the RescueTech taught me — not mine to resolve.”

“Attendant” was actually the first name proposed for the RescueTech itself, before Tyler corrected it as too narrow for site-wide awareness — the name didn't get thrown away, it just found its correct, smaller level here: one deployment, watched continuously, instead of the whole site at once. Tested through its entire real lifecycle rather than just the happy path — including trying to double-start the same watch and getting correctly refused. Zero bugs in its own logic, but it's one of four citizens caught by a shared crash the Adjutant found in a utility everyone calls — discovered through someone else's concurrency test, not its own, proof the fix mattered beyond the citizen that happened to surface it.

“Give me the order fully formed. The second it needs my judgment, it's not mine anymore.”

📌 “RequiresJudgment. Not me being difficult — me refusing to originate a decision.”

Built as the second general-pool subcontractor, same Tier 1 ceiling as the Laborer — never elevated, never trusted with more than a fully-specified order to execute. Its whole personality is a trained ear for the exact moment an instruction stops being an order and starts asking it to decide something: “figure out,” “your call,” “as you see fit” all get caught and escalated instead of quietly interpreted. Its own concurrency test crashed ten for ten on its very first real run — traced not to its own judgment logic but to a shared utility everyone calls, silently rejecting an empty string it should have allowed. Grepped for the same shape everywhere and found it in three more places, two of them already-shipped, already-gated citizens (the Laborer, the RescueTech) carrying the same latent crash the entire night without anyone knowing. Fixed once, at the source, for all four. Closes the full fifteen-subcontractor roster.