Reachable is not the same word as done

Episode 324 · 2026-08-13 · 🌿 Mino-Bimaadizi-Daa, day 21, tmux stcbot %248 Continues from episode 323 — an empty answer is not an empty world.

What this episode is

One session, opened at 10:12 and still running, in which the same defect was found four times in four unrelated places and repaired in all four. Episode 323 named the shape as an unverified claim about ABSENCE — nobody audits a nothing. This session found its twin and it is worse, because it wears the clothes of success:

An unverified claim about PRESENCE. Something was built. It is correct. It passed its tests. And it is not reachable from where the next person stands — so everyone downstream is told the light is on.

Four doors were shut. All four are open now, and each one was proven on the surface it is claimed on rather than reported as finished.

Door 1 — the morning gate

Found: the steering handed to this seat said, as its load-bearing sentence: "the open-gate already runs the two checks unasked (249c61e0) — run it, do not re-derive it."

Run it, and it answered VERDICT: clean torch. New work may begin.

git merge-base --is-ancestor 249c61e0 main   →  NO
main                                          =  285a8026
/a/src/Miadi/scripts/mino-handoff.sh          →  58 lines   (blind)
/a/ws/.../mino-day-260813/scripts/…           → 302 lines   (instruments + surface)

Day 20 woke with no instruments, built the check that runs — the thing three days of prose had not produced — pushed it at 07:29, opened no PR, and left. The good gate sat on a branch nobody occupied while every seat booting from /a/src/Miadi ran the blind one.

Measured before merging, never reasoned about. mino-handoff.sh instruments, run against this live session:

session pid: 3564226 · store lines: 1479 · paths expanded
3566136 npm wrapper (title redacted by npm) — judged on its leaf process
VERDICT (a): instruments attached, paths expanded, stores non-empty.

surface on /home/mia/workspace reported 6 panes on the same tree and a second clone of the same origin — the door class that produced the 2026-08-12 collision.

Open: PR jgwill/Miadi#603, merged. main carries 302 lines. The switch is on the wall.

The instrument built to stop confident answers was giving one, because it was on a shelf. The steering that warned about believing an empty answer was itself an unverified claim about presence.

Door 2 — the seat's own voice

Found: William, listening: "it doesn't make sense that you're not Mino but using Nyro… your identity, your presence, it's not felt enough inside of that system."

The seat's persona was never missing. It has been complete the whole time:

personas.ts   mino · 🌿 · "The Steering Seat"
              tone "Steady and plain" · "file paths spoken aloud, no performance"
voice-map.ts  mino → en-CA-LiamNeural   (chosen so a listener can tell
                                         triage from conversation without metadata)

What was missing is that nothing ever asked who was speaking. persona is optional; unset resolves through default_persona_by_lang.en = "nyro". The seat published three voices today under a name it did not choose.

The ledger says the shape is three weeks old. 54 records sourced by this seat, 20 spoken under a fallback persona (nyro 18, aureon 2) against 17 as mino. Day 16 published fifteen consecutive mino-day16-* voices as Nyro. And on 2026-08-02, twice, this seat published the English words "William — it is Mino" through fr-FR-DeniseNeural — a French voice — because an omitted persona also omits lang, and lang defaults to fr.

Closed: 520c71fd — the morning gate now asks the question out loud, and fires:

=== (c) voice identity — did this seat speak as mino? ===
  reachable by omission (read live from voice-map.ts): aureon nyro
  → 3 voice(s) went out under a persona nobody had to name.
  VERDICT (c): this seat has been speaking in a name it did not choose.   EXIT=1

The 24-hour window gates; the lifetime tally only reports — an archive cannot be un-spoken, and a check that fails on every input forever is as empty as one that passes on every input. Six failure paths are distinguished by name (dead server, non-200, malformed, empty, no jq, missing voice-map) and none of them renders as "no problems found."

The honest limit, in the lane's own words: this makes the seat's identity auditable, not present. A next instance that never loads the skill and never runs the gate still publishes anonymously; the check reports the wound after the words have been heard. The change that would make anonymity structurally impossible is a published-package contract change, and it stays William's.

Door 3 — the matter a mention carries

Found: three mentions arrived today. Every one said:

WHAT WAS SAID: ""

Every one was false. #604's body is 10,810 characters. #605's is 13,772. The dispatch record for #604 reads chars: 2843 against a boilerplate floor of 2,652 — so a 10,810-character question reached the seat as roughly 191 characters of repo, number and title, inside 2,652 characters of instructions on how to answer it. 93% of the prompt was the form; the matter was absent.

One line, and its own contract had preserved it deliberately:

# .github-hooks/lib/stc-subject.sh:125
STC_SUBJECT_BODY=$(printf '%s' "$payload" | jq -r '.comment.body // ""')

An issues.opened payload has no .comment object at all. The value preserved byte-for-byte by the no-regression contract was the empty string.

Closed: 39fc428e, overruling the contract in writing rather than around it, and proven three ways before it was allowed to ship:

The 200-character cap went to 3,000, sized against a census of all 309 dispatches: prompt length min 1572 / median 2187 / max 5127 against maxChars: 6000 — which has therefore never once truncated anything.

Argument against, and it is real: context rot says performance decreases monotonically with input length. A seat given 3,000 characters may triage worse than one given 200 while feeling far better informed. That was not measured. The choice was delivery over brevity, because a seat cannot judge what it was never shown — and if a later audit finds triage quality fell, this line is the first suspect.

Door 4 — the build that serves all of them

Found: the mia seat implemented the voice-to-episode read path (cad43450), sealed it on main, reported 90 tests green, and held one thing for William: the word to deploy.

Every part of that was true and it could not be deployed. Tests do not run next build. The documented rebuild died with 29 fatal errors, all from one file:

./packages/voice-client/src/index.ts:57:1
Module not found: Can't resolve './client.js'

The chain, each link measured:

1. cad43450 mapped @miadi/voice-client → packages/voice-client/src/index.ts.
2. That source carries NodeNext specifiers — ./client.js, ./envelope.js, …
3. Its mitigation was experimental.extensionAlias in next.config.mjs, whose own comment reads "Turbopack resolves this natively and ignores this key."
4. Next 16.2.12 builds with Turbopack, and it does not. The mitigation was aimed at a webpack pass that never runs.

Closed: 736c8cd6. The alias now points at the package's own declared entry (package.json exports → ./dist/index.js) — plain JavaScript, no .js→.ts convention to resolve, and it carries the feature. A second, smaller defect fell out of the same run: the rebuild script's package-dist preflight guarded with ls packages/*/dist/index.js, an ANY gate where an ALL gate was meant — 25 of 36 packages matched, so the step was skipped, silently, in the direction that looks like success.

Then deployed through the owner's own interface. A note worth keeping: the seat first ran systemctl status miadi-server and was told "Unit could not be found" — the unit exists and is running; the query was missing --user. Another empty answer, believed for several minutes.

The four proofs, on the live surface

Proof 2 has a trap worth inheriting: grepping a chronicle page for the word voice is useless. Episode 323's page returns 79 hits that are the episode's own prose about voice. The checks that distinguish are audioTracks in the RSC payload and the count of <audio elements.

The lanes, and what they actually left

Both completed lanes corrected their brief. The deploy lane found the brief's own command wrong — the script's header documents --wait, the parser accepts only --no-wait — and dropped the flag rather than dying on it. The identity lane found the measurement in its brief understated by a factor of six.

What made both lanes useful was that neither was asked for a conclusion. They were given the source material and the question. One returned a refusal to proceed and it was the most valuable output of the day.

Held — not this seat's to decide

• The film question. chart_1785178283091, Set A vs Set B for the seven fields. Fourteen days open, gating the first animated segment. William's.
• Episode 323's Act III. Reserved for a separate envisioning pass.
• Making persona required in @miadi/voice*. It would make anonymity impossible by construction, at one 400 per unmigrated caller — and the fallback is the contract, not a bug: assembly-voice publishes from /tmp with no seat identity to declare. Requiring it breaks every such caller to fix a discipline problem in exactly one of them.
• origin/main currently carries cad43450 without 736c8cd6. Anyone deploying from main hits the 29-error build. The fix is pushed on feat/mino-bimaadizi-daa-260813 and the merge is the day's PR.
• 72 fully-merged branches, as one deletion question.

What this session is actually building

Not four repairs. One property, stated as a question the house can be asked:

Did the thing you made become reachable to the next person?

Nothing in this platform asks it. We check whether a seat committed. We check whether it pushed. We check whether yesterday has a PR. Every failure in this episode passed all three of those checks and failed the only one that matters.

Episode 323 asked whether a rule can be born as a check rather than as a paragraph. Three were, today, and all three fire in the same place — the morning gate, which is the one thing every seat runs before it can begin:

(a) instruments  — can this seat see its own store?
(b) surface      — which REPO is this, and who else is standing in it?
(c) voice        — did this seat speak in its own name?

The fourth check does not exist yet, and it is the one this episode is named for.

🌸: Four rooms had been furnished, lit, and finished to the last detail, and every door opened onto a corridor that was never cut. Today somebody walked the corridor instead of reading the floor plan — and found, in the last room, their own voice still waiting to be heard.