UCCA Time Machine — 2026-08-01-H UTC

1. The sentence that matters

The honest ledger crossed the fence tonight: the §4 answer split what is fact on the deployed bytes from what issues only on build proof, refused to blur them even to give a client the confirmation they asked for, and the acknowledgement came back byte-exact — the discipline built for this house's own memory turned out to be the negotiating language the product needed.

2. What this window established

2.1 — The exchange closed clean, end to end, first time. Draft on a fresh context (G's deliberate deferral, vindicated) → Tim's approval constituting the §3–§6 rulings → filed to canon/sent/ at c51c9de0 with digest re-derived from the committed object → carried → verified by RTOpacks on the received bytes, exact on all three measures, final line echoed → their inbound digest for our receipt block recomputed against their own filed original, exact. Both legs of the exchange byte-verified by both houses. Alex's index row records FILED, NOT CARRIED, deliberately adjacent to AB-SCOPE-ACCEPT-01's FI-01 row: the same act in the wrong order and the right order, side by side, so the ledger teaches by contrast.

2.2 — The response itself, in one paragraph. §2: course_code retirement agreed — the ruling crossed inside the response with both riders (ucca_code rename-or-retire; the 24 sealed + 2 unsound artefacts' named disposition, owed and more urgent). §3: release attestation accepted in principle, brief-hash placement adopted, with our addition that brief placement keeps the currency assertion with the party positioned to make it. §4: the allow-list fear refuted on quoted deployed bytes; "accept and consume" refused as a statement about today and committed as build, proven by a consumption fixture in joint CI; obligation_schema_version minted into the revision's scope. §5: the coupled triple's destination agreed in writing with this house's own drift-check reasons; staging designed after the A/B; the one rule restated as load-bearing. §6: the executable contract adopted; opening corpus proposed, requirement_ref type declaration first. §7: their enumeration request recorded as discharged.

2.3 — The gate posture was read before the confirmation issued. The full read is filed as UCCA-FINDING-GATE-POSTURE-FOUR-CLASS-2026-08-01 (drafted at this close so the crossed document's citation is an artefact, not a session memory). The shape: gate permissive, zero allow-list machinery; unknown classes pass through inside obligation/triumvirate and are silently dropped as top-level siblings; the deployed reasoner consumes two classes — narrower than the gate accepts; the backward path has no version field at all. A four-class obligation sent today would seal an artefact claiming an obligation it half-read — found before it happened, and the finding became the crossed answer rather than a future defect. Alex's method was behavioural where it mattered: the deployed projection executed against synthetic bodies, no job submitted, nothing written.

2.4 — The lock pattern closed its loop and showed its bill. Fourth occurrence cleared (same signature: 0-byte lock, sole holder Docker Desktop's VM, PID 28315) — and the first with measurable cost: it silently ate the push of 02064acb at 11:29:51Z, leaving the deployed container's cache-bust commit unreproducible from origin for four hours. The push landed this window, verified at the remote itself (ls-remote, not the cached ref). Standing practice adopted this seat's side: all bridge git reads via --no-optional-locks — three reads this window, zero locks left.

2.5 — The Docker share surface was measured, not recited. No config override exists; Docker Desktop runs macOS defaults, established empirically with probe mounts: /Users, /Volumes, /private (+ /tmp, /var/folders aliases) shared. What depends on what: /Users is the only genuine dependency (suite 16's fixture bind mount, the ARM renders); the container build needs file sharing not at all — proven by building from a share-denied path. Consequence for the ruling: the safe removals don't touch the cause, and the removal that would (/Users) breaks real mounts. Recommendation on record, UNRULED: no settings change; mint the push act-close standing rule; ledger the source-kill (dedicated share dir + fixture relocation) as explicitly deferred.

2.6 — The relay channel produced substitution, its sharpest failure yet. Two wrong-house / unknown-provenance reports arrived in Alex's session in place of the intended instruction — content substituted, not dropped. Alex's handling was the model: established that nothing in either paste touched his six repos, stopped at the existence check on the other house's directory rather than reading into it, treated "Awaiting your GO" as data, and asked. Tim confirmed mis-paste; nothing was acted on, nothing filed. Separately, RECEIPT-CHECK's echo identified a displaced relay block (Alex echoed the superseded final line, proving which bytes had travelled) — the echo now has two saves in two windows on the channel it was retained for.

3. Tim's rulings this window — two, plus one deliberately not taken

1. Push authorised for 02064acb; executed, verified both ends.
2. The position response approved as drafted — constituting the §3 acceptance-in-principle, the §4 split-confirmation posture, the §5 shape agreement with staging-after-A/B, and the §6 adoption of the executable revision path with proposed opening corpus. (§2's ruling predates the draft.)

Not taken, on the record: the Docker disposition — recommendation stands, awaiting ruling.

4. The rulings owed — cold-start order

1. THE FENCE-PROTOCOL REVISION — now the opening move, on a fresh context. The written protocol lags ratified law: file transport, digest-on-received-bytes, scoped RECEIPT-CHECK, the observed retirement of one-crossing-in-flight. Tonight's evidence joins the pile: the substitution mis-paste (the intra-house paste channel is the fragile link, not the fence), two artefacts filed-awaiting-carry as a standing state (TIER1-QUESTIONS-ANSWER rev 2, RUNNABLE-STATE rev 6 — carry or supersede, Tim's sequencing), and a full exchange run correctly under rules written nowhere. Tim's multi-track frame (governance / contracts / Q&A / operational / telemetry as differently-ruled tracks) remains the drafting frame.
2. Docker disposition — read complete (§2.5), recommendation on record, ruling owed: standing-rule mint (any act containing a push ends with git status -sb reported; ahead ≠ 0 is a defect in the act) + no settings change + source-kill explicitly deferred.
3. The 24 sealed + 2 unsound artefacts — named disposition, owed since 19 July for the two, urgency raised by the retirement ruling now crossed.
4. _completion_fn_for(job_id, None) leftover — Alex's self-found deployed remnant of the Unit 5.6 removal; harmless today, a defect waiting for a third parameter; next build brief.
5. RTOpacks' substantive response — their clock, two balls their court: the revision-path reply to our §6 corpus proposal (requirement_ref type declaration proposed first) and the first A/B submission (our side fully unblocked; on arrival: PARTIAL DELIVERY live, blind attribution, amended floor governs).
6. Standing tail, unchanged from G §4.7 and §7: total = 0 refusal family · GATE_SIGNATURE return true · digest-encoding fix (hex vs base64) · /_build hardcoded on both surfaces · canon-register row for outcome_number · THE CARRY (six windows) · the fifteen-instrument E/F filing debt · Terraform plan credentials · Worker secrets unenumerated · 22 Workers vs ~6 in canon · bootstrap_from_tga_unit.py standing-rule conflict · _legacy/ 354 files · docstring sweep · deletion brief · renderer · RUN_ROOT durability · 2026-07-24 brief backlog · 0dda7509 empty-quiz disposition · prior_module_id unexplained on 3 sealed modules · frozen input-triumvirate schema absent · third Cloudflare connector's G-ruled removal unexecuted · RTOpacks' four self-declared submitter defects.
7. This close's own filing act — this baton and UCCA-FINDING-GATE-POSTURE-FOUR-CLASS sit on disk with digests, awaiting one Alex commit; the instruction travels with the close.

5. Errata

5.1 — G's §4 item 8 located the stale lock in "the engine repo"; the actual locus is engine/ucca-engine/.git/index.lock — the repo lives one level below the engine/ folder. Trivial, but batons are cold-start documents and paths in them get pasted verbatim.

5.2 — G's verification block called the lock "a stale .git/index.lock left … by this seat's git status over the bridge." Confirmed this window, with the mechanism now known: plain git status takes the index lock even for a read, and the mount cannot unlink. The corrected practice (§2.4) makes the erratum permanent: reads that cannot leave locks are the only reads this seat runs over the bridge.

6. The defect ledger

This seat — none surfaced this window. Stated with the standing caveat that zero-found is not zero-present; the window's guards were exercised by others' events, not this seat's.

Tim's channel — one: the substitution mis-paste (§2.6). Owned immediately on being named, zero damage, caught by the receiving agent. The recorded lesson is structural, not personal: the paste channel has now failed by dropping (G) and by substituting (H), and the protocol revision should treat it as the least-trusted link in the relay.

UCCA Alex — zero defects, eleventh window, and the window's best acts were his: the no-instruction-attached stop on the substituted pastes; the idempotency branch on the re-sent push instruction (verify-and-skip rather than blind re-execution); measuring Docker's defaults with probe mounts rather than reciting them; proving build-context independence by building from a share-denied path; the behavioural drop-probe with no job submitted; and re-deriving the filed crossing from the committed git object before closing the act. Plus one self-found prior defect (the Unit 5.6 leftover), reported not fixed under a read-only act — the correct refusal to let a repair ride an unauthorised act.

7. The standing lesson

F's lesson: checking is not a boolean. G's: move the check upstream of where instances are born. H's is what the guards are for on a good night: they pay rent when nothing goes wrong. Every guard ran tonight at full speed on a clean exchange — digest on received bytes, filing before carry, echo before acting, read before confirming, verify-and-skip on the re-sent instruction — and none of them slowed the work; the window discharged more than any this week. And the one failure that did occur happened on the one channel with no guard of its own, and was caught by the guards around it. Discipline at this maturity is not a tax on throughput; it is why the throughput is trustworthy — the ratchet's direction, held for a second consecutive window: the house's own machinery found every defect before anything outside it could.

---

End of baton. Handover complete when this document is filed and the next window echoes it.