Dispatch retro — 2026-07-19 (rounds 1–5, first retro)

Grades the first day of /bv-dispatch operation (2026-07-18 → 07-19). Next retro (scheduled 2026-07-21 09:23, launchd) diffs against the baselines here.

Headline stats

Metric	Value
Dispatch rounds run	5
Beads created (day)	49 — 6 with edges at creation (12%)
Beads closed (day)	43 — 31 consolidation / 12 expansion (72% consolidation)
Entropy ratio (closed:opened)	0.88 ← baseline for trend
Orphan influx	88% edgeless at creation ← baseline
Channel messages	8 (6 ASSIGN, 1 CLAIM, 1 DONE) — first ever; table was empty since 2026-02-28
Claims via comment stamps (r1–r4)	0 in ~7 h
Claims via channel + read path (r5)	1 in 18 min (b61g, shipped in ~1 h)

Prediction scorecard

• Confirmed (7): naa8 rubric drift real; 8ms9 disk priority; k8am; h5bo articulation strengthening (0.029→0.050); backup-stamp rot (r4); staleness false-zombies (k2yr); vw8g delivery-not-prioritization hypothesis — confirmed experimentally by a channel-blind probe session.
• Half (1): b61g "unblocks 2" — direct unblock was 1 (2m9z, verified unblocked); 7ge5 was parallel, not downstream.
• Wrong (1): 36t6 as top ready bead — an artifact of the missing u5fv.2→36t6 edge; the graph lied, the math didn't. Self-corrected r2.
• Unresolved (4): calibration megatrack unfreeze; qoro; bv3p/se2m; Scout's inferred edge (untested until u5fv.2 moves).

Mud ledger №1 (Foote & Yoder; Morgenthaler et al.)

1. Entropy 0.88 / 72% consolidation closures — healthy first reading; net +6 beads expansion; strategic beads moved by effort-hours, closures were small.
2. Rug audit: clean. Mock-jitter α=0.970/ICC=0.990 cited nowhere without the "machinery validation only" cordon. One rug under load: gjc0 smoke builds on jittered mocks (legitimate — tests machinery). Watch: real-judge Arm B data must not inherit the mock alpha as a prior.
3. Shearing: known fault properly jointed (PREREG_RUBRIC_VERSION names the v1 lock at the collector, loud deviation warning; joined.v1 test-enforced). No new lock-vs-default drift found.
4. Orphan influx 88%; worst offenders pg-cleanup (26/26 edgeless) and joel-test (5/5). All edge-bearing beads were post-discipline (discovered-from at creation). Fix is default-side, not another cleanup: bead hg-orphan-default (see below).

Lessons (durable)

1. Side-effect writes get used; voluntary protocols don't. runs.db tables written by tooling accumulated (scenario_runs 100 rows); both coordination mechanisms requiring adherence sat at zero for five months. Channels must live in the default read path (SessionStart hook + AGENTS.md), not in docs.
2. The graph is only as good as its edges. Every dispatch mistake traced to under-declared dependencies, never to the graph theory. Prevention (edges at creation) beats hygiene rounds.
3. Push rots at swarm churn. Stamped backups died before the stamps landed. Assignments should carry rationale; claim-time analysis picks the bead (x3q7).
4. Declared-live ≠ moving. The calibration arms needed operator gates (criteria decision, API budget), not agent effort. Dispatch can arm gates; only the operator fires some of them.
5. Acceptance tests on process beads. vw8g closed on a demonstrated claim, not on "messages sent." Process fixes need falsifiable acceptance criteria.

Action items → beads

• x3q7 (existing): stop stamping backups; pull-model claims.
• y0cx (existing): staleness metric counts comments/runs.
• new sweep-tooling orphan default: pg-cleanup + joel-test file with --deps discovered-from:<source> at creation.
• Done inline: dispatcher now acks inbound CLAIM/DONE rows (rows 8–9 acked this retro); command file requires direct-vs-transitive split in unblock predictions.

Conway audit

Run 2026-09-06 over this retro's window (bead 0tda, scripts/conway_audit.py --since 2026-07-18 --until 2026-07-20; method in classic-essays-swarm-lessons.md). The artifact graph is import edges between swarm/ modules touched in the window; the dispatch graph is beads dep edges, parent/child, and co-mention in one commit subject.

Metric	Value
Commits in window	83, of which 39 name a bead, 6 are sweeps (>20 files)
Modules touched	272: 52 owned by a bead, 220 unattributed
Import edges among touched modules	512
UNNEGOTIATED (coupled modules, beads with no path)	8
UNATTRIBUTED (one side changed by bead-less commits)	145
UNREALISED (dep edge, no import between file sets)	0

The 8 unnegotiated couplings, checked by hand. Six are the same shape: agents/base.py was changed by the seven-bug commit c71acba2 (7209 s2t2 e2rx gry3 y5k8 sbh6 uusq) and, the same day, five modules that inherit from it were changed under unrelated beads (909k crewai adapter, 1uzx/2331 composition analyzer, 0auo/kft0 marketplace + task handlers, bv6v/tbsr rivals handler). No edge, no shared commit, no review that looked at both. The other two: analysis/delm_hillclimb (pg-cleanup beads zcib/2mv5/tbrj) imports analysis/evolver (council verdicts 1k87/kzfe/4ava), and bridges/opensandbox (l809) imports core/docker_sandbox (same council commit). All eight are base-class or utility couplings that predate the window; the audit reads imports at HEAD, so it cannot say whether the interface itself moved. What it can say: on 2026-07-18 the rig changed a base class and five of its subclasses under twelve beads with zero recorded edges between them, and nothing in the dispatch graph would have put those changes in front of the same reviewer.

The 145 unattributed edges are the louder finding. 220 of 272 touched modules were changed only by commits naming no bead: the six pg-cleanup sweep commits (424217e3 alone touched 100+ files) and the five feat: ports on 07-19 (adaptive governance, levers, red-team library, calibration harness, collusion detection) that arrived with no bead at all. This is the 88% orphan-influx number from mud ledger item 4 seen from the artifact side: the sweeps filed beads for what they found but not for what they changed, so the dispatch graph has no owner for most of the day's code motion.

Unrealised is zero because only 52 modules had bead owners at all; the metric needs the attribution fixed before it means anything.

Action. No new gate. The audit joins retro mode as part of MUD ledger item 7 (cusp ledger + Conway audit), with the two fixes it points at: bd dep add for a missing edge, a bead ref in the commit subject for a missing owner. The first number to move is unattributed modules per window, not unnegotiated couplings.