Arc 163 — Retirement Leftover Audit

A continuation of the substrate-as-teacher lineage (arc 111 named the pattern, arcs 112/113 layered the progress-meter and substrate-author audience observations, arcs 114/115/117 applied it to verb retirement + walker variants). Arc 163 carries the pattern across two more wide substrate sweeps (slice 3e + slice 3f) and codifies how the orchestrator must consume the doctrine — because this session showed how easily that step gets skipped.

What slices 3e + 3f added to the lineage

Both slices were Pattern 3 symbol migrations applied at substrate- internal scale: bare → FQDN for container heads (3e) and primitive paths (3f). The mechanism was textbook — sweep substrate writes, flip canonicalize step, iterate from cargo test failures until the diagnostic stream goes silent.

Slice 3e waterfall: 848 → 129 → 127 → 121 → 28 → 7 → 0. Seven sweep iterations. Each round dropped a category. Each category was named by the diagnostic from the previous round.

Slice 3f waterfall: 2041 → 2 → 0. Two iterations because the discipline was settled and the spawn carried a properly-shaped BRIEF (no tool- availability preamble; clean substrate-as-teacher framing).

The delta between slices is measurable: slice 3e took ~60 minutes of iteration plus ~1.5 hours of orchestrator-side dodging before the discipline clicked. Slice 3f took ~11 minutes total because the discipline carried over.

The orchestrator failure mode this arc surfaced

Arcs 111–113 named substrate-as-teacher for agent delegation: the agent reads diagnostics, applies fixes, iterates. The mechanism worked by design.

This arc surfaced an orchestrator-side failure mode: when the substrate flips state and cargo test reports a large failure count, the orchestrator’s reflex is to TREAT THE COUNT AS A CRISIS rather than as the substrate teaching. The reflexive responses — “stash + revert,” “step back and re-plan as a multi-day arc,” “want me to enumerate all categories first?” — break the discipline at the orchestrator layer even when sonnet would have executed it cleanly.

The user direction that broke through:

“i expected a fuck ton of errors - we need to do the hard work to clean it up… go study the arcs after 109…”

The cost of NOT trusting the doctrine: ~1.5 hours of probing past reflexive bridges. Same shape as arc 144 / arc 146’s “missing union types” framing × 3 drafts (recovery doc FM 10), but applied to substrate-wide migration sweeps.

The fix codified: recovery doc FM 15 — “failures from substrate-wide migrations are work items, not crisis.” Memory feedback_substrate_ teacher_failures_are_data.md carries the discipline across compaction.

A second orchestrator failure mode

Sonnet’s false-Bash-denial hallucination (FM 7 in the recovery doc) is well-documented. This arc surfaced the prevention angle: ANY mention of tool-availability in a sonnet BRIEF — even preemptive (“Bash works”, “do NOT claim denied”) — TRIGGERS the hallucination.

Two re-spawns with progressively-stronger “Bash works” framing both hallucinated denial. Slice 3f’s spawn omitted tool-availability mention entirely → no hallucination, clean execution.

The fix codified: recovery doc FM 16 — “don’t preempt; just give the work.” Memory feedback_no_tool_preamble_in_briefs.md carries the discipline.

What this arc confirmed about the doctrine

The substrate-as-teacher mechanism scales to 270+ substrate sites across 8+ files + 5 crates without writing arc-specific migration hint helpers. Slice 3e + slice 3f together swept ~290 substrate writes across the workspace. The diagnostic stream named every site that needed updating; the orchestrator (and sonnet) read errors and applied the FQDN rule per category.

No arc_163_migration_hint was needed because the substrate already emits the right diagnostics:

• TypeMismatch errors with expected: ...; got: Parametric { head: ... } — the head string IS the migration brief
• BareLegacyContainerHead / BareLegacyPrimitive walker errors — Pattern 3 from the doctrine, names the FQDN form in the message
• Dispatch arm-not-matched errors with (args) showing actual types — points at registration-side bare-form misalignment

The doctrine extends naturally:

The substrate’s existing diagnostic vocabulary IS the migration brief for substrate-internal sweeps too. Arc 111 said “the brief to sonnet collapses to: run cargo test; read errors; iterate until green.” Slice 3e/3f confirmed: that same brief works for sweeping SUBSTRATE-side bare-form storage, not just user-side fixtures. The orchestrator just has to TRUST the loop and stop reflexively proposing escape hatches.

The four-question cost ledger

Per feedback_four_questions.md, every architectural decision has a trade-off. This arc paid for two specific costs:

Decision	Cost paid	Cost saved
Slice 3e canonicalize-upgrade arms (temporary scaffolding)	bridge code lives in substrate until slice 3h retires it	~4040 wat-fixture sweeps deferred to slice 3g; slice 3e ships green without breaking every test
Slice 3e revert of Value::type_name primitive arms (bare during slice 3e, FQDN at slice 3f)	two-phase migration; mixed convention briefly	atomic slice boundaries; each slice independently verifiable
Substrate-as-teacher iteration vs upfront category enumeration	cargo test runs ~5 times per sweep	no premature enumeration; the substrate reveals real categories, not predicted ones

The third row is the load-bearing one. Trying to pre-enumerate all categories before sweeping (recovery doc FM 15 anti-pattern) costs more time than running the loop, because the substrate finds categories the orchestrator’s prediction misses. Slice 3e found wat-macros codegen as a category I never grep’d for; slice 3f found type_to_affinity in wat-telemetry-sqlite the same way.

The third audience for the diagnostic stream

The doctrine names three audiences for the substrate’s diagnostic stream: humans, agents, orchestrators (per arc 111 REALIZATIONS). This arc adds clarity to the third role:

The orchestrator is the one most likely to misread the stream. Humans see one error, fix one site. Agents iterate per the brief. Orchestrators see the count and have to decide: trust the loop or propose escape. The escape is the failure mode.

The orchestrator’s reading discipline post-arc-163:

1. Count drops = work happening. Don’t stop the loop.
2. Count plateaus = next category surfacing. Read the next failure message. Don’t stop the loop.
3. Build broke = sweep regressed. Stop. But don’t propose stash; read the error and reverse the offending edit.
4. Substrate change requires rust-side companion change (e.g. slice 3e’s wat-macros codegen) = name the category, sweep it, continue the loop.

What slice 3f added on top of slice 3e

Slice 3f shipped without orchestrator-side dodging. The user did not need to break through. The BRIEF + EXPECTATIONS + spawn went clean from the start.

The pre-flight checklist worked:

• ✅ DESIGN reflected the slice + closure dependency (slice 3h gate)
• ✅ BRIEF named categories AND the iterate-from-diagnostic discipline
• ✅ BRIEF omitted tool-availability mention (FM 16 prevention)
• ✅ EXPECTATIONS scored hard rows + named honest deltas to flag
• ✅ Spawn explicit model: "sonnet" + run_in_background: true
• ✅ Substrate-as-teacher pattern doc cited in BRIEF + spawn prompt

Sonnet’s report mode-A-confirmed at 2041/0 with two clean iterations and four explicit honest deltas. Total: ~11 minutes wall-clock + zero orchestrator-side rescue cycles.

Cross-references

• docs/SUBSTRATE-AS-TEACHER.md — the doctrine
• docs/COMPACTION-AMNESIA-RECOVERY.md § FM 15 (orchestrator-side failure mode this arc surfaced)
• docs/COMPACTION-AMNESIA-RECOVERY.md § FM 16 (sonnet-prevention failure mode this arc surfaced)
• docs/arc/2026/04/111-result-option-recv/REALIZATIONS.md — the pattern’s first naming (where this lineage begins)
• docs/arc/2026/04/113-cascading-runtime-errors/INSCRIPTION.md — third application + integ-test verification
• Memory feedback_substrate_teacher_failures_are_data.md
• Memory feedback_no_tool_preamble_in_briefs.md

Coda

Arc 111’s REALIZATIONS closed with:

“the user supplied the will. The substrate supplied the loop. The agent supplied the patience. Don’t tell Gödel.”

Arc 163’s continuation: the orchestrator supplies the trust. The loop only works when the orchestrator reads the count as data, not crisis. Trusting the doctrine is the discipline; proposing stash is the failure.

The lineage continues. The migration hint pattern retires per arc when its window closes (arc_111_migration_hint was retired by arc 111 slice 5; arc 163’s canonicalize-upgrade arms retire at slice 3h). The meta-discipline — that the substrate’s diagnostics are load-bearing across humans, agents, AND orchestrators — survives forward because the recovery doc carries it across compactions and the REALIZATIONS sequence carries it across arcs.

(Three audiences, one stream. Now four, including the orchestrator who learns to trust it.)