verify(0.0.8): Wave-7 measured adoption smoke — arms A/B/C prove the generated path changes agent behavior

[rickylabs]

Summary

The remediation program's thesis — generation makes the idiomatic path the easiest path — is
falsifiable exactly once the 0.0.7 generators and 0.0.8 runtime-truth surfaces exist in a
published canary. This issue owns the one measured unfamiliar-agent smoke that tests it, as the
exit gate of 0.0.8. It consumes, and must not duplicate, the existing measurement chain
(#1102/#1201 capability, #1197 re-measurement, #1090 observational boxes).

Evidence

• Six consecutive measured runs with zero MCP/doctor/otel adoption (agentic: the agent-init harness had zero adoption on 0.0.4 — 0 MCP calls, 0 plugin doctor, 0 aspire otel, 0 skills across a full 452-call agent run #1197 body; corpus
  research/wave-5-6-plans.md [Wave 5] Package Quality — Apps Layer (sdk · service · fresh · fresh-ui) — 5a–5d ✅ MERGED · Structural Consolidation A–E ✅ · 3× IMPL-EVAL APPROVED #17).
• Wave-6 R3's only-GO run is confounded by the supervisor-enforced init gate, model, and canary
  (research/wave-6-runs.md Wave 1 — Contracts & schemas #7) — the strongest untested causal lever in the corpus.
• Full design: fable-5-remediation-plan/WAVE7-AND-AGENT-ADOPTION.md (this run).

Current surface

Wave-6 harness (brief v3 + PLAN-WAVE6): capability-map rows, executed-command census,
contamination rules. No arm structure; no generated-slice measurement rows (the verbs do not
exist yet); wrapper-exit gates only.

Target contract

Three arms on identical brief/budget: A = post-0.0.8 canary, no init enforcement; B = post-0.0.8,
R3-style enforced init gate; C = pre-remediation 0.0.6 canary (control). Two runs per arm,
different frontier models, blind gap-audit scoring on the Wave-6 rubric. New measured rows:
generated-slice verb usage (used / rejected-with-reason / silent = harness failure), consumer
no-any gate verdict, runtime-truth probes (receipt handling, child liveness, stream restart)
read from persisted state + one correlated trace.

Acceptance

• Arm design (A/B/C, 2 runs each) executed on published canaries with pinned versions recorded
• Generated-slice usage measured per run: used or explicitly rejected; zero silent rows
• Consumer-side type-escape gate executed against every product (not framework exports)
• Runtime-truth rows proved from persisted state + correlated OTEL, not wrapper exits
• Arm A vs C delta reported on frontend-composition and runtime-truth capability rows
• Negative path: a run that skips a generated verb without recording rejection is scored as a
  harness failure, not silently excused
• Verdict + confirmed residual gaps filed as issues (not folded into harness thickening)
• [post-merge] The verdict comment on this issue names the canary versions and links raw
  measurement artifacts

Boundaries

#1102/#1201 own MCP retrieval/corpus capability; #1197 owns the re-measurement method and
extraction script; #1090 owns the wave-five observational boxes (this issue may satisfy them and
says so per box, but does not absorb them). Harness texts live in .llm/harness/ + the owner's
wave folders, not on the board. This issue does not script product choices.

Docs/consumer proof

The published verdict is the program's go/no-go for advancing the train past remediation
(MILESTONE-TRAIN.md §5). A GO here is the evidence base for any future "0.1.0" claim.

Provenance

Seed run plan-fable5-remediation-roadmap--seed, PR #1347, 2026-08-08. Supervisor-authored
(Fable 5 · high); design in WAVE7-AND-AGENT-ADOPTION.md.

---

Filed from planning seed PR #1347 · source Draft-ID T7-01 · live issue #1381.

[rickylabs]

Milestone audit: this is explicitly a post-publish measured adoption run against 0.0.8 artifacts, so it cannot honestly block the artifact it measures. Moved to 0.0.9, where Wave 7 can compare the published 0.0.8 generated path while the frontend contribution layer is active.