From f1a0b0c62c090cf7f2f76dabf3e46740ff49c8c4 Mon Sep 17 00:00:00 2001
From: Eva
 <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
Date: Tue, 11 Aug 2026 20:55:46 -0400
Subject: [PATCH] bench: promote reply guard to lineage v3 (GEPA Q1 verdict)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Q1 probe on qemu-startup (the silent-stall exemplar), 6 arms x k=4:
control 3/4 with one narcoleptic finish (end_turn, no DONE, idle to the
900s wall); +stop_hooks 4/4; +require_reply 4/4; mia-v3 prompt 3/4 (its
fail was a grind, and Mie still got zero turns — the dispatch wording
does not activate, so v3-prompt goes back to the bench); b+c 3/4 (fail
was an agent crash mid-prompt, infra class); b+d 4/4.

k=8 confirm of require_reply alone: 8/8. Twelve passing trials total
and the guard never once fired — every pass sent DONE unprompted — so
this is armed insurance, not behavior change, and the honest claim is
'no regression, kills the only observed stall vector at its gate'.

require_reply wins over stop_hooks as the lineage change because it is
the smaller mechanism: one env var, in-process, no hook-server
dependency, nag budget shared with the existing stop_rejections cap.

tb-gepa-twins-v3 = tb-gepa-twins-v2 + require_reply:true on Mia only.
Prompts unchanged (mia-v2/mie-v0, same pinned sha256s). Condition hash
3eb6102eaf2a7e69, distinct from v2 8ad83a22af1a0bb3 and probe arm
b3a00123cc7cc1ab. orchestra: ruff clean, 288 tests pass.

Co-authored-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
Signed-off-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
---
 .../manifests/gepa/tb-gepa-twins-v3.yaml      | 46 +++++++++++++++++++
 1 file changed, 46 insertions(+)
 create mode 100644 benchmarks/harbor-buzz-orchestra/manifests/gepa/tb-gepa-twins-v3.yaml

diff --git a/benchmarks/harbor-buzz-orchestra/manifests/gepa/tb-gepa-twins-v3.yaml b/benchmarks/harbor-buzz-orchestra/manifests/gepa/tb-gepa-twins-v3.yaml
new file mode 100644
index 0000000000..6cb1e5e8c3
--- /dev/null
+++ b/benchmarks/harbor-buzz-orchestra/manifests/gepa/tb-gepa-twins-v3.yaml
@@ -0,0 +1,46 @@
+# GEPA lineage v3 = v2 + require_reply on Mia (reply guard, BUZZ_AGENT_REQUIRE_REPLY=1).
+# Promoted from Q1 per-problem probe on qemu-startup (silent-stall / "narcoleptic
+# finish" family: end_turn without DONE, then idle to the trial wall — 6+ of the
+# v2 full-89 timeouts share the shape).
+# Evidence: arm C went 4/4 (k=4 probe, jobs/gepa-q1-c) + 8/8 (k=8 confirm,
+# jobs/gepa-q1-c-confirm) on eva-team, control 3/4 with one narcoleptic death.
+# The guard never actually fired in a passing trial (zero nags in all 12
+# transcripts) — it is armed insurance, chosen over stop_hooks (also 4/4)
+# because it is the more minimal mechanism: one env var, in-process, no hook
+# server dependency, and its nag budget shares the existing stop_rejections cap.
+# v2 baseline: full-89 59/89 (66.3%) vs v0 49/89 — KEEP.
+schema_version: "1"
+condition: tb-gepa-twins-v3
+timing_signal: true
+roster:
+  - id: mia
+    kind: orchestrator
+    role: conn
+    require_reply: true
+    count: 1
+    endpoint: deepseek-ai/Deepseek-V4-Flash
+    model_revision: deepseek/deepseek-v4-flash-20260731
+    prompt:
+      path: personas/bench/gepa/mia-v2.md
+      sha256: cce6b223a0ac651ff762c4f50037b14f41d304571353a1e0797e30fffcb4b94d
+    generation:
+      thinking_effort: high
+  - id: mie
+    kind: worker
+    role: twin
+    count: 1
+    endpoint: deepseek-ai/Deepseek-V4-Flash
+    model_revision: deepseek/deepseek-v4-flash-20260731
+    prompt:
+      path: personas/bench/gepa/mie-v0.md
+      sha256: a4924e4c4507ac466a7500bb34a04a782618df41dba850b6f0f729787c48439b
+    generation:
+      thinking_effort: high
+prices:
+  deepseek-ai/Deepseek-V4-Flash:
+    input_per_million_usd: 0.14
+    cached_input_per_million_usd: 0.14
+    output_per_million_usd: 0.28
+    cache_read_rate: 0.0
+trial_budget:
+  timeout_seconds: 36000