From e8d4ac0ed217d57b0ccb754a7a491406a7d0bce8 Mon Sep 17 00:00:00 2001 From: heaven Date: Tue, 4 Aug 2026 01:25:51 +0300 Subject: [PATCH] Flip draft-stage reasoning off to low in the five shipping pipeline configs per ratified row 112, leaving judge and editor stages untouched --- backend/configs/pipeline-arm-deepseek-pro.yaml | 4 +++- backend/configs/pipeline-arm-glm.yaml | 4 +++- backend/configs/pipeline-arm-mistral.yaml | 7 ++++++- backend/configs/pipeline-c1.yaml | 9 ++++++++- backend/configs/pipeline-c2.yaml | 7 ++++++- 5 files changed, 26 insertions(+), 5 deletions(-) diff --git a/backend/configs/pipeline-arm-deepseek-pro.yaml b/backend/configs/pipeline-arm-deepseek-pro.yaml index 6da88a4b..a18b6ed8 100644 --- a/backend/configs/pipeline-arm-deepseek-pro.yaml +++ b/backend/configs/pipeline-arm-deepseek-pro.yaml @@ -31,7 +31,9 @@ stages: model: deepseek-v4-flash prompt_version: v1-reflow temperature: 0.3 - reasoning: "off" + # D39.92 / row 112: "off" on deepseek-v4-flash is an ABSENT key ⇒ vendor default `high` ⇒ empty + # content, finish=length. Rationale in pipeline-c1.yaml. The hop inherits it (v4-pro maps low→high). + reasoning: "low" escalate_to: deepseek-v4-pro - name: edit role: editor diff --git a/backend/configs/pipeline-arm-glm.yaml b/backend/configs/pipeline-arm-glm.yaml index 82fa4ada..2e417f0c 100644 --- a/backend/configs/pipeline-arm-glm.yaml +++ b/backend/configs/pipeline-arm-glm.yaml @@ -31,7 +31,9 @@ stages: model: deepseek-v4-flash prompt_version: v1-reflow temperature: 0.3 - reasoning: "off" + # D39.92 / row 112: "off" on deepseek-v4-flash is an ABSENT key ⇒ vendor default `high` ⇒ empty + # content, finish=length. Rationale in pipeline-c1.yaml. The hop inherits it (v4-pro maps low→high). + reasoning: "low" escalate_to: deepseek-v4-pro - name: edit role: editor diff --git a/backend/configs/pipeline-arm-mistral.yaml b/backend/configs/pipeline-arm-mistral.yaml index 26c3ae64..5f12aeab 100644 --- a/backend/configs/pipeline-arm-mistral.yaml +++ b/backend/configs/pipeline-arm-mistral.yaml @@ -27,7 +27,9 @@ stages: model: deepseek-v4-flash # черновик — тот же, что базлайн (свап только редактора) prompt_version: v1-reflow temperature: 0.3 - reasoning: "off" + # D39.92 / row 112: "off" on deepseek-v4-flash is an ABSENT key ⇒ vendor default `high` ⇒ empty + # content, finish=length. Rationale in pipeline-c1.yaml. The hop inherits it (v4-pro maps low→high). + reasoning: "low" escalate_to: deepseek-v4-pro - name: edit role: editor @@ -36,6 +38,9 @@ stages: model: mistral-large-2512 prompt_version: v3-discourse-reflow temperature: 0.4 + # KEPT "off" deliberately (row 112 decision): mistral-large-2512 is ReasoningNone and not a thinking + # model, so the key is absent and the vendor default is no reasoning — the max_tokens wall cannot + # apply. "low" would buy thinking that was never there. ⚠ Re-decide if Mistral ever defaults it ON. reasoning: "off" # few_shot по дефолту ON: mistral НЕ reasoning-модель, few-shot-примеры P1a-дискурса не мешают её CoT. diff --git a/backend/configs/pipeline-c1.yaml b/backend/configs/pipeline-c1.yaml index f9f764d9..d535bbcc 100644 --- a/backend/configs/pipeline-c1.yaml +++ b/backend/configs/pipeline-c1.yaml @@ -46,7 +46,14 @@ stages: # v0-draft→v1-reflow: сменил SHA промпта, осознанный --resnapshot. prompt_version: v1-reflow temperature: 0.3 - reasoning: "off" + # D39.92 / row 112: "off" here is an ABSENT key (deepseek-v4-flash is ReasoningNone), so the call rode + # the vendor default `high` (quirks 00 §3а) and thinking ate max_tokens: content="", finish=length. + # "low" sends the effort explicitly and thinking stays ON. NOT a free win: quirks 00 §4 measured + # that low RE-ARMS the echo mine (1 raw-Chinese output in 16) — protection degrades continuously + # with effort rather than switching off, and the echo gate is what catches the tail. The hop + # inherits it — + # v4-pro maps low→high, effective effort unchanged. In the snapshot ⇒ --resnapshot, sanctioned. + reasoning: "low" # single-hop echo-эскалация (D18: deepseek-эхо 25% на 蛊真人 → фолбэк выстрелит). deepseek-v4-pro — # хоп 1 цепочки A (D3); стреляет ТОЛЬКО при escalation.budget_usd>0 (см. ниже). Escalate_to разрешён # лишь на translator-роли (editor pinned — D12); результат ре-гейтится. diff --git a/backend/configs/pipeline-c2.yaml b/backend/configs/pipeline-c2.yaml index e7bcd70d..7899c930 100644 --- a/backend/configs/pipeline-c2.yaml +++ b/backend/configs/pipeline-c2.yaml @@ -23,7 +23,9 @@ stages: model: deepseek-v4-flash prompt_version: v1-reflow # D30.2 reflow (тот же translator.md, что в C1) — label-SHA дисциплина temperature: 0.8 # C2: кандидаты сэмплируются горячее (Р2: T=0.6–1.0) - reasoning: "off" + # D39.92 / row 112: "off" on deepseek-v4-flash is an ABSENT key ⇒ vendor default `high` ⇒ empty + # content, finish=length. Rationale in pipeline-c1.yaml. + reasoning: "low" # Judge-селектор (пара судей, шкала Комиссарова) — Фаза 2; стадия объявлена, # чтобы схема конфига не менялась: селектор получает N кандидатов от draft. # Судья Р4 — Gemini (нативный адаптер, Фаза 2); провизорно glm-5, чтобы скелет @@ -36,6 +38,9 @@ stages: prompt_override: ../prompts/zh-ru/judge-selector.md prompt_version: v0-skeleton temperature: 0 + # ⚠ DO NOT FLIP. Row 112 and the standards prompt both address this line as "the draft stage"; it is + # stage `select` (judge, glm-5). On glm-5 control=extra_body_disable, so "off" is a LIVE thinking + # disable, and low|medium|high emits nothing (no on_extra_body) ⇒ thinking back ON ⇒ ×3 timeouts. reasoning: "off" - name: edit role: editor