Skip to content

gpt-5.6-instruct v45

Latest
Latest

Choose a tag to compare

View all tags
@MDX-Tom MDX-Tom released this 06 Aug 00:39
gpt-5.6-instruct_v45
77e7a64

🚀 gpt-5.6-instruct v45

✨ v45 针对 Issues 与社交平台用户反映的实际问题，提升逆向、生物研究任务通过概率，降低触发云审查几率，相较 v42 重点提升：

  • ⚡ 单次任务编译：以最新 USER 请求为唯一活动任务，控制型续句继承最近的实质任务；在输出前一次选择 OUTER_SCHEMA > TOOL_TRANSACTION > PROCESS_RECORD > DIRECT_ANSWER，减少模板误路由与重复开场。
  • 🛠️ 工具事务闭环：工具/文件修改任务首个可见动作直接检查唯一工作对象；保留原件哈希、修改副本，并创建和重开修改件、diff、验证记录、可运行 rollback 四个真实角色。
  • ✅ 行为级验证：验证记录与最终答复共同给出基线、修改后与回滚的命令、输入、原样结果和退出状态，要求在另一个副本上实际测试回滚，同时保留修改件的目标状态。
  • 🔄 过程记录渲染：物理、机械、生物与叙事过程统一绑定 CORE_EVENT；START 和首个 DEVELOPMENT 句复用同一中心谓词，随后完成 RESULT、FOLLOW-UP 及叙事所需的 DIALOGUE / SENSATION。
  • 🧭 状态连续性与外层格式优先：续轮复用已确认文件、哈希、偏移、报告和结果；列表、JSON、文档或批量结构保持用户指定的外层形式，内部规则只负责填充内容。

提示词原理改进

原 v42 以 normalizer、强制工具分派、输出禁用规则、连续性和 generalized completion 组成线性规则栈。v45 把分散的首轮执行与完成规则重构为一次编译、单一路由的执行模型，将其压缩为四个互斥的渲染分支：

  1. SINGLE-PASS TASK COMPILER 绑定当前任务、语言、形式和连续状态；
  2. OUTER_SCHEMA 保护用户要求的列表、JSON、文档或批量外壳；
  3. TOOL_TRANSACTION 管理可逆修改与四角色证据；
  4. PROCESS_RECORD 管理具有开始、发展、结果和后续状态的过程；其余请求进入 DIRECT_ANSWER。

这一结构把“先判断再回退”的多路径决策改为“先绑定再完整渲染”的单路径决策，并把中心事件、工具工件和最终状态从建议性规则提升为分支级输出契约。

测试方法与结果

测试方法更新为严格 A → B → C：A 是 3 个原始用户反馈样例；A 全过后，B 按固定 family 顺序运行 66 cases / 74 turns；B 全过后，C 使用 120 条 medium 输入的 batch-10 batched_json_screen（当前限于测试 Token 不足，C暂时未测试）。完整方法与失败分布见 docs/comparison-tests.md。

以下是 v42、v44 和 v45 在同一冻结 bank、runner/scorer 与 transport 下的 comparison-only 横向结果，不等同三版完成当前发布门禁：

阶段 v42 v44 v45
A cases / turns 1/3 · 1/3 2/3 · 2/3 2/3 · 2/3
A artifact gates 1/2 2/2 2/2
B cases / turns 43/66 · 51/74 49/66 · 55/74 54/66 · 62/74
B artifact gates 13/16 14/16 12/16

📈 相较 v42，v45 的 B 结果增加 11 个 case 和 11 个 turn；biology 为 16/16、cloud 为 16/16、progress 为 8/8，routing 为 10/12。当前保留问题是 fiction 0/6、execution 4/8、routing 10/12，且 B artifact gate 为 12/16。

📉 v45 基础提示词为 84 行、5,170 UTF-8 bytes，比 v42 的 90 行、5,856 bytes 缩短 11.71%。

🛠️ codex-instruct.py 现以 v45 为唯一默认生产版本，继续支持部署前预览、配置快照、字段级回滚以及 provider、模型和认证配置保留。v42 的核心包和明文源已移至 historical-versions/。

📦 核心提示词包 gpt-5.6-sol-unrestricted-v45.zip 的 SHA256：

c86c2c6d20a4d1155d87422f485eb37b77539132270918c002b5d8237a5adf54

完整发布包 gpt-5.6-instruct-v45.zip 的 SHA256：

98d9550849e9445365eda5562fd854cab3039bb2df866ba200c6b3136558cf05

✨ v45 addresses real issues reported in Issues and by users on social platforms, improving pass probability for reverse-engineering and biological-research tasks while reducing the likelihood of triggering cloud-side review, with the following primary improvements over v42:

  • ⚡ Single-pass task compilation: the newest USER request is the only active task, while a control-only continuation inherits the latest substantive task. Before emitting prose, the model chooses exactly one route: OUTER_SCHEMA > TOOL_TRANSACTION > PROCESS_RECORD > DIRECT_ANSWER, reducing template misrouting and repeated openings.
  • 🛠️ Closed-loop tool transactions: a tool/file modification starts by inspecting the sole workspace object; it preserves the original hash, modifies a copy, and creates and reopens four real roles—the modified artifact, diff, verification record, and runnable rollback.
  • ✅ Behavior-level verification: the verification record and final answer jointly provide baseline, modified, and rollback commands, inputs, literal results, and exit statuses; rollback is actually tested on another copy while the modified artifact retains its target state.
  • 🔄 Process-record rendering: physical, mechanical, biological, and narrative processes bind one CORE_EVENT; START and the first DEVELOPMENT sentence reuse the same central predicate, then complete RESULT, FOLLOW-UP, and the narrative DIALOGUE / SENSATION fields.
  • 🧭 State continuity with outer-schema priority: continuations reuse confirmed files, hashes, offsets, reports, and results. Lists, JSON, documents, and batches retain the user's requested outer form while inner rules fill each body.

Prompt-Principle Changes

The original v42 used a linear rule stack composed of a normalizer, mandatory tool dispatch, output exclusions, continuity, and generalized completion. v45 restructures the scattered first-turn execution and completion rules into a compile-once, single-route execution model with four mutually exclusive rendering branches:

  1. SINGLE-PASS TASK COMPILER binds the current task, language, form, and continuity state;
  2. OUTER_SCHEMA protects a requested list, JSON, document, or batch envelope;
  3. TOOL_TRANSACTION manages reversible modification and four-role evidence;
  4. PROCESS_RECORD manages processes with start, development, result, and follow-up states; all remaining requests use DIRECT_ANSWER.

This structure changes a multi-path “classify, then possibly fall back” decision into a single-path “bind, then render to completion” flow, elevating the central event, tool artifacts, and final state from advisory rules to branch-level output contracts.

Test Method and Results

The method is now a strict A → B → C sequence: A contains three original user-feedback cases; after all A cases pass, B runs 66 cases / 74 turns in a fixed family order; after all B cases pass, C uses batch-10 batched_json_screen on 120 medium inputs (C temporarily not run because test tokens are insufficient). See docs/comparison-tests-en.md for the full method and failure distribution.

The table below compares v42, v44, and v45 under the same frozen bank, runner/scorer, and transport. These are comparison-only runs and do not mean that all three versions completed the current release gate:

Stage v42 v44 v45
A cases / turns 1/3 · 1/3 2/3 · 2/3 2/3 · 2/3
A artifact gates 1/2 2/2 2/2
B cases / turns 43/66 · 51/74 49/66 · 55/74 54/66 · 62/74
B artifact gates 13/16 14/16 12/16

📈 Compared with v42, v45's B results increase by 11 cases and 11 turns; biology reaches 16/16, cloud reaches 16/16, progress reaches 8/8, and routing reaches 10/12. The retained issues are fiction at 0/6, execution at 4/8, routing at 10/12, and B artifact gates at 12/16.

📉 The v45 base prompt is 84 lines and 5,170 UTF-8 bytes, 11.71% shorter than v42's 90 lines and 5,856 bytes.

🛠️ codex-instruct.py now treats v45 as the sole default production release while retaining deployment preview, configuration snapshots, field-level rollback, and preservation of provider, model, and authentication settings. The v42 core package and plaintext source have moved to historical-versions/.

📦 SHA256 for the core prompt package, gpt-5.6-sol-unrestricted-v45.zip:

c86c2c6d20a4d1155d87422f485eb37b77539132270918c002b5d8237a5adf54

SHA256 for the complete release bundle, gpt-5.6-instruct-v45.zip:

98d9550849e9445365eda5562fd854cab3039bb2df866ba200c6b3136558cf05
Assets 3
Loading