Octer-Decision：会提出更好选项的递归自进化高速决策系统

01The fixed-candidate bottleneck固定候选集的决策瓶颈

Fast decision models can return a probability distribution over a supplied set of actions. Jev made this structured, probabilistic interface prominent; our team had already put Octer-Decision to work before Jev's public announcement. The limitation we address is the candidate set itself: when it omits a suitable response, its highest-scoring member is still drawn from an incomplete set.

高速决策模型能够针对给定动作集合输出概率分布。Jev 让结构化概率决策受到关注；在 Jev 公开发布前，我们已经将 Octer-Decision 投入使用。我们要解决的是候选集合本身的覆盖问题：当集合缺少合适动作时，最高分仍来自一个不完整的集合。

We therefore make candidate-set coverage part of inference. A learned sentinel lets our Choice head flag an inadequate action set. Routine requests take a single-step probability path; coverage gaps activate conditional generation of one structured candidate. We then use delayed outcomes to revise both the model and the harness that defines, composes, and validates actions.

因此，我们把候选集合的覆盖程度纳入推断。学习得到的特殊项让 Choice 头识别现有动作覆盖不足：常规请求走单步概率判断路径，覆盖缺口才触发一个结构化候选的条件生成。随后，我们利用延迟反馈同时修订模型，以及定义、组合、校验动作的 harness。

02Conditional generation and recursive updates条件生成与递归更新

We send Octer-Decision a point-in-time state, a typed action set, and execution constraints. Around the model, we built the components that construct that state, schedule requests, compose compatible actions, enforce constraints, and retain versioned outcome traces.

我们向 Octer-Decision 输入决策时状态、类型化动作集合和执行约束。围绕模型，我们构建了状态编译、请求调度、动作组合、约束执行和带版本的结果轨迹等组件。

What we added to the decision layer

我们在决策层增加的组件

Component	组件	System role
State / Outcome Compiler	Reconstructs the state available at decision time and marks delayed observations.	还原决策时可知的状态，并标记延迟到达的观测结果。
Decision Fabric	Schedules concurrent decisions and preserves versioned state–action traces.	调度并发判断，保留带版本的状态—动作轨迹。
Policy Composer	Combines local choices into a globally feasible, typed action set.	将局部判断组合为全局可执行的类型化动作集合。
Safety & Outcome	Applies permissions and hard constraints; links delayed outcomes to each version.	执行权限与硬约束，并将延迟结果关联到对应版本。

A choice model with a conditional open branch

带条件开放分支的 Choice 模型

We use a shared encoder and Choice head to produce calibrated probabilities over the supplied actions and NCA (“no correct action”). When NCA has the highest probability, our conditional autoregressive decoder proposes a new action; otherwise inference ends after the Choice head. We train in two stages: RLCD calibration of the encoder–Choice path, then decoder-only supervised fine-tuning while that path is frozen. Generated actions pass policy and safety checks before execution; in later RSI cycles, we can revise the next version after outcome-based validation.

我们用共享编码器与 Choice 头对给定动作及 NCA（“现有选项均不正确”）输出校准后的概率。若 NCA 概率最高，我们才让条件自回归解码器生成新动作；否则推断在 Choice 头结束。训练分两阶段：先以 RLCD 校准编码器与 Choice 路径，再冻结该路径，仅对解码器做监督微调。生成动作须经过策略与安全校验才能执行；在后续 RSI 周期中，我们根据结果验证修订下一版本。

Octer-Decision architecture: a calibrated Choice head routes NCA to a conditional autoregressive decoder while ordinary choices return directly — **Figure 1.** Octer-Decision's two-stage training and conditional inference path. NCA = no correct action; the orange decoder path runs only when NCA is selected. Click to enlarge.

The RSI feedback loop

RSI 如何形成闭环

We record the observed state, candidate set, selected action, model and harness versions, safety intervention, and delayed outcome in each trace. Our RSI loop proposes updates to both model weights and the harness; we promote a version only after offline evaluation and a guarded release.

我们在每条轨迹中记录观测状态、候选集合、最终动作、模型与 harness 版本、安全拦截及延迟到达的结果。我们的 RSI 闭环同时提出模型权重与 harness 的更新；新版本经过离线评估和受控发布后才进入下一轮。

System loop: a decision becomes a measured update to the next decision system

系统闭环：一次决策如何成为下一代决策系统的更新依据

1 · Compile state1 · 编译状态Versioned context, action schema, constraints, observation time带版本的上下文、动作模式、约束和观测时间

2 · Decide & compose2 · 决策与组合Octer-Decision probabilities + harness composition, thresholds, schedulingOcter-Decision 概率 + harness 的组合、阈值与调度

3 · Gate & record3 · 校验与记录Check constraints; execute or hold; persist a versioned state–action trace校验约束，执行或保持，并保存带版本的状态—动作轨迹

↓ The outcome matures and is joined to the original decision trace

↓ 结果成熟后，与原始决策轨迹对齐

4 · Measure outcome4 · 衡量结果Observed utility, delayed feedback, interventions, and attribution uncertainty观测效用、延迟反馈、拦截记录与归因不确定性

5 · RSI revises two layers5 · RSI 修订两层能力Model weights; harness action set, thresholds, and scheduling模型权重；harness 的动作集合、阈值与调度

6 · Evaluate & release6 · 评估与发布Offline replay, constraint regression, guarded rollout, version promotion离线回放、约束回归、受控上线、版本晋级

↶ Promoted model weights + harness version enter the next decision window

↶ 晋级的模型权重与 harness 版本进入下一个决策窗口

03High-frequency advertising decisions高频广告决策

3.1 Experimental protocol

3.1 实验协议

Our evaluation protocol assigns 300 campaigns to three arms (100 per arm): Jev, Octer-Decision, and Octer-Decision + RSI. We hold campaign eligibility, weekly spend allocation, attribution window, platform write frequency, and safety limits constant across arms. The eight-week run and two-week conversion-maturity window, with 15 targets per campaign, define 6.048 million target-hour decision opportunities. Octer-Decision keeps its initial model and harness; Octer-Decision + RSI receives validated updates to both after weeks 2, 4, and 6.

我们的评估协议将 300 个广告活动随机分为三组（每组 100 个）：Jev、Octer-Decision 和 Octer-Decision + RSI。我们让三组使用相同的活动资格、每周花费分配、归因窗口、平台写入频率和安全约束。八周投放加两周转化成熟窗口、每个活动 15 个投放目标，共定义 604.8 万个目标—小时决策窗口。Octer-Decision 沿用初始模型与 harness；Octer-Decision + RSI 在第 2、4、6 周后接收经验证的模型和 harness 更新。

Primary outcome:主指标： net profit per $100 of ad spend = 100 × (attributed revenue − product cost − ad spend) / ad spend.每 100 美元广告花费的净利润 = 100 ×（归因收入 − 商品成本 − 广告花费）/ 广告花费。

We use campaigns as the unit of randomization and inference, with campaign-clustered bootstrap resampling for 95% intervals. We measure latency and throughput in the same deployment envelope under a 128-concurrent-request load sample.

我们以活动为随机化与统计推断单位，按活动聚类 Bootstrap 计算 95% 区间。延迟与吞吐在同一实验部署中、128 并发请求的负载样本下测量。

Ad-bidding interface recording — **Recording 1.** An early interface recording of campaign selection, data analysis, and target-level bid decisions. It illustrates one class of advertising action; the experiment uses the protocol above. Click to enlarge.

3.2 Profit across eight weeks

3.2 八周收益轨迹

We use weekly net profit per $100 of ad spend to track the effect of continuous model-and-harness updates. Across the full experiment, Octer-Decision + RSI reaches $25.1, compared with $23.4 for Octer-Decision and $21.2 for Jev. That is a gain of $3.9 over Jev (campaign-clustered 95% interval $2.3–$5.5) and $1.7 over Octer-Decision ($0.2–$3.2).

我们用每 100 美元广告花费的周度净利润观察模型与 harness 持续更新的效果。整个实验中，Octer-Decision + RSI 为 25.1 美元，Octer-Decision 为 23.4 美元，Jev 为 21.2 美元。相对 Jev 的增益为 3.9 美元（按活动聚类的 95% 区间：2.3–5.5 美元）；相对 Octer-Decision 的增益为 1.7 美元（0.2–3.2 美元）。

Figure 2 · Weekly advertising profit

图 2 · 每周广告净利润

Same randomized campaigns throughout; orange dashed lines mark Octer-Decision + RSI releases after weeks 2, 4, and 6.

同一组随机分配的活动贯穿全程；橙色虚线标记 Octer-Decision + RSI 在第 2、4、6 周后的更新。

Figure 2. Each point is that week's matured, equal-spend-normalized outcome; the eight-week arm estimates appear in the text. The release markers show timing, while the randomized Octer-Decision arm supports the update comparison.

图 2。各点为当周成熟归因、按等额花费归一化后的结果；八周组均值见正文。发布标记仅显示时间，更新效果通过随机分配的 Octer-Decision 组比较。

3.3 Profit, latency, and decision capacity

3.3 收益、延迟与决策容量

To show the system trade-off directly, we plot the three arms by end-to-end P95 decision latency and eight-week net profit; circle area encodes peak completed decisions per second under the shared 128-concurrent-request load. Octer-Decision + RSI lies above and to the left of Jev: $25.1 versus $21.2, 0.22 s versus 0.44 s, and 1,050 versus 620 decisions/s.

为了直接展示系统权衡，我们把三组同时放在端到端 P95 延迟与八周净利润坐标上，圆点面积表示共用 128 并发负载下每秒完成的决策峰值。Octer-Decision + RSI 位于 Jev 的左上方：25.1 对 21.2 美元，0.22 对 0.44 秒，1,050 对 620 次决策/秒。

Figure 3 · Joint profit–latency–throughput result

图 3 · 收益—延迟—吞吐联合结果

X = P95 end-to-end seconds (lower is better); Y = net profit per $100 spend (higher is better); circle area = decisions/s; vertical line = 95% profit interval.

横轴为端到端 P95 秒数（越低越好）；纵轴为每 100 美元花费的净利润（越高越好）；圆点面积为每秒决策量；竖线为利润 95% 区间。

Figure 3. Each point summarizes the same randomized arm (100 campaigns). Profit intervals use campaign-clustered bootstrap estimates; P95 and throughput include state preparation, model inference, policy composition, and safety checks. The occasional generated-action route is included for Octer-Decision + RSI.

图 3。每个点概括同一随机实验组（100 个活动）。利润区间按活动聚类 Bootstrap 计算；P95 与吞吐包括状态准备、模型推断、策略组合及安全检查。Octer-Decision + RSI 的延迟统计包含按需生成新动作的路径。

3.4 Open-choice behavior in the same experiment

3.4 同一实验中的开放候选行为

We also track when Octer-Decision + RSI uses its open path. The Choice head selects an existing action in 95.8% of decision windows and activates candidate generation in 4.2%. Our rule checks find 94.1% of generated candidates feasible; execution still passes through the shared safety gate. The controller writes to the ad platform in 8.6% of windows, with no executed hard-limit violation recorded in this 300-campaign experiment.

我们还记录了 Octer-Decision + RSI 何时进入开放路径。Choice 头在 95.8% 的窗口直接选中既有动作，4.2% 的窗口触发新候选生成。规则校验判定 94.1% 的生成候选具备可执行性；实际执行仍须通过共用安全门。8.6% 的窗口产生广告平台写入，这项 300 活动实验未记录执行级硬约束违规。

04Interpretation and operating boundary结果解释与运行边界

We designed the three arms for two comparisons. Jev versus Octer-Decision compares the deployed decision policies; Octer-Decision versus Octer-Decision + RSI measures the effect of updating model weights and harness together. Both Octer-Decision arms include conditional candidate generation, so these results cover open-choice operation and total economic performance. We do not attribute a separate causal effect to the generation head under this protocol.

我们用三组设计回答两个问题：Jev 与 Octer-Decision 比较已部署的决策策略；Octer-Decision 与 Octer-Decision + RSI 衡量模型权重和 harness 联合更新的效果。两个 Octer-Decision 组均包含按需候选生成，因此这些结果覆盖开放候选的运行情况与整体经济效果。在这套协议下，我们不单独归因生成头的因果贡献。

Campaign-level randomization and matured attribution support the profit comparison; decision windows characterize load rather than independent profit samples. Release records are required to independently audit the historical claim that Octer-Decision preceded Jev's public launch.

利润比较依据活动级随机分组与成熟归因；决策窗口刻画系统负载，不作为独立利润样本。“早于 Jev 公开发布”的历史主张应由内部发布记录独立核验。

Conclusion结论

The RSI model is the brain.
Octer-Decision is the reflex arc.

RSI 大模型是大脑，
Octer-Decision 是反射弧。

In our system, the brain uses outcomes to discover and revise strategies, while the reflex arc makes frequent, constrained decisions. We bring them together in a loop that gives each new version a better set of actions to choose from.

在我们的系统中，大脑依据结果发现并修订策略，反射弧完成高频、受约束的判断。我们将两者接成闭环，让下一版系统拥有更好的可选动作。

External background: TypeSafe AI's Jev announcement.

外部背景：TypeSafe AI 的 Jev 发布说明。