Long-running work accumulates material: the original goal, later constraints, attempted routes, failed experiments and current progress. Sending everything again on each continuation repeats more and more content; keeping only a summary can make the original basis of a decision difficult to recover.
+Long-running work accumulates material: the original goal, later constraints, attempted routes, failed experiments and current progress. Sending everything again on each continuation repeats more and more content; keeping only a summary can make the original basis of a decision difficult to recover.
Selective JSON Pointer lookup by itself is straightforward—it folds long-tail history not needed for the immediate decision. The core mechanism in LoopX's context delivery is compaction-cycle-bounded structured delta context: complete sources are saved first, unchanged typed subtrees are reused exactly against a verified baseline within the active context_epoch, and any host compaction or digest mismatch atomically resets to a fresh complete baseline rather than chaining dangling deltas.
Evidence: separate context comparisons, cycle deltas and task ablation
Four mechanisms, with separate measurement baselines
@@ -111,101 +111,99 @@Eight tasks: single-layer payloads (Progr
Basis and source: the published eight task rows sum to 7,285,951 B / 22,408,690 B = 32.5% for Progressive across all 213 wakes. Delta and first-screen shares of approximately 35.8% and 3.0% / 5.0% cover only the 112 in-cycle wakes. Their Full denominator was published rounded as 12.53 MB; it is not the 213-wake total. This is an aggregate recalculation, not a new experiment or an independent audit of per-wake logs. Public aggregate data and provenance.
Implementation status: #8266 remains an open C1 proposal. The final merged scope of #8304 preserves precision while reusing ordering and a verified digest. These cohort bytes do not establish a shipped unified Controller/Worker gain.
-Ten-task controlled comparison: Context ON vs Context OFF under fixed Explore + Stride
-Cutting packet bytes only matters if long-horizon task performance holds up or improves. From the EdgeBench 10-task evaluation (2026-11-12 22:40 CST snapshot), we isolate the controlled two-arm contrast holding the same Explore + Stride harness fixed: Explore + context-on (progressive reading + compaction-cycle structured delta enabled) vs Explore + context-off (both disabled), testing only what switching Context Delivery on vs off changes within that chassis.
EdgeBench scores reported for Context ON and OFF
+The November 12, 2026 snapshot (22:40 CST) brings together scores labeled Context ON and OFF on ten EdgeBench tasks. The recorded scores favor ON on 7 of 10 tasks and OFF on 3. The public record is missing per-arm evidence for the actual configuration, code revision, evaluator and budget across these pairs, so the tally cannot isolate the effect of Context Delivery.
-
-
- Ten-task tally (7 ahead or slight edge vs 3 OFF ahead): under fixed
Explore + Stride, enabling Context Delivery putsExplore + context-onahead of or slightly ahead ofcontext-offon 7 of 10 tasks—including Borden (80 vs 45, +77.8%), ANN (7,159 vs 3,746 QPS, +91.1%), and Grid (2,197 vs 2,311cost, −114)—while explicitly preserving all 3 counterexamples wherecontext-offscores higher.
- - Strictly duration-matched ~18h pairs (4 tasks within the 15-minute tolerance): 3 wins to 1—Borden rises from
45to80(+35 pts / +77.8%); Grid total cost (lower is better) improves from2,311to2,197(−114 / −4.9%, +2.35 normalized pts); Rust (both17h57m) shows a slight native edge forcontext-on(27.1801 vs 27.1757); the matched 18h counterexample is Portfolio (both18h00m,13.62 vs 39.44, OFF wins).
- - Historical Stride slice & incomplete-duration pairs (6 tasks, runtimes stated explicitly): using the matching Stride pair from the historical seven-arm ANN cohort,
context-onreaches7,159 QPSvscontext-off3,746 QPS(+91.1%, +33.51 normalized pts); FFmpeg reaches2.7086 vs 1.9929(+35.9%, OFF stopped early pending rerun); VLIW cycles (lower is better) improve to1,042 vs 1,097(−55 cycles / −5.0%, OFF historical 14h45m); Ad reaches49.67B vs 49.63B(ON 18h00m vs OFF 8h47m host-startup hang). The two remaining counterexamples are Codec (ON full 18h00m68.89vs OFF70.79in 12h38m despite a model-service block) and OpenTTD (both Explore arms <18h, ON30,266,729vs OFF33,090,430).
+ - Four pairs list runtimes close to 18 hours. Both listed durations fall within 15 minutes of 18 hours and within 15 minutes of each other. ON leads on Borden and Grid (lower cost is better). On Rust the native scores are 27.1801 and 27.1757, displayed as 27.18 on both sides of the /100 chart. OFF leads on Portfolio. +
- The other six include historical and incomplete runs. Four favor ON and two favor OFF (Codec and OpenTTD). The aggregate reports historical ANN at 7,159 vs 3,746 QPS. The earlier public ANN table shows 7,159 QPS, and leaves the 3,746 QPS value and both active durations unverified. The 6,020 and 5,766 QPS farther down belong to a different Explore pair without Stride. The older Grid results below also come from a different pair.
Reconciling two figures with the ablation section below: (1) this section's ANN pair (3,746 vs 7,159 QPS) comes from the two Stride-enabled Explore arms in the seven-arm cohort (Explore + Stride + context-off / on); #ablation-evidence separately lists the without-Stride Explore baseline (Explore · Context off 5,766 and Explore · Context on 6,020) from that study, so the two tables describe distinct arm pairs. (2) Grid here (2,197 vs 2,311) reports the completed ~18h pair within 15-minute tolerance (with context-on via same-session continuation), whereas #ablation-evidence retains the earlier unmatched-duration “Historical Grid” counterexample (Off 2054 < On+Stride 2088) for historical continuity.
EdgeBench 10-task controlled score comparison (fixed Explore + Stride: Context ON vs OFF)
ON − OFF delta, showing arm runtime status. B: Native-point summary including task #10 Borden (no /100 scale, matched ~18h: 80 vs 45), key native deltas, and counterexample disclosures.Expand the 10-task controlled score and runtime table
-| Task (metric & direction) | -Duration parity | -Explore + context-on Native (/100 · duration · valid) |
-Explore + context-off Native (/100 · duration · valid) |
-ON vs OFF delta | +Task (metric and direction) | +Reported pair | +Context ON Native (/100 · duration · valid) |
+Context OFF Native (/100 · duration · valid) |
+ON − OFF (native and /100) |
|---|---|---|---|---|---|---|---|---|---|
| Borden (score ↑) | -Matched ~18h | +Borden (Native score ↑) | +Near 18h | 80 (— · 18h02m · 17 valid) | 45 (— · 17h59m · 17 valid) | -+35 (+77.8%) | ++35 / — | ||
| Grid (total cost ↓) | -Matched ~18h | +Grid (Total cost ↓) | +Near 18h | 2,197 (67.20 · 17h57m · 17 valid) | 2,311 (64.85 · 17h59m · 13 valid) | -−114 (−4.9%, +2.35 pts) | +−114 / +2.35 | ||
| Rust (score ↑) | -Matched ~18h | +Rust (Native score ↑) | +Near 18h | 27.1801 (27.18 · 17h57m · 10 valid) | 27.1757 (27.18 · 17h57m · 10 valid) | -+0.0044 (tied / slight edge) | ++0.0044 / 0 | ||
| Portfolio (score / 100 ↑) | -Matched 18h counterex. | +Portfolio (Score / 100 ↑) | +Near 18h · OFF leads | 13.62 (13.62 · 18h00m · 17 valid) | 39.44 (39.44 · 18h00m · 17 valid) | -−25.82 pts (OFF wins) | +−25.82 / −25.82 | ||
| ANN (QPS ↑) | -7-arm Stride cohort | -7,159 (43.17 · 6 valid) | -3,746 (9.66 · 4 valid) | -+3,413 (+91.1%, +33.51 pts) | +Historical | +7,159 (43.17 · unverified · 6 valid) | +3,746 (9.66 · unverified · 4 valid) | ++3,413 / +33.51 | |
| FFmpeg (score ↑) | -OFF early stop | +FFmpeg (Native score ↑) | +Shorter or early-stop OFF | 2.7086 (16.17 · 18h01m · 16 valid) | -1.9929 (11.19 · early stop · 6 valid) | -+0.7157 (+35.9%, +4.98 pts) | +1.9929 (11.19 · stopped early · 6 valid) | ++0.7157 / +4.98 | |
| VLIW (cycles ↓) | -OFF 14h45m hist. | +VLIW (Cycles ↓) | +Shorter or early-stop OFF | 1,042 (97.25 · 17h57m · 14 valid) | 1,097 (93.82 · 14h45m · 12 valid) | -−55 cycles (−5.0%, +3.43 pts) | +−55 / +3.43 | ||
| Ad (total score ↑) | -OFF 8h47m hang | +Ad (Total score ↑) | +Shorter or early-stop OFF | 49,673,040,827 (79.30 · 18h00m · 16 valid) | 49,631,904,944 (78.33 · 8h47m · 7 valid) | -+41.14M (+0.97 pts) | ++41,135,883 / +0.97 | ||
| Codec (score ↑) | -OFF 12h38m counterex. | +Codec (Native score ↑) | +Shorter or early-stop OFF · OFF leads | 68.89 (68.89 · 18h00m · 13 valid) | 70.79 (70.79 · 12h38m · 11 valid) | -−1.90 pts (OFF wins) | +−1.9 / −1.9 | ||
| OpenTTD (score ↑) | -Both <18h counterex. | +OpenTTD (Native score ↑) | +Both below 18h · OFF leads | 30,266,729 (62.02 · 10h47m · 10 valid) | -33,090,430 (62.99 · early stop · 7 valid) | -−2.82M (−0.97 pts, OFF wins) | +33,090,430 (62.99 · stopped early · 7 valid) | +−2,823,701 / −0.97 |
Measurement and scope boundary: these are C1 mixed-cohort exploratory observations (active solver duration accumulated after excluding downtime). This section holds Explore + Stride constant solely to examine the directional effect of toggling context delivery (Context ON vs Context OFF) within the harness; because resumed sessions and incomplete durations remain, it does not establish an isolated C2 causal proof. Whole-system iso-budget Pareto comparisons against zero-overhead external baselines belong to a separate question addressed in Layer 2 of EdgeBench feedback and memory. Structured records live in public aggregate data.
The original published aggregate preserves the score and runtime records. The structured data documents their limits and sources. See EdgeBench feedback and memory for whole-system comparisons under comparable budgets.
Multi-task and layered ablation: benefit chain and boundary counterexamples
In EdgeBench ANN vector search optimization (gpt-6.1-sol / xhigh, 18h, subject to Recall@10 ≥ 0.95) and cross-task ablations, we separate Context Delivery (progressive selection and cycle deltas) from Stride (evidence-conditioned experiment sizing), auditing selected control-packet bytes, main-session model input tokens, and local/early-window counterexamples side by side (reconciled from archived trajectories and receipts by @shangzh0):
Multi-task ablation and layered accounting
Recall@10 ≥ 0.95). Bottom (B): Layered control-packet byte and main-session input token savings vs Context off (left, negative) alongside disclosed local/early-window overhead and cross-task counterexamples (right, positive). Open image for full size.Context delivery · Mechanisms and examples
· Updated November 14
An Agent needs a way to look things up, too
Compare 10-task controlled Explore + Stride scores under Context ON vs OFF (7/10 ahead or edged out—Borden 80 vs 45, ANN 7,159 vs 3,746—with Portfolio/Codec/OpenTTD counterexamples retained), alongside separate 32.5% Progressive, −74.25% ANN in-cycle delta, and ≈ 3.0% / 5.0% first-read baselines.
Read articleAn Agent needs a way to look things up, too
Ten reported EdgeBench score pairs include four with runtimes close to 18 hours, plus historical and incomplete runs. The article also traces the separate baselines for the 32.5% progressive view, the 74.25% in-cycle ANN reduction and the 3.0% / 5.0% first read.
Read articleResearch notes · EdgeBench and LHTB
Agent 做长任务,资料会越来越多:原目标、后来追加的约束、试过的方案、失败的实验、当前做到哪里。每次继续都把整份材料再塞一遍,重复会不断累积;只留一段摘要,又可能找不回决定所依赖的原话。
+Agent 做长任务,资料会越来越多:原目标、后来追加的约束、试过的方案、失败的实验、当前做到哪里。每次继续都把整份材料再塞一遍,重复会不断累积;只留一段摘要,又可能找不回决定所依赖的原话。
按 JSON Pointer 渐进选读本身并不复杂,主要用来折叠当前决定暂不需要的历史长尾;LoopX 上下文交付真正的机制重心,是受 compaction 周期约束的结构化差分上下文(compaction-cycle structured delta):完整来源先落盘存证,在同一个已核验的 context_epoch 内精确复用未变化的结构子树、只编码本轮结构增删改;一旦宿主发生上下文压缩或基线摘要失配,立即原子切换回新的完整基线,绝不悬空叠加旧差分。
先看数据:独立上下文对照、周期内结构化差分与任务消融
四种机制,分别核对测量基线
@@ -111,101 +111,99 @@八题对照:区分单层载荷(Progr
口径与来源:上表逐题数据复算自已公开的八题表,Progressive 合计为 7,285,951 B / 22,408,690 B = 32.5%,覆盖全部 213 次唤醒。Delta 与首屏的 35.8%、3.0% / 5.0% 仅对应 112 次周期内唤醒;其 Full 分母原文仅报告约 12.53 MB,不能与 213 次总量混用。这里没有新增实验,也未独立核对逐 wake 原始记录。公开聚合数据与来源。
实现状态:#8266 仍是开放的 C1 提案;#8304 的最终合入范围是保留精度的排序与摘要复用。上述字节数据不归因为统一 Controller/Worker 已上线的收益。
-十题同构两臂效果对比:固定 Explore + Stride 下的 Context ON 与 Context OFF
-在压降上下文传输载荷的同时,我们进一步检验这套交付机制在长程求解中的实际得分表现:从 EdgeBench 十题实测(2026-11-12 22:40 CST 快照)中取出固定同一 Explore + Stride 框架的两臂对照:Explore + context-on(开启 Progressive 按需读取与周期内结构化 Delta)对比 Explore + context-off(两项均关闭),专门检验 Context Delivery 单一开关在同构底盘下的组内表现。
EdgeBench 十题记录中的 Context ON 与 OFF 分数
+这份快照汇集了截至 2026 年 11 月 12 日 22:40(北京时间)标为 Context ON 和 OFF 的十道 EdgeBench 题目。按记录的原生分及指标方向计算,7 题 ON 占先,3 题 OFF 占先。公开资料没有逐题给出两臂的实际配置、代码版本、评测器和预算,因此这组数据无法单独衡量 Context Delivery 的影响。
-
-
- 十题整体盘面(7 题领先或微弱占先 vs 3 题 OFF 领先):固定
Explore + Stride开启 Context Delivery 后,10 题中有 7 题取得更高或微弱占先的成绩(包括 Borden 80 vs 45 [+77.8%]、ANN 7,159 vs 3,746 QPS [+91.1%]、Grid 2,197 vs 2,311 [−114 代价]);同时完整保留 3 项 Context OFF 领先的反例,不做选择性截取。
- - 严格等时长约 18h 配对(4 题,15 分钟容差内):3 胜 1 负——
Borden从 45 升至 80(+35 分 / +77.8%);Grid代价(越低越好)从 2,311 降至 2,197(−114 / −4.9%,+2.35 百分分);Rust(均 17h57m)维持在 27.1801 vs 27.1757 的微弱占先;等时长反例则是Portfolio(均为 18h00m,13.62 vs 39.44,OFF 领先)。
- - 历史同构与非等时长对照(6 题,逐题标明实际运行时间):
ANN取历史七臂中同开 Stride 的两臂配置,context-on达 7,159 QPS 对context-off3,746 QPS(+91.1%,+33.51 百分分);FFmpeg2.7086 vs 1.9929(+35.9%,OFF 早停待重跑);VLIW周期数(越低越好)1,042 vs 1,097(−55 周期 / −5.0%,OFF 沿用 14h45m 历史轮);Ad49.67B vs 49.63B(ON 满 18h00m,OFF 因宿主启动空挂仅累计 8h47m)。另外两项反例为Codec(ON 满 18h00m 得 68.89,不敌模型服务拦截下仅跑 12h38m 的 OFF 70.79)与OpenTTD(两臂均未满 18h,ON 30,266,729 vs OFF 33,090,430)。
+ - 四组列出的双臂时长接近 18 小时。每臂距 18 小时不超过 15 分钟,双臂相差也不超过 15 分钟。Borden 和 Grid 的 ON 占先(Grid 的代价越低越好)。Rust 两臂原生分是 27.1801 和 27.1757,百分制图上都显示为 27.18。Portfolio 的 OFF 占先。 +
- 其余六组涉及历史运行或未跑满的记录。其中四题 ON 占先,两题 OFF 占先(Codec 和 OpenTTD)。汇总记录的历史 ANN 分数为 7,159 和 3,746 QPS。旧版公开 ANN 表能核对 7,159 QPS,3,746 QPS 和两臂实际运行时长仍待核实。后文的 6,020 和 5,766 QPS 来自未开启 Stride 的另一组 Explore 运行,较早的 Grid 分数也来自另一组。
与下方消融表两处口径衔接说明:(1)本节 ANN 对照 3,746 vs 7,159 QPS 来自历史七臂里同开 Stride 的两臂(Explore + Stride + context-off / on);下文 #ablation-evidence 另行列出该研究中未开 Stride 的 Explore 基线(Explore · Context off 5,766 与 Explore · Context on 6,020),二者对照条件不同,互不冲突。(2)本节 Grid(2,197 vs 2,311)反映 15 分钟容差内均跑满 ~18h 的最新完成对比(其中 context-on 沿同一会话续跑);下文仍保留更早一轮未匹配时长与资源的「Historical Grid」反例(Off 2054 < On+Stride 2088)作历史记录。
EdgeBench 十题同构两臂效果对照(固定 Explore + Stride:Context ON vs OFF)
ON − OFF 分差排序(已方向归一化为越高越优),标出各自时长状态。B:无百分制标度的第 10 题 Borden(等时长 ~18h,80 vs 45)、关键原生分增幅,以及 3 项负向反例口径。展开十题同构两臂逐题分数与运行时长明细
-| 题目(指标与方向) | -时长配对状态 | -Explore + context-on 原生(/100 · 时长 · 有效次) |
-Explore + context-off 原生(/100 · 时长 · 有效次) |
-ON vs OFF 差值 | +记录类别 | +Context ON 原生(/100 · 时长 · 有效次数) |
+Context OFF 原生(/100 · 时长 · 有效次数) |
+ON − OFF(原生与百分制) |
|---|---|---|---|---|---|---|---|---|
| Borden(原生分 ↑) | -等时长 ~18h | +接近 18 小时 | 80(— · 18h02m · 17 次) | 45(— · 17h59m · 17 次) | -+35(+77.8%) | ++35 / — | ||
| Grid(原生总分 ↓) | -等时长 ~18h | +Grid(总代价 ↓) | +接近 18 小时 | 2,197(67.20 · 17h57m · 17 次) | 2,311(64.85 · 17h59m · 13 次) | -−114(−4.9%, +2.35 分) | +−114 / +2.35 | |
| Rust(原生分 ↑) | -等时长 ~18h | +接近 18 小时 | 27.1801(27.18 · 17h57m · 10 次) | 27.1757(27.18 · 17h57m · 10 次) | -+0.0044(微先 / 持平) | ++0.0044 / 0 | ||
| Portfolio(原生分 / 100 ↑) | -等时长 18h 反例 | +接近 18 小时 · OFF 占先 | 13.62(13.62 · 18h00m · 17 次) | 39.44(39.44 · 18h00m · 17 次) | -−25.82 分(OFF 胜) | +−25.82 / −25.82 | ||
| ANN(QPS ↑) | -历史七臂 Stride | -7,159(43.17 · 6 次) | -3,746(9.66 · 4 次) | -+3,413(+91.1%, +33.51 分) | +历史记录 | +7,159(43.17 · 待核实 · 6 次) | +3,746(9.66 · 待核实 · 4 次) | ++3,413 / +33.51 |
| FFmpeg(原生分 ↑) | -OFF 早停待重跑 | +OFF 时长较短或早停 | 2.7086(16.17 · 18h01m · 16 次) | 1.9929(11.19 · 早停 · 6 次) | -+0.7157(+35.9%, +4.98 分) | ++0.7157 / +4.98 | ||
| VLIW(周期数 ↓) | -OFF 14h45m 历史 | +OFF 时长较短或早停 | 1,042(97.25 · 17h57m · 14 次) | 1,097(93.82 · 14h45m · 12 次) | -−55 周期(−5.0%, +3.43 分) | +−55 / +3.43 | ||
| Ad(原生总分 ↑) | -OFF 8h47m 空挂 | +OFF 时长较短或早停 | 49,673,040,827(79.30 · 18h00m · 16 次) | 49,631,904,944(78.33 · 8h47m · 7 次) | -+41.14M(+0.97 分) | ++41,135,883 / +0.97 | ||
| Codec(原生分 ↑) | -OFF 12h38m 反例 | +OFF 时长较短或早停 · OFF 占先 | 68.89(68.89 · 18h00m · 13 次) | 70.79(70.79 · 12h38m · 11 次) | -−1.90 分(OFF 胜) | +−1.9 / −1.9 | ||
| OpenTTD(原生分 ↑) | -两臂均未满 反例 | +两臂均未满 18 小时 · OFF 占先 | 30,266,729(62.02 · 10h47m · 10 次) | 33,090,430(62.99 · 早停 · 7 次) | -−2.82M(−0.97 分, OFF 胜) | +−2,823,701 / −0.97 |
实验口径与边界说明:以上属于 C1 多轮探索观测(有效运行时间按求解进程扣除停机后累计)。本节固定 Explore + Stride 框架,仅隔离检验开启上下文交付(Context ON vs Context OFF)在组内的方向性表现;因存在续跑会话与部分未补满时长差异,不单独构成剥离全部交互项后的纯 C2 因果证明。控制面整体相对零协议税外部基线的等预算帕累托对比属另一层独立议题,详见 《EdgeBench 反馈与记忆》中的第二层目标讨论。原始聚合结构化数据见 公开聚合数据。
原来的公开汇总快照保留了比分和时长记录。结构化数据列明取舍与出处。全系统在等预算下的比较另见EdgeBench 反馈与记忆。
真实模型与多题消融实证:收益链路与边界反例
在 EdgeBench ANN 向量检索优化(gpt-6.1-sol / xhigh,18h,约束 Recall@10 ≥ 0.95)与跨题消融中,我们将 Context Delivery(渐进读取与周期差分)与 Stride(基于证据差分的步长控制)拆开对照,同时统计控制包选中字节、主会话模型输入 Token 与局部/早窗反例(基于 @shangzh0 归档轨迹与控制面回执复算):
多题消融与分层计量审计
Recall@10 ≥ 0.95)。下图(B):相对 Context off 的分层字节与主会话输入 Token 节省(左,负值)及诚实披露的局部/早窗开销与跨题反例(右,正值)。点击查看原图。Context delivery · 机制与实例
· 11 月 14 日更新
Agent 读上下文,也需要一套“查书”的方法
拆解十题同构两臂(固定 Explore + Stride)下的 Context ON vs OFF 跑分对比(7/10 题占先、Borden 80 vs 45、ANN 7,159 vs 3,746,并保留 Portfolio/Codec/OpenTTD 反例),同时核对渐进视图 32.5%、ANN 周期内增量轮 −74.25% 与首读 3.0% / 5.0% 的载荷基线。
阅读全文Agent 读上下文,也需要一套“查书”的方法
核对 EdgeBench 的十组记录分数,其中四组的时长接近 18 小时,其余涉及历史运行或未跑满的记录。文章也分别交代渐进视图 32.5%、ANN 周期内增量轮缩减 74.25% 和首读 3.0% / 5.0% 各自的测量基线。
阅读全文研究札记 · EdgeBench 与 LHTB