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
@@ -111,6 +111,101 @@
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.
+
+
Ten-task tally (7 ahead or slight edge vs 3 OFF ahead): under fixed Explore + Stride, enabling Context Delivery puts Explore + context-on ahead of or slightly ahead of context-off on 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 2,311 cost, −114)—while explicitly preserving all 3 counterexamples where context-off scores higher.
+
Strictly duration-matched ~18h pairs (4 tasks within the 15-minute tolerance): 3 wins to 1—Borden rises from 45 to 80 (+35 pts / +77.8%); Grid total cost (lower is better) improves from 2,311 to 2,197 (−114 / −4.9%, +2.35 normalized pts); Rust (both 17h57m) shows a slight native edge for context-on (27.1801 vs 27.1757); the matched 18h counterexample is Portfolio (both 18h00m, 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-on reaches 7,159 QPS vs context-off3,746 QPS (+91.1%, +33.51 normalized pts); FFmpeg reaches 2.7086 vs 1.9929 (+35.9%, OFF stopped early pending rerun); VLIW cycles (lower is better) improve to 1,042 vs 1,097 (−55 cycles / −5.0%, OFF historical 14h45m); Ad reaches 49.67B vs 49.63B (ON 18h00m vs OFF 8h47m host-startup hang). The two remaining counterexamples are Codec (ON full 18h00m 68.89 vs OFF 70.79 in 12h38m despite a model-service block) and OpenTTD (both Explore arms <18h, ON 30,266,729 vs OFF 33,090,430).
+
+
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)
A: Nine tasks with normalized EdgeBench scores (/100, direction-normalized so higher is better), ordered by 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
+
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.
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):
Read progressive views and compaction-cycle deltas with their own measurement baselines: 32.5% across 213 wakes, −74.25% on ANN in-cycle wakes, and ≈ 3.0% / 5.0% eight-task first-read entry. Compaction resets the complete baseline; these cohorts are separate comparisons.
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.
diff --git a/apps/presentation/site/scripts/generate-long-horizon-audit.py b/apps/presentation/site/scripts/generate-long-horizon-audit.py
index 4acda82a8c..48d837c18a 100644
--- a/apps/presentation/site/scripts/generate-long-horizon-audit.py
+++ b/apps/presentation/site/scripts/generate-long-horizon-audit.py
@@ -323,3 +323,166 @@ def save(fig, relative):
"v2: 372 records, 249 reported transitions; 7 outcomes unreconciled from published counts. Latest: 25 reviews, 17 scored transitions.\nRow counts: improved / flat / reported regression / unreconciled. Published aggregates do not establish causal benefit.")
fig.text(.03,.025,footer,fontsize=10,color="#475569",linespacing=1.6)
save(fig, f"images/edgebench-feedback-and-memory/stride-yardstick-closed-loop-{language}.svg")
+
+ # 4. EdgeBench Controlled Two-Arm Explore + Stride: context-on vs context-off score comparison (10 tasks)
+ two_arm = DATA["stride_context_on_off_comparison"]
+ arm_tasks = {row["task"]: row for row in two_arm["tasks"]}
+ ordered_norm_names = ["ANN", "FFmpeg", "VLIW", "Grid", "Ad", "Rust", "OpenTTD", "Codec", "Portfolio"]
+ norm_rows = [arm_tasks[name] for name in ordered_norm_names]
+
+ fig = plt.figure(figsize=(14.6, 7.8))
+ gs = fig.add_gridspec(
+ 1, 2, width_ratios=[1.25, 1.05],
+ left=.18, right=.98, bottom=.175, top=.805, wspace=.11,
+ )
+ left = fig.add_subplot(gs[0, 0])
+ right = fig.add_subplot(gs[0, 1])
+
+ y_pos = list(range(len(norm_rows)))
+ on_vals = [r["context_on_norm"] for r in norm_rows]
+ off_vals = [r["context_off_norm"] for r in norm_rows]
+ off_bar_color = "#94a3b8"
+
+ left.barh([y - .18 for y in y_pos], on_vals, height=.32,
+ color=TEAL, label=("Explore + Stride + context-on" if not zh else "Explore + Stride + context-on(开启上下文交付)"))
+ left.barh([y + .18 for y in y_pos], off_vals, height=.32,
+ color=off_bar_color, label=("Explore + Stride + context-off" if not zh else "Explore + Stride + context-off(关闭上下文交付)"))
+
+ tag_map_zh = {
+ "ANN": "ANN (QPS↑ · 历史7臂Stride)",
+ "FFmpeg": "FFmpeg (原生↑ · OFF早停)",
+ "VLIW": "VLIW (周期↓ · OFF 14h45m)",
+ "Grid": "Grid (总分↓ · 等时长 ~18h)",
+ "Ad": "Ad (总分↑ · OFF 8h47m)",
+ "Rust": "Rust (原生↑ · 等时长 ~18h)",
+ "OpenTTD": "OpenTTD (原生↑ · 均<18h)",
+ "Codec": "Codec (原生↑ · OFF 12h38m)",
+ "Portfolio": "Portfolio (分↑ · 等时长 18h)",
+ }
+ tag_map_en = {
+ "ANN": "ANN (QPS↑ · 7-arm Stride)",
+ "FFmpeg": "FFmpeg (score↑ · OFF early stop)",
+ "VLIW": "VLIW (cycles↓ · OFF 14h45m)",
+ "Grid": "Grid (cost↓ · ~18h matched)",
+ "Ad": "Ad (score↑ · OFF 8h47m)",
+ "Rust": "Rust (score↑ · ~18h matched)",
+ "OpenTTD": "OpenTTD (score↑ · both <18h)",
+ "Codec": "Codec (score↑ · OFF 12h38m)",
+ "Portfolio": "Portfolio (score↑ · 18h matched)",
+ }
+ left.set_yticks(y_pos, [(tag_map_zh if zh else tag_map_en)[r["task"]] for r in norm_rows], fontsize=9.7)
+ left.invert_yaxis()
+ left.set_xlim(0, 134)
+ left.set_xticks([0, 25, 50, 75, 100])
+ left.set_xlabel("EdgeBench 百分制归一分 (/100,越高越优;含 ↓ 指标已方向归一化)" if zh else "EdgeBench normalized score (/100, higher is better)", fontsize=9.5)
+ left.set_title(
+ "A. 固定 Stride 下 context-on vs context-off 百分制对照" if zh else "A. Fixed Stride: context-on vs context-off (/100)",
+ loc="left", pad=14, fontsize=11.0, fontweight="bold",
+ )
+ left.set_axisbelow(True)
+ left.grid(axis="x", color="#e2e8f0", linewidth=.7)
+ left.tick_params(axis="both", length=0)
+
+ for idx, r in enumerate(norm_rows):
+ diff = r["context_on_norm"] - r["context_off_norm"]
+ anchor = max(r["context_on_norm"], r["context_off_norm"])
+ if r["task"] == "Rust":
+ txt = f"{r['context_on_norm']:.2f} vs {r['context_off_norm']:.2f} (+0.0044)"
+ col = TEAL
+ fw = "bold"
+ elif diff > 0:
+ txt = f"+{diff:.2f} ({r['context_on_norm']:.2f} vs {r['context_off_norm']:.2f})"
+ col = TEAL
+ fw = "bold"
+ else:
+ txt = f"{diff:.2f} ({r['context_on_norm']:.2f} vs {r['context_off_norm']:.2f})"
+ col = "#b45309"
+ fw = "bold"
+ left.text(anchor + 2.2, idx, txt, va="center", fontsize=8.8, fontweight=fw, color=col)
+
+ right.axis("off")
+ right.set_title(
+ "B. 原生关键差值(含 Borden)与反例边界" if zh else "B. Native score deltas (Borden) & counterexamples",
+ loc="left", pad=14, fontsize=11.0, fontweight="bold",
+ )
+ borden = arm_tasks["Borden"]
+ cards_two_arm = (
+ [
+ (
+ "第 10 题(无 /100 标度):Borden 等时长 ~18h\n"
+ f"• context-on(18h02m):{borden['context_on_native']}(17 次有效)\n"
+ f"• context-off(17h59m):{borden['context_off_native']}(17 次有效)\n"
+ "→ +35 分(+77.8%,同构等时长显著拉升)。",
+ "#f0fdfa", "#99f6e4",
+ ),
+ (
+ "关键原生分数增幅(固定 Explore + Stride)\n"
+ "• ANN (QPS ↑):7,159 vs 3,746(+91.1%, +33.51 百分分)\n"
+ "• FFmpeg (原生 ↑):2.7086 vs 1.9929(+35.9%, OFF 早停)\n"
+ "• VLIW (周期 ↓):1,042 vs 1,097(−55 周期 / −5.0%)\n"
+ "• Grid (代价 ↓):2,197 vs 2,311(−114 / −4.9%, ~18h)",
+ "#f8fafc", "#cbd5e1",
+ ),
+ (
+ "10 题总盘:ON 7 项占先 vs OFF 3 项领先\n"
+ "• 等时长 ~18h(4 题):3 胜(Borden/Grid/Rust 微先),\n"
+ " 1 反例(Portfolio 13.62 vs 39.44)\n"
+ "• 非等时长或历史对照(6 题):ON 领先 4 题,\n"
+ " OFF 领先 2 题(Codec 70.79 / OpenTTD 33.09M)",
+ "#fffbeb", "#fcd34d",
+ ),
+ ]
+ if zh else
+ [
+ (
+ "Task #10 (no /100 scale): Borden (~18h matched)\n"
+ f"• context-on (18h02m): {borden['context_on_native']} (17 valid)\n"
+ f"• context-off (17h59m): {borden['context_off_native']} (17 valid)\n"
+ "→ +35 pts (+77.8%, duration-matched gain).",
+ "#f0fdfa", "#99f6e4",
+ ),
+ (
+ "Key native deltas under fixed Explore + Stride\n"
+ "• ANN (QPS ↑): 7,159 vs 3,746 (+91.1%, +33.51 pts)\n"
+ "• FFmpeg (score ↑): 2.7086 vs 1.9929 (+35.9%, OFF early stop)\n"
+ "• VLIW (cycles ↓): 1,042 vs 1,097 (−55 cycles / −5.0%)\n"
+ "• Grid (cost ↓): 2,197 vs 2,311 (−114 / −4.9%, ~18h)",
+ "#f8fafc", "#cbd5e1",
+ ),
+ (
+ "10-task summary: 7 ON ahead/edge vs 3 OFF ahead\n"
+ "• Matched ~18h (4): 3 ON wins (Borden/Grid/Rust),\n"
+ " 1 counterexample (Portfolio 13.62 vs 39.44)\n"
+ "• Historical / <18h (6): 4 ON ahead, 2 OFF ahead\n"
+ " (Codec 68.89 vs 70.79; OpenTTD 30.27M vs 33.09M)",
+ "#fffbeb", "#fcd34d",
+ ),
+ ]
+ )
+ y_starts = [.98, .63, .28]
+ for y_box, (txt, face, edge) in zip(y_starts, cards_two_arm):
+ right.text(.01, y_box, txt, transform=right.transAxes, va="top", fontsize=9.1,
+ linespacing=1.50, bbox={"boxstyle": "round,pad=.60", "facecolor": face, "edgecolor": edge})
+
+ fig.suptitle(
+ "EdgeBench 十题同构效果对照:固定 Explore + Stride 下的 Context ON vs OFF"
+ if zh else
+ "EdgeBench 10-task controlled comparison: Context ON vs OFF under Explore + Stride",
+ fontsize=14.6, fontweight="bold", y=.955,
+ )
+ sub_two_arm = (
+ "固定 Explore + Stride 框架(context-on 开启 Progressive 与周期内差分)。10 题中 context-on 7 题占先,完整保留 3 项反例与真实时长差。"
+ if zh else
+ "Holding Explore + Stride constant (context-on adds progressive reads + cycle deltas). Ahead on 7/10 tasks, retaining 3 counterexamples and runtime gaps."
+ )
+ fig.text(.5, .88, sub_two_arm, ha="center", fontsize=9.9, color="#475569")
+ fig.legend(*left.get_legend_handles_labels(), loc="lower center", ncol=2, frameon=False, bbox_to_anchor=(.45, .078), fontsize=9.6)
+ footer_two_arm = (
+ "口径说明:截取自 EdgeBench 同构两臂实测(2026-11-12 22:40 CST 快照)。Borden 无百分制换算故在右栏 B 单列原生分(80 vs 45)。\n"
+ "Grid、Borden、Rust、Portfolio 四题双臂均已跑满约 18h(15m 容差内);ANN 取自历史七臂对应 Stride 配置;其余 5 题存在至少一臂未满 18h。本图仅检验开关的组内方向性效果。"
+ if zh else
+ "Basis: EdgeBench controlled two-arm snapshot at 2026-11-12 22:40 CST. Borden has no /100 scale and appears in native points in Panel B (80 vs 45).\n"
+ "Grid, Borden, Rust, and Portfolio are ~18h duration-matched (within 15m); ANN uses the matching Stride pair from the 7-arm study; remaining 5 pairs have <18h runs. Checks within-harness direction only."
+ )
+ fig.text(.03, .018, footer_two_arm, fontsize=8.9, color="#475569", linespacing=1.45)
+ save(fig, f"images/context-delivery/stride-context-on-vs-off-{language}.svg")
diff --git a/examples/blog-bilingual-index-smoke.mjs b/examples/blog-bilingual-index-smoke.mjs
index d99c731653..1c2235fdb6 100644
--- a/examples/blog-bilingual-index-smoke.mjs
+++ b/examples/blog-bilingual-index-smoke.mjs
@@ -132,6 +132,52 @@ export async function validateBilingualBlog(blogDir) {
"utf8",
);
assertIncludes(multitaskSvg, "100.0%", `Eight-task byte audit chart must anchor at 100% Full: ${langSuffix}`);
+ const twoArm = audit.stride_context_on_off_comparison;
+ deepStrictEqual(
+ [
+ twoArm?.summary?.total_tasks,
+ twoArm?.summary?.context_on_ahead_or_edge,
+ twoArm?.summary?.context_off_ahead,
+ twoArm?.tasks?.length,
+ ],
+ [10, 7, 3, 10],
+ "Controlled two-arm comparison summary must match the 10-task tally (7 ahead/edge, 3 OFF ahead)",
+ );
+ assertIncludes(
+ article,
+ '
',
+ `Both published editions must expose the controlled Context ON vs OFF score section: ${locale}`,
+ );
+ assertIncludes(
+ article,
+ `stride-context-on-vs-off-${langSuffix}.svg`,
+ `Both published editions must embed the controlled Context ON vs OFF chart: ${langSuffix}`,
+ );
+ const twoArmSvg = await readFile(
+ resolve(blogDir, `images/context-delivery/stride-context-on-vs-off-${langSuffix}.svg`),
+ "utf8",
+ );
+ const twoArmSection = article.split('