← J++ 案例 / J++ cases

Harness 的世界A world compiled from the Harness ledger

这是把一个真实运行了三个月的 Agent 系统——Harness,约 89 万条事件——的账本,用 J++ 编译成可看的三维世界的第一批骨架。J++ 在里面做的事:生成器为每类对象提候选,JEV 逐个对象挑身体;为每个任务挑结构、材料、颜色、顶;为每一步挑工序。数量(规模、层数、开窗)由代码照实算,物理先决条件由世界强制。五页都是真机运行:判断次数与费用在页面右下角。

This compiles the ledger of a real Agent system that ran for three months — Harness, about 890,000 events — into a viewable three-dimensional world with J++. What J++ does inside it: a generator proposes candidates for each object class, and JEV picks a body for each object, then picks structure, material, color and roof for each task, then picks a procedure for each step. Quantities (scale, floor count, window openings) are computed by code as recorded; physical preconditions are enforced by the world itself. All five pages are live runs: judgment counts and cost sit in the bottom-right corner of each page.

01

一个夏天的 48 座48 buildings across one summer

建设链最长的 48 个真实任务,跨 7–9 月;账本里只有 1 座被判成封顶,其余 47 座没有封顶记录,照实呈现。The 48 real tasks with the longest build chains, spanning July–September; the ledger records only 1 building as topped out, the other 47 have no topping-out record, shown as recorded.

  • 2,045 次判断、82 次生成,约 $0.07 / 2,045 judgments, 82 generations, about $0.07
  • 251 次顺序纠正 / 251 sequence corrections
  • 账本里 1 座判成封顶,47 座没有封顶记录 / the ledger records 1 building topped out, 47 with no record of it
打开 · Open ↗
02

五座在建的东西Five structures under construction

2026-09-07 一天里 5 个任务、258 条事件合成 53 道工序,JEV 从候选里挑结构与工序:三联塔判到 1.00,判据桥判到 0.98。One day, 2026-09-07: 5 tasks and 258 events compiled into 53 procedure steps, with JEV picking structure and procedure from candidates — 三联塔 (a triple-tower structure) scored 1.00, 判据桥 (a criterion-bridge structure) scored 0.98.

  • JEV 选结构:三联塔 1.00、判据桥 0.98… / JEV picked structures: 三联塔 (triple tower) 1.00, 判据桥 (criterion bridge) 0.98…
  • 53 道工序 / 53 procedure steps
  • 20 次顺序纠正 / 20 sequence corrections
打开 · Open ↗
03

一次修复One fix

一个修复席 15:45–15:54 的真实事件链,60 秒里 6 次判断,5 次命中缓存,只有 1 次真花钱:$0.0000481。A real event chain from one fix session, 15:45–15:54: 6 judgments in 60 seconds, 5 served from cache, only 1 actually cost money — $0.0000481.

  • 60 秒,6 次判断 / 60 seconds, 6 judgments
  • 5 次命中缓存 / 5 cache hits
  • 1 次真调用:$0.0000481 / 1 live call: $0.0000481
打开 · Open ↗
04

放大进一座楼Zoom into one building

一沙一世界:同一镜头从 48 座的城一路放大——外墙连续让开,露出楼里 441 间房(一个会话一间),房里 1,713 件东西(一条步骤一件),再放大到一件东西的原文与 JEV 对这一步的题和读数。原文里的任务代号是门,点开飞向那座楼。A world in a grain of sand: one continuous camera push from the 48-building city — walls part in sequence to reveal 441 rooms (one room per session), holding 1,713 items across them (one per step), zooming down to one item's original text and the question JEV was asked about that step, with its reading. Task codes in the text are doors; click one to fly to its building.

  • 441 间房、1,713 件东西 / 441 rooms, 1,713 items
  • ?tour=1 自动示范 38 秒 / ?tour=1 runs a 38-second auto-demo
  • 层次全部由数据给出 / Every level comes from the data
打开 · Open ↗
05

港与城Harbor and city

一个夏天:编排器 orchestrator.py 是港城——391 个函数按行数定大小,调用铺成街,30 个片区由 JEV 判角色。16,117 次派发是船,JEV 判性质:卡滞重派 7,216(45%)、闸门拦截 2,289、接力推进 2,180、例行心跳 1,566、会话收尾 1,418、送信通知 665、异常告警 645、人工触发 135;真正起了会话的 2,701 艘驶向它任务的楼。7/19 轮询守护进程停下:灯塔暗、闸门放下、6 盏灯留在闸口,此后船从接力水道出港。One summer: the orchestrator, orchestrator.py, rendered as a harbor city — 391 functions sized by line count, calls paved into streets, JEV judging the role of 30 districts. 16,117 dispatches are ships; JEV judges what each one is: stall-retry 7,216 (45%), gate-block 2,289, relay-advance 2,180, routine heartbeat 1,566, session close-out 1,418, message delivery 665, alarm 645, manual trigger 135. Of these, 2,701 actually opened a session and sailed to their task's building. When the polling daemon stopped on 7/19: the lighthouse went dark, the gate dropped, 6 lights stayed at the gate, and ships have sailed out through the relay channel since.

  • 2,701 艘真正起了会话 / 2,701 ships actually opened a session
  • 6–9 月出港/到楼:489/24、896/83、82/0、1,234/150 / Monthly ships out / arrivals, Jun–Sep: 489/24, 896/83, 82/0, 1,234/150
  • 235 秒 / 235 seconds
打开 · Open ↗

判断分工的两次实证Two tests of what should be judged versus decided by code

把只有闸门检查项的 CommitAccepted 交给 JEV 判「修根因还是绕过」,1,460 条里 1,379 条给出同一个答案。这批材料本身只有闸门检查字段,没有给判断留下可区分的证据。

Handing CommitAccepted events that only carry gate-check fields to JEV, asking it to judge "fixed the root cause or worked around it": 1,379 of 1,460 came back with the same answer. The material handed to JEV carried only gate-check fields — no evidence that could tell the two answers apart.

把颜色、顶当抽象选项直接问 JEV,48 座里 41 座选了青灰。修法:代码能定的用代码判;需要 JEV 判的,让生成器先为每个候选写一句「什么样的任务适合它」,判断器才有证据可依据,而不是在裸选项里凭空选。

Asking JEV to pick color and roof as bare abstract options, 41 of 48 buildings came back the same blue-gray. The fix: let code decide whatever code can decide; where JEV must judge, have the generator first write one sentence per candidate on what kind of task it fits, so the judge has evidence to work from instead of choosing among bare labels.

程序The programs

三个 J++ 程序各不到 60 行:compile.jpp、compile-steps.jpp、compile-build-v2.jpp;港城的片区角色与船的性质由 compile-harbor.jpp 判,不到 40 行。程序源码在 world 仓库里,暂未随本仓库发布,随下一次同步一并发布;这次先发运行产物——五段录制好的三维世界。

The three J++ programs are each under 60 lines: compile.jpp, compile-steps.jpp, compile-build-v2.jpp; the harbor's district roles and each ship's nature are judged by compile-harbor.jpp, under 40 lines. Source lives in the world repository and is not yet published here; it ships with the next sync. This drop is the recorded worlds themselves.

建议用桌面 Chrome 打开;空格键重放,拖动画面可以接管镜头。
Best viewed in desktop Chrome. Press space to replay; drag to take over the camera.

数据来自 Nature 自己的 Harness 账本;页面上的任务名、事件原文是真实记录。

Data comes from Nature's own Harness ledger; the task names and event text shown are real records.