Epistemic Fold
Epistemic Fold for DeepSeek Harness
A contract-preserving context runtime for long-horizon agents.
面向长周期 Agent 的、保持契约的上下文运行时。
What is this? · 这是什么？
Epistemic Fold (EF) is a compaction backend plugin for the
DeepSeek Harness that treats
conversation history, memory, and working context as three different things:
History ≠ Memory ≠ Context
Its core principle:
An agent may fold a trajectory out of the working context only after its
externally relevant epistemic effects have been materialized, preserved, and
made recoverable.
EF 是 DeepSeek Harness
的一个 compaction backend 插件。它把会话历史、记忆与工作上下文当作三件不同
的事，核心命题：
只有当一段轨迹对未来仍有意义的知识效应已经被稳定对象吸收、关键状态与
边界得到保留、且原始证据可以恢复时，该轨迹才有资格被折叠出工作上下文。
Concretely, when the runtime folds a span of conversation it ·
运行时折叠一段会话时：
- archives the exact model-visible messages into an immutable,
hash-verified
CheckpointBundle before any lossy surface replacement
commits (BundleDurable ≺ SurfaceLoss) —
先归档精确的模型可见消息到不可变、哈希校验的 CheckpointBundle，
在任何有损 surface 替换提交之前；
- never re-folds what is already frozen — a monotonically advancing
Fold Frontier separates frozen checkpoints from the open trajectory, so
the cached prefix stays byte-stable across folds —
绝不重折叠已冻结的内容 —— 单调前进的 Fold Frontier 分隔冻结
checkpoint 与开放轨迹，缓存的 prefix 在多次折叠间逐字节稳定；
- derives the current state deterministically from raw session events —
objectives, constraints, decisions, values, evidence, failures, and
obligations, with full provenance and a hard rule that narrative summaries
can never verify state (
Raw Events → State and Raw Events → Summary
run in parallel; Raw → Summary → State is forbidden) —
当前状态由原始 session 事件确定性派生，全部携带完整 provenance；
语义摘要永远不能验证状态（Raw → Summary → State 被禁止）；
- recovers exactly —
context_search / context_recall serve bounded,
paginated, provenance-checked recall of anything that left the working
set —
精确恢复 —— context_search / context_recall 提供有界、分页、
经 provenance 校验的召回。
The optimization goal is not maximal compression. It is correctness first —
then information density, prefix-cache locality, long-horizon state
consistency, and stable closed-loop continuation.
优化目标不是最大压缩率，而是 correctness first —— 在硬正确性约束下提高信息
密度、prefix cache 局部性、长期状态一致性与闭环延续稳定性。
Status · 状态
Research implementation against DeepSeek Harness 0.1.7-rc.2 (verified
baseline 477b4f420553e8a52c2fbccc464d7561b239c443) · 对照 DSH
0.1.7-rc.2 核验基线的研究性实现。
| Milestone 里程碑 | Scope 范围 | Gate |
|---|
| M0 | Exact archive / recall closure, atomic bundle store, deterministic fallback · 精确归档/召回闭环、原子 bundle store、确定性 fallback | ✅ C0.1–C0.5 |
| M2 | Fold Frontier, leaf/root folds, prefix fingerprint proofs · Fold Frontier、leaf/root 折叠、prefix 指纹证明 | ✅ C2.1–C2.4 |
| M3a | Deterministic current state (anchors, authority, projection) · 确定性当前状态 | ✅ ALR=0 · SSR=0 · provenance=100% · bounded |
| R0-A | Cross-layer integration closure: marker protocol, frontier hard invariant, recall isolation, commit records, authority gate, disjoint rendering · 跨层组合正确性闭合 | ✅ 65 tests |
| R0-B | Native plugin entry, package manifest, composition smoke (mount/fold/restart), CI lanes · 原生插件入口 + 组装冒烟 + CI | ✅ |
| R0-C | Corrected metrics, keyless boundary corpus (12 hard + 6 exploratory), paired runner (B1/E3a0/E3aR), long-horizon economics 32/64/128 · 指标修正 + 边界语料 + 配对续跑 + 长程经济学 | ✅ 115 tests · report |
| R1-A | Token source attribution, versioned economics profiles (asOf + source, user-overridable), provider cache telemetry · 词元来源归因 + 版本化经济模型 + 缓存实况遥测 | ✅ 24 tests |
| R1-B | W1–W5 workloads × 32/64/128, counterfactual ROI lab (Delta / M1 / M5 / Adaptive Root) · 五类负载矩阵 + 反事实 ROI 实验台 | ✅ 14 tests · gate |
| R1-D | Provider-aware amortized rebase policy compiler (break-even horizon, hard overrides win) · 模型感知摊销式 rebase 策略编译器 | ✅ 12 tests |
| R1-E | Pareto frontier over cost / footprint / success, correctness as a filter · 帕累托前沿 + 正确性作为过滤器 | ✅ 9 tests · report |
|
M1 (verified ingress reduction), M3b (negative knowledge / uncertainty), M4
(dependency graph) and beyond are deliberately not implemented — each
requires observed failure evidence from benchmarks first (see
05_EF_DECISIONS_AND_OPEN_QUESTIONS.md,
section F) · M1/M3b/M4 及之后刻意未实现——每项都需要 benchmark 先观察到
对应的失败证据（见 05 文档第 F 节）。
R1 followed the same rule and rejected production Delta Leaf on measured
ROI (3.7–16.6% upper bound) in favour of an amortized root policy ·
R1 遵循同一规则：依据实测 ROI（上限 3.7–16.6%）否决了生产化 Delta Leaf，
改走摊销式 root 策略。
Evaluation evidence (R0-C, keyless + deterministic): EF keeps more history
cache-warm at every horizon, and the leaf/root maintenance loop works. The
honest long-horizon finding is that frozen checkpoints charge a recurring
prompt cost that can exceed Basic's rewrite cost as folds accumulate —
root rebase cuts that load ~45% and is the effective lever. EF's cache
locality converts into cost advantage only while the provider's cache
discount is deep (ρ break-even curve). Full numbers, including the cases where
EF does not win, are in
the R0-C evaluation report ·
评估证据（R0-C，keyless + 确定性）：EF 在每个时间尺度上都保留更多
cache-warm 历史，leaf/root 维护回路工作正常。诚实的长期发现是：
frozen checkpoint 的重复 prompt 成本会随折叠累积，可能超过 Basic 的重写
成本——root rebase 能把该负载降低约 45%，是有效的杠杆。EF 的 cache 局部性
只有在 provider 缓存折扣足够深时才转化为成本优势（ρ break-even 曲线）。
包含 EF 并不占优情形的完整数据见 R0-C 评估报告。
R1 evidence (keyless, deterministic) locates the real cost precisely, and it
is not where the architecture expected: ~87% of a leaf checkpoint node is
repeated framing preamble, and because the frozen prefix is monotonically
non-decreasing, once it alone exceeds the pressure threshold every step
folds — EF folds 51× against Basic's 21× on an identical workload and digest,
a 3.3× total-token gap. With task success held constant, Basic dominates EF on
cost and footprint on every workload; the
R1 report states that plainly and states
equally plainly that it is not a verdict — equal success is the one premise
the keyless tier cannot test, and EF's entire justification is that success is
not equal. No production default is flipped on economics alone ·
R1 证据（keyless、确定性）精确定位了真实成本，而它并不在架构预期之处：
leaf checkpoint 节点约 87% 是重复的 framing 前导文本；且由于 frozen
prefix 单调不减，一旦其自身超过压力阈值，每一步都会折叠——相同负载与
相同摘要文本下 EF 折叠 51 次而 Basic 仅 21 次，总词元相差 3.3 倍。在任务
成功率相同的前提下，Basic 在每个负载上都优于 EF；R1 报告
如实陈述这一点，并同样明确说明这不是最终结论——"成功率相同"正是 keyless
层唯一无法验证的前提，而 EF 的全部理由就是成功率并不相同。任何生产默认值
都不会仅凭经济性而改变。
The live tier then ran, and it did not find a behavioral advantage. 64
paired trials against the configured model scored EF 32/32 vs Basic 31/32 —
one trial, which is not evidence. The null result is reported as null:
at this scale Basic's lossy summary retained every fact the cases probed. Two
things were decisive, though: measured cache realization h = 0.910 (not the
assumed 1.000), making the effective ratio 5.9× worse than the headline ρ;
and a production bug no keyless test could see — failure anchors rendered
without their description, so the model saw that something was unresolved but
not what. Full detail in
the live results ·
实模型层随后运行，并未发现行为优势。 对当前配置模型完成 64 组配对试验，
结果为 EF 32/32 对 Basic 31/32——仅差一次，不构成证据。零结果即如实报告
为零结果：在该规模下，Basic 的有损摘要保留了全部被测事实。但有两项结论是明确
的：实测缓存命中率 h = 0.910（而非假定的 1.000），使有效价格比比标称 ρ
差 5.9 倍；以及一个keyless 测试完全无法发现的真实缺陷——failure anchor
渲染时丢失了描述，模型只能看到"有未解决项"却看不到"是什么"。详见
实模型结果。
R2 then attacked the price side, and did not finish. The fold-every-step
loop is fixed (folds 50 → 24, peak context −68%) and EF's cost excess over
Basic was cut from BCR 1.276 to 1.149 — but the R2 objective was BCR < 1,
and 0 of 40 measured cells are cheaper than Basic. The cause is now precise
rather than suspected: EF's cold-token overhead is nearly zero, and its entire
disadvantage is that it carries 1.5×–85× more warm tokens, because the
frontier invariant forbids re-folding a frozen checkpoint while Basic re-folds
its single one. No rebase cadence reaches the objective (best: 1.096). Full
detail in the R2 report ·
R2 随后转向价格侧，但未能完成。 逐步折叠循环已修复（折叠次数 50 → 24，
峰值上下文 −68%），相对 Basic 的成本超额从 BCR 1.276 降到 1.149——但 R2
的目标是 BCR < 1，而 40 个实测单元中 0 个比 Basic 便宜。原因现已精确定位
而非猜测：EF 的 cold token 开销几乎为零，全部劣势在于它携带多出 1.5–85 倍的
warm token，因为 frontier 不变量禁止重折叠已冻结的 checkpoint，而 Basic 会
重折叠它唯一的那个。任何 rebase 节奏都达不到目标（最好为 1.096）。详见
R2 报告。
R3–RC0 closed the price gap, then found the shipped defaults were not
calibrated for it. R3 removed the framing tax (BCR 1.149 → 0.986 on the
economy workloads) and R4 built the real-recall workload, realized billing, and
the presets — reaching all five gate components satisfied. RC0 then hardened the
release contract and changed the headline conclusion: at the shipped
defaults the full-task realized ratio has a median of 0.96 but a CI upper bound
of 1.353, so the gate is OPEN because of dispersion, not because the typical
run is dearer. The default flip stayed blocked by design ·
R3–RC0 收窄了价格差距，随后发现出厂默认值并未针对它校准。 R3 消除了 framing
税（经济负载上 BCR 1.149 → 0.986），R4 建成了真实召回负载、实测计费与预设，
五个 gate 组件全部满足。RC0 随后加固发布契约，并改写了主结论：在出厂默认值
下全任务实测比中位数为 0.96，但置信区间上界 1.353——gate 因离散度而 OPEN，
并非因为典型运行更贵。默认翻转按设计保持阻断。
RC1 then normalized the policy instead of adding mechanism. It named the
binding constraint (thresholdRatio: 0.8 reads as an 80% trigger; the shipped
reserve makes the real trigger 49.6%), measured the safety reserve
(9,733 tokens needed vs 65,536 shipped — 6.73×), moved parameter search
off the API and onto a replay simulator that calls the production policy
functions, exonerated the frozen budget by measurement (0 of 135 shadow
evaluations were profitable-but-blocked), and classified the RC0 outlier as
provider cache state — at comparable prompt sizes the two request shapes
cache identically (reuse ratio 1.000). Running it live exposed two instrument
defects larger than the experiment itself: the live endpoint accepts a
top-level system field, returns HTTP 200, and silently ignores it (so every
earlier live measurement that relied on system-prompt assembly was not sending
one — which matters most for R3's framing change, whose entire saving is earned
there), and the paired driver shared provider cache between runs and arms,
so its null test read 1.606 where it must read ~1. With both fixed the null test
behaves (1.021) and the noise floor is 1.005 — but the cost verdict still does
not reproduce at n=8 (0.965 vs 1.104), so the gate stays OPEN for dispersion
on a sound instrument. The deliverable is a certified operating profile for
the measured route, not a global default: an uncertified route falls back to
legacy, never to economy.
Full detail in the RC1 report ·
RC1 随后选择规范化策略而非增加机制。 它命名了真正的约束（thresholdRatio: 0.8 读作 80% 触发，而出厂 reserve 使真实触发点落在 49.6%），实测了安全
reserve（需要 9,733 token，出厂 65,536，相差 6.73 倍），把参数搜索从
API 搬到调用生产策略函数的 replay 模拟器上，用实测为 frozen budget 免责
（135 次 shadow 评估中 0 次「有利可图却被阻断」），并把 RC0 的离群值归类为
provider 缓存状态——在可比 prompt 长度下两种请求形状缓存表现完全一致
（复用比 1.000）。实机运行还暴露了一个比实验本身更重要的缺陷：实机端点接受
顶层 system 字段、返回 HTTP 200、却静默忽略它，因此此前所有依赖 system
prompt 组装的实机测量都没有真正发送 system prompt——这对 R3 的 framing 改动
影响最大，因为该改动的全部收益正来自那里。最终交付物是针对已测路由的经济模式
认证档案，而非全局默认值：未认证路由回退到 legacy，绝不回退到 economy。
详见 RC1 报告。
RC1.1 then reconciled the evidence, and withdrew one RC1 conclusion. Three
defects did not survive audit: the certified profile reported certified: true
while the cost gate was OPEN (a cache reuse ratio was standing in for a
price measurement), a ~30K-scale safety estimate was reported as a production
headroom, and the replay varied the trigger and the retention together — which is
what produced RC1-H's "the shipped reserve helps" conclusion. Holding retention
fixed reverses the sign, so that conclusion is withdrawn. With both confounds
removed the region Ω is wide (10 robust cells, 7 below parity across all cache
scenarios) and the mature production values sit in its widest band, but the live
cost verdict still does not reproduce at n=8. The live smoke also measured what looked like a
product boundary: the economy preset preserves declared state across folds and
its marker-only checkpoint has no prose in it.
The project state is now Economy mechanics certified; route-level cost
recommendation still open, with no new architecture added to flip a default.
RC1.2 then closed that boundary question, and corrected it. RC1.1's smoke
never executed a tool call, so it could only show that the surface lacks
undeclared prose — not that EF cannot recover it. With a real agent loop
(model → tool-call → ToolRuntime → tool/result → next model step) and a
deterministic keyless proof, PARSE-7741 is absent from the surface and present
in what context_search → context_recall returns, under both semantic modes.
So semanticMode does not gate recall, and the correct contract is declared
state is hot; undeclared history is recoverable through bounded recall. The
preset keeps semanticMode: 'none' — not because rationale is unaffordable (it
costs +6.4%, measured against its real ~7.4K-token call rather than the 512 the
simulator had assumed) but because nothing requires it.
RC1.2.1 then corrected the comparison itself. The three-arm live smoke had
passed plugin: true unconditionally, and since the harness returns early on
that, its engine: 'basic' spread was dead code — the "Basic" arm was really
EF, which is why it reported facts retrievable 2/5 on a metric undefined for
real Basic. The harness now refuses that combination outright, and the
corrected comparison was: true Basic scores 3.00/3 with zero tool calls, while
economy-none scores 2.40/3 with a tool returning the facts in 5/5 runs — so the
gap is retrieval policy and model tool use, not storage or mechanism.
RC1.3 then closed that gap, and it was retrieval ergonomics on both sides.
An un-hinted baseline (no "use context_search" in the probe) plus a per-fact
failure taxonomy found the only failure mode was no-search — never a
recall-miss, so no product-side retrieval defect — and two concrete defects
underneath it. First, context_search returned a 240-character excerpt centred
on the match, which cut a 230-character message's other facts off the front:
the model searched for one value, got an excerpt naming its neighbours, and
answered the rest unknown. A hit was manufacturing a search-miss. Short
messages now come back whole, and hits report matchKind, a verbatim excerpt, and
the context_recall page that holds the match. Second, the test filler numbered
its units in decimal, so unit 64 and unit 90 satisfied the fact matchers from
retained filler — a confound that also affected RC1.2's numbers. With both fixed
the baseline read 2.33/3, and one measured change to the framing section — search
the folded history before answering or reporting something unknown — took it to
3.00/3, matching Basic, reproduced in two independent n=9 runs with search in
9/9. The hint is now worth 0.00: a deployment no longer needs to tell the
agent how to use its own tools. Economy remains a retrieval-dependent low-cost
mode (still ~17× cheaper on the n=5 sample) and is not promoted to a generic
default replacement.
RC1.3.1 then froze the retrieval layer. RC1.3's probe put every fact in ONE
message of ONE checkpoint — the easy case, and one that hid three chronological
defects which appear only in a long task. First, list() returns bundles in
ascending createdAt and search applied its limit mid-scan, so the checkpoints
dropped were the newest: the current value was the first thing lost. Second,
locate returned the first matching message, so a superseded value shadowed
its own correction. Third, exactPageOffset was page-aligned for the default page
size, so it only "contained the match" for a caller that happened to use that
size. Each makes a model read an old value as current — the most damaging
class of error for an agent working from folded history, because the answer looks
well-sourced. Now hits are ordered newest-first by conversation position
(source.orderedSurfaceSeqs, not a clock that can run backwards or collide), so
limit means "the most recent N"; locate returns the latest match and
reports matchCount (above 1 means the fact has history) plus
earliestMatchedMessageIndex (the superseded value stays precisely reachable);
and matchedMessageIndex replaces the aligned offset, valid at any page size.
20 keyless tests pin this, and the live suite was re-run to confirm no regression
(still 3.00/3, searching 9/9). The retrieval layer is now frozen — compression
architecture, recall correctness, and retrieval ergonomics are all CLOSED, leaving
only the route-level realized cost gate OPEN, which affects whether economy is
recommended or defaulted and must not drive the architecture further. Full detail
in the RC1.3.1 report ·
the RC1.3 report ·
the RC1.2 report ·
RC2 then built the product surface, and honestly failed to rank its own tiers.
Three modes now ship — economy, balanced, quality — as a ladder that varies
ONE lever per rung (the semantic face, then retention), with everything else held
identical so a difference is attributable. economy is the rung with measured
end-to-end evidence (RC1.3); the two steadiness rungs are marked HYPOTHESIS on
every surface that describes them, because paying for steadiness is a claim this
project has not measured. /context status registers on the real DSH command
plane, so it costs zero model tokens, and it labels every figure measured or
estimated — an unknown figure prints unknown, never 0. A real-task
comparison then ran 24 times across the four modes over real filesystem and
execution tools, with clean transport (0/261 provider calls lost). It did not
discriminate the modes: TaskQuality saturated at 1.00 for every arm, and the one
steadiness loss was not reproducible on re-run (0.75 → 1.00 for the same cell). So
the ladder's benefit stays a hypothesis, and the reason is the instrument, not the
code — the tasks are too easy and n=1 cannot separate a mode effect from model
variance. What the run does establish positively is the fold/recall contract under
restart and model-switch: every arm still produced the artifact to the
corrected spec after folding and resuming.
RC2.1 then corrected the surface and asked the one question that mattered.
Four /context status defects were fixed, two of them the familiar error of a
number claiming more than it measured: an absent reading was reported as a
measured 0 pressure (and 0.0% occupancy computed from it), and the archived
token figure — a chars/4 heuristic — was labelled measured while only the
message count really was. Usage is now read from DSH's own tokenUsage
projection rather than hand-summed from compaction/summary events, which had
missed every ordinary assistant turn; and current checkpoints (on the surface
now) are separated from lifetime folds (ever committed, from the bundle
store), because RC2 reported the first under the second's name. The RC2 legacy
arm was also renamed to what it was — EF legacy, not Basic: it mounted the EF
plugin with the legacy policy, which still folds through the EF engine, writes
Bundles and exposes the recall tools, so "versus Basic" was not supported. A real
Basic arm now exists. The ladder was reordered retention first (the stronger
and cheaper mechanism), and one critical A/B isolated retention alone: across 12
runs it moved steadiness 0.90 → 0.90 (delta 0.00) while costing +0.003,
with both arms losing the identical probe. So balanced buys no measurable
steadiness on this instrument, and its status stays HYPOTHESIS — the honest
reading is "this instrument could not see the benefit", not "the benefit is
absent". Full detail in
the RC2.1 report ·
the RC2 report ·
RC1.2 随后闭合了该边界问题，并作出修正。 RC1.1 的 smoke 从未执行工具调用，因此
只能说明表面不携带未声明叙述，而非 EF 无法恢复它。通过真实 agent loop
（model → tool-call → ToolRuntime → tool/result → 下一模型步）与确定性的 keyless
证明，PARSE-7741 不在表面上，却存在于 context_search → context_recall 的返回中，
且在两种 semantic mode 下均如此。因此 semanticMode 并不限制召回，正确的契约是
已声明状态热保存；未声明历史可通过有界召回恢复。preset 保持 semanticMode: 'none'
——并非因为 rationale 负担不起（实测其真实约 7.4K token 的调用后仅增加 6.4%，而非
模拟器假设的 512），而是因为没有证据表明需要它。
RC1.2.1 随后修正了比较本身。 三臂实机 smoke 无条件传入了 plugin: true，而 harness
在该选项上提前返回，因此其 engine: 'basic' 展开是死代码——所谓 "Basic" arm 实为
EF，这正是它报出 facts retrievable 2/5 的原因（该指标对真正的 Basic 并无定义）。
harness 现在直接拒绝该组合。修正后的比较为：真正的 Basic 以零工具调用取得
3.00/3，而 economy-none 取得 2.40/3，且工具有 5/5 的运行返回了事实。因此差距在于
检索策略与模型工具使用，而非存储或机制。
RC1.3 随后闭合了这一差距，而它两侧都是检索工效学问题。 通过无提示 baseline
（probe 中不再出现 "use context_search"）与逐事实的失败分类，发现唯一的失败模式是
no-search——从未出现 recall-miss，因此不存在产品侧的检索缺陷——其下有两个具体缺陷。
其一，context_search 返回以命中位置居中的 240 字符摘录，会把一条 230 字符消息的
其他事实截掉：模型检索某个值时拿到只提到其邻居的摘录，于是把其余问题答成
unknown。命中本身在制造 search-miss。 现在短消息整体返回，命中同时报告
matchKind、逐字摘录，以及承载该命中的 context_recall 页码。其二，测试填充文本以
十进制编号单元，于是保留的填充里 unit 64、unit 90 直接满足了事实匹配器——这一
confound 同样影响了 RC1.2 的数字。两者修正后 baseline 为 2.33/3；对 framing section
做一处可测量的改动——在回答或将某事报告为未知之前，先检索已折叠历史——使其达到
3.00/3，与 Basic 持平，并在两次独立的 n=9 运行中复现（9/9 均发生检索）。提示词现在
的增量为 0.00：部署方不再需要告诉 agent 如何使用它自己的工具。economy 仍是
依赖检索的低成本模式（n=5 样本上仍约便宜 17 倍），并未被提升为通用默认
替代品。详见 RC1.3 报告 ·
RC1.2 报告。
RC1.3.1 随后冻结了检索层。 RC1.3 的 probe 把全部事实放在一个 checkpoint 的
一条消息里——这是最简单的情形，它掩盖了三个只在长任务中才会出现的时序缺陷。
其一，list() 按 createdAt 升序返回，而 search 在扫描途中就按 limit 截断，于是被丢弃的
恰恰是最新的 checkpoint：当前值最先消失。其二，locate 返回第一条匹配消息，因此被
取代的旧值会遮住它自己的更正。其三，exactPageOffset 是按默认页大小对齐的，只有在调用方
恰好使用该页大小时才"包含命中"。三者都让模型把旧值当作当前值——对基于折叠历史工作的
agent 而言，这是最有害的一类错误，因为答案看起来有据可依。修正后：命中按会话时序
（source.orderedSurfaceSeqs，而非会回拨、会碰撞的墙钟）新→旧排序，limit 因此意味着
"最近 N 条"；locate 返回最新匹配并给出 matchCount（大于 1 即表示该事实有历史）与
earliestMatchedMessageIndex（旧值仍可精确取回）；matchedMessageIndex 取代页对齐偏移，
在任意页大小下都有效。20 个 keyless 测试钉住这些行为，并重跑了实机套件确认无回归
（仍为 3.00/3，9/9 检索）。至此检索层正式冻结——压缩架构、召回正确性、检索工效学三者
全部 CLOSED，仅剩路由级实际成本 gate 仍为 OPEN，且它只影响是否推荐/默认启用 economy，
不再反向驱动架构增长。详见 RC1.3.1 报告。
RC2 随后构建了产品表面，并如实承认未能区分自己的三档。 现提供三档模式——
economy、balanced、quality——构成一个阶梯，每上一档只变动一个杠杆
（先是语义面，再是保留比例），其余全部保持不变，因此差异可归因。economy 是唯一
有端到端实测证据的一档（RC1.3）；两个稳定性档在所有描述它们的界面上都标注为
HYPOTHESIS，因为"为稳定性付费"这一主张本项目尚未测量。/context status 注册在真实
DSH 命令层，因此不消耗任何模型 token，且每个数值都标明 measured 或 estimated
——无法确立的数值显示 unknown，绝不为 0。随后在真实文件系统与执行工具上，对四种
模式各跑了 24 次真实任务，传输干净（261 次 provider 调用 0 次丢失）。但结果未能区分
各模式：TaskQuality 在所有 arm 上都饱和于 1.00，唯一一次稳定性丢失在重跑后不复现
（同一单元 0.75 → 1.00）。因此阶梯的收益仍为假设，原因在测量工具而非代码——任务
过于简单，且 n=1 无法把模式效应与模型方差分开。该运行确实正面确认的一点是 restart 与
model-switch 下的折叠/召回契约：每个 arm 在折叠并恢复后，仍按更正后的规格产出了
工件。
RC2.1 随后修正了表面，并只问了一个真正重要的问题。 修复了 /context status 的四处
缺陷，其中两处属于本项目熟悉的那类错误——数值宣称的比实际测到的更多：缺失读数被报成
已测量的 0 压力（并据此算出 0.0% 占用率）；归档 token 数（chars/4 启发式）
被标为 measured，而真正被测量的只有消息条数。用量现在直接读取 DSH 自身的
tokenUsage projection，而不是从 compaction/summary 事件手工累加（后者漏掉了每一次
普通 assistant 回合）；并且把当前 checkpoints（此刻在表面上）与生命周期 folds
（自会话开始累计、来自 bundle store）分开——RC2 曾用后者的名义报告前者。RC2 的 legacy
arm 也被正名为它本来的样子——EF legacy，而非 Basic：它挂载的是 EF 插件 + legacy
策略，仍经由 EF 引擎折叠、仍写 Bundle、仍暴露 recall 工具，因此"对比 Basic"并不成立。
现在有了真正的 Basic arm。阶梯改为保留量优先（机制更强、代价更低），并用一次关键
A/B 单独隔离保留量：12 次运行中稳定性 0.90 → 0.90（delta 0.00），而成本 +0.003，
两个 arm 丢失的探针完全相同。因此在本测量工具下 balanced 买不到可测的稳定性，其状态
仍为 HYPOTHESIS——诚实的读法是"该工具看不见这项收益"，而不是"收益不存在"。详见
RC2.1 报告。
RC2 报告。
Full detail in the RC1.1 report ·
RC3 随后把它变成了真正的插件，并纠正了一处说法。 RC2.1 之前的一切都跑在手工搭建的
new Context() 里，而不是 DSH 中——provider 与工具是真的，宿主不是；而且当时 EF 根本
无法挂载到真实宿主：入口是裸 TypeScript，Node 无法执行。RC3 增加了构建步骤产出可加载的
lib/、插入 EF 并禁用 compaction-basic 的 dsh.bundle patch（EF 持有
ctx.compaction），以及 loader 读取的入口。挂载进本机真实的 DSH 0.2.0-rc.2 profile 后，
EF 确实发生了折叠：两次已提交的 compaction 事务、真实表面上的 EF1 L cp:… 标记，
以及磁盘上经哈希校验的 bundle。/context mode economy|balanced|quality 可切换运行中的
引擎，且控制面会拒绝"中途改变 framing"的切换，而不是半途生效。真实宿主还暴露了一个测试
harness 无法发现的 bug：在 Cordis 中 ctx.get('tools') 会抛错，除非 tools 在
inject 中声明，而所有 harness 都预先挂载了 ToolRuntime，因此这条坏探测永远不可达——
已改用 ctx.inject(['tools'], …)。仍有一项未决：Sidebar 面板需要打包的客户端产物
（核心 sidebar 仅浏览器端；真实 sidebar 插件带 1.1 MB 客户端 JS），EF 是自建浏览器构建
还是依赖第三方 sidebar 服务，属于范围决策。详见
RC3 报告。
三处缺陷未能通过审计：认证档案在
成本 gate 为 OPEN 时报告 （复用比冒充了测量）；
一个约 30K 采样尺度的安全估计被当作生产 headroom；重放同时变动了 trigger 与
retention——而这正是 RC1-H「出厂 reserve 有益」结论的来源。固定 retention 后符号
，该结论予以撤回。两个 confound 去除后，区域 Ω 足够宽（10 个稳健格点，
7 个在所有缓存情景下均低于平价），成熟的生产参数位于最宽区间内，但实机成本结论在
n=8 下仍不可复现。实机 smoke 还测出了一条真实产品边界：经济模式预设能跨折叠保留
状态，但承载未锚定的叙述（可复现地 0/3，而 Basic 为 1/3），因为
marker-only checkpoint 里没有叙述文本——rationale checkpoint 能恢复其中大部分，
因此这是预设自身的选择而非架构限制。项目状态现为，且未为翻转默认值新增任何架构。详见
。
Frozen stages · 已冻结阶段
These are closed and should not be re-litigated in future reports. They reopen
only when a new incident corpus produces corresponding evidence (RC1 §46) ·
以下阶段已关闭，后续报告不再重复讨论；仅当新的故障语料产生对应证据时才重新开启
（RC1 §46）：
M1 CLOSED / deferred by evidence
M3b CLOSED / deferred
M4 CLOSED / deferred
M5 CLOSED / no measured need
RecallPrune CLOSED / no measured problem
DeltaLeaf REJECTED (R1-B gate)
The three layers RC1.3.1 closes are frozen the same way · RC1.3.1 闭合的三个层次
以同样方式冻结：
Compression architecture CLOSED (fold/frontier; unchanged since R3)
Recall correctness CLOSED (RC1.2 mechanism proof; exact recall bounded)
Retrieval ergonomics CLOSED (RC1.3 probe + RC1.3.1 temporal guard)
semanticMode none (nothing required rationale)
RC2 adds the product surface without reopening any of them · RC2 增加了产品表面，
且未重新开启上述任何一层：
Product surface SHIPPED economy / balanced / quality + /context status
Tier steadiness benefit HYPOTHESIS (not confirmed by the RC2 task sample)
Only one question remains open, and it is a pricing question that must not
drive the architecture · 仅剩一个未决问题，且它属于定价问题，不得反向驱动架构：
route-level realized cost gate OPEN (dispersion at n=8)
Development · 开发方式
This repository is a standalone plugin source tree. Tests run the vendored DSH
sources directly (the same source-level resolution the DSH monorepo uses),
so no build of the plugin itself is needed ·
本仓库是独立的插件源码树。测试直接运行 vendor 的 DSH 源码（与 DSH
monorepo 相同的源码级解析方式），插件本身无需构建。
Prerequisites · 前置要求：Node ^22.19 || >=24, pnpm 11.7.x, npm.
# 1. Vendor the DSH monorepo at the verified baseline · 在核验基线上 vendor DSH monorepo
git clone https://github.com/deepseek-ai/deepseek-harness.git vendor/deepseek-harness
cd vendor/deepseek-harness
git checkout 477b4f420553e8a52c2fbccc464d7561b239c443
pnpm install
# 2. Build declaration output for the packages EF consumes · 构建 EF 消费包的声明输出
node --max-old-space-size=8192 ./node_modules/typescript/bin/tsc -b \
packages/compaction/compaction-basic packages/core/tools packages/util/atomic-write
# 3. Back in the plugin repo: install tooling and (re)generate the
# src/type resolution maps into vitest.config.ts + tsconfig.json
# 回到插件仓库：安装工具链并重新生成两张解析表
cd ../..
npm install
node scripts/generate-maps.cjs
# 4. Run everything · 全量运行
npx vitest run # the whole suite · 全量测试套件
npx tsc --noEmit -p tsconfig.json # type check against vendor declarations
The suite grows every stage, so this file deliberately does NOT state a test
count: a hardcoded number goes stale the moment a spec is added, and a stale
count is read as an unfinished suite. vitest run prints the authoritative
totals. The live tiers are opt-in (EF_LIVE=1) and SKIP without a resolved
route, so an unmeasured behavior is never reported as a passing one.
本文件刻意不写死测试数量：每个阶段都会增加用例，写死的数字一经添加新用例即过期，
而过期的数字会被误读为套件不完整。权威数字以 vitest run 的输出为准。
Live 层为可选（EF_LIVE=1），没有可用路由时跳过，绝不把「未测量」报成「通过」。
scripts/generate-maps.cjs extracts the @deepseek-ai/* path table from the
vendored monorepo and emits two maps: scripts/vendor-paths.json (sources,
consumed by the vitest alias table) and scripts/vendor-types-paths.json
(built declarations, consumed by tsconfig). Re-run it after rebasing the
vendor clone onto a new DSH version ·
该脚本从 vendor monorepo 提取 @deepseek-ai/* 路径表，产出两张映射：
scripts/vendor-paths.json（源码，供 vitest alias 表使用）与
scripts/vendor-types-paths.json（构建产物声明，供 tsconfig 使用）。vendor
clone 升级到新 DSH 版本后重新运行即可。
Plugin usage · 插件用法
Mount the composite plugin into a cordis context — it owns ctx.compaction,
registers the deterministic state projection, provides the authority-gated
anchor service as ctx.epistemicFold, and registers the context_search /
context_recall tools when a ToolRuntime is present ·
将复合插件挂载到 cordis context——它持有 ctx.compaction、注册确定性状态
projection、提供 ctx.epistemicFold（authority 门控的 anchor 写入通道），
并在存在 ToolRuntime 时注册 context_search / context_recall：
import { EpistemicFoldPlugin } from 'dsh-epistemic-fold/plugin'
await ctx.plugin(EpistemicFoldPlugin, {
auto: true,
// Point bundles at the profile's persistence root (cleaned with the profile).
bundleRoot: `${profileRoot}/epistemic-fold`,
semanticMode: 'rationale', // or 'none' for a zero-LLM deterministic runtime
})
DSH profile YAML（等价形态 · equivalent form）:
- name: dsh-epistemic-fold
config:
bundleRoot: <profile persistence root>/epistemic-fold
semanticMode: rationale
Configuration modes · 配置模式
The product exposes three tiers, plus legacy (the engine's own default)
· 产品提供三档模式，另有 legacy（引擎自身默认值）：
- name: dsh-epistemic-fold
config:
bundleRoot: <profile persistence root>/epistemic-fold
mode: economy # economy | balanced | quality | legacy (default)
| tier | the lever it adds | evidence |
|---|
economy | — (default retention) | MEASURED — RC1.3: parity with Basic at ~1/17 the cost |
balanced | a larger verbatim tail (retention 0.16 → 0.24) | HYPOTHESIS — the RC2.1 A/B measured no gain |
quality | + narrative checkpoints (semanticMode: rationale) | HYPOTHESIS — not measured |
The ladder varies one lever per rung — retention first, then the semantic
face — with everything else held identical, so a difference between adjacent
rungs is attributable. Retention comes first because it is the stronger and
cheaper mechanism: a fact inside the retained tail never leaves the surface,
while a rationale checkpoint still has to be read. Run /context status in a
session to see the active mode, its evidence status, and the live figures ·
阶梯每上一档只变动一个杠杆（先保留比例、再语义面），其余保持不变。保留量在前，
因为它更强也更省：留在保留尾部的事实根本不会离开表面，而 rationale checkpoint 仍需
被读取。在会话中运行 /context status 可查看当前模式、其证据状态与实时数值。
mode: economy fills in the policy keys R3 measured as cheaper
(leafAdmission: economic, rootPolicy: economics, semanticMode: none,
framingMode: system-dedup). Every tier is a named set of values, not a
branch: expansion happens before resolution, so the engine cannot tell a preset
from the same keys written by hand, and an explicit setting always wins over
the tier. mode: economy 展开为 R3 实测更省的策略键组合；每一档都是「一组具名取值」
而非分支，展开在解析之前完成，因此引擎无法区分 preset 与手写配置，且显式设置始终
覆盖 preset。
framingMode: system-dedup requires the DSH frameCheckpoint seam. Without it
the engine refuses to start rather than silently running with per-checkpoint
framing — which would report an economy saving the deployment does not get.
Apply the seam with node scripts/apply-framing-seam.mjs <dsh-root>, upgrade
DSH, or set framingMode: legacy explicitly.
system-dedup 依赖 DSH 的 frameCheckpoint seam；缺失时引擎拒绝启动，而不是静默退回
per-checkpoint framing（那会报告一个实际并未获得的节省）。
reliability is deliberately not offered: there is no live evidence for what
the right reliability configuration is, and naming one would assert a conclusion
the project does not have. The two steadiness rungs are named balanced and
quality and carry an explicit HYPOTHESIS status instead of borrowing an
unearned name. reliability 暂不提供：尚无实测证据确定最优 reliability 配置，命名它
等于断言一个项目尚未得到的结论。两个稳定性档命名为 balanced 与 quality，并显式标注
HYPOTHESIS 状态。
Installing into a real DSH · 安装到真实 DSH
EF is a real DSH plugin as of RC3: it builds to loadable JS and ships a bundle
patch. A profile adds it as a dependency and lists it as a bundle ·
自 RC3 起 EF 是真正的 DSH 插件：构建为可加载 JS 并附带 bundle patch。profile 将其作为
依赖并列为 bundle：
// <DSH_HOME>/profiles/<name>/package.json
{
"dependencies": { "dsh-epistemic-fold": "file:/path/to/dsh-epistemic-fold" },
"dsh": { "profile": { "bundles": ["@deepseek-ai/dsh-base", "dsh-epistemic-fold"] } }
}
Then pnpm install in the profile directory and boot with
dsh --profile <name>. The patch disables compaction-basic, because EF
owns ctx.compaction and Cordis allows one registration per service ·
在 profile 目录执行 pnpm install，然后 dsh --profile <name> 启动。patch 会
禁用 compaction-basic：EF 持有 ctx.compaction，而 Cordis 每个服务只允许一次注册。
The patch ships mode: legacy, deliberately. Every tier selects
framingMode: system-dedup, which requires the frameCheckpoint seam — and
released DSH does not have it. A tier on an unpatched harness refuses to
start rather than silently running the costlier framing. Select a tier in
config on a seam-carrying build, or at runtime with /context mode.
patch 有意使用 mode: legacy：三档都需要 frameCheckpoint seam，而已发布的 DSH
没有它。未打补丁的宿主上选择档位会拒绝启动，而不是静默使用更贵的 framing。
Continuous Integration runs two lanes on every push: the pinned DSH
baseline (mandatory) and DSH master (allowed-to-fail compatibility
probe) · CI 每次推送跑两条 lane：pinned 基线（必过）与 DSH master
（allowed-to-fail 兼容性探测）。
Repository layout · 仓库结构
src/
engine.ts EpistemicFoldEngine — Basic's transaction + EF compile hook · Basic 事务 + EF compile hook
policy.ts plugin config resolution + routed-model pressure math · 配置解析 + 压力数学
candidate.ts pending fold-candidate identity (single slot per session) · 每 session 单槽
bundle-store.ts FileBundleStore — atomic, hash-verified, 0600 permissions · 原子写、哈希校验
compiler.ts input splitting, canonical bundle build, checkpoint rendering · 输入切分、bundle 构建
frontier.ts Fold Frontier: locate/re-derive from the CURRENT surface · 从当前 surface 重定位
leaf-policy.ts [frontier+1, bestEnd] span selection + frozen budget load · span 选择 + 冻结预算
root-policy.ts root rebase advisories (manual /compact = Root Fold) · rebase 建议
state.ts deterministic StateReducer: anchors, supersession, lifecycles · 确定性状态
authority.ts which event kinds may ground which authority domains · authority 模型
projection.ts wires the reducer into ctx.sessionProjections · 接入投影注册表
renderer.ts structured checkpoint: Current/Evidence/Open/Rationale/Recall · 结构化 checkpoint
recall.ts context_search + context_recall (bounded, paginated, exact) · 有界精确召回
tools.ts registers the recall tools against ctx.tools · 注册召回工具
hash.ts canonical JSON + SHA-256 digests · canonical JSON + 摘要
economics-profile.ts versioned cost model: ρ, ρ_eff, cache realization, break-even · 版本化成本模型
policy-compiler.ts amortized rebase policy: break-even horizon, hard overrides · 摊销式 rebase 策略
eval/src/
token-attribution.ts per-request token buckets that reconcile exactly · 词元来源归因
provider-telemetry.ts realized cache rate from provider usage · 缓存实况遥测
counterfactual.ts oracle arms measuring ROI upper bounds · 反事实 ROI 上界
pareto.ts frontier over cost/footprint/success; correctness is a filter · 帕累托前沿
eval/workloads/ W1–W5 workload matrix · 五类负载
profiles/economics/ versioned provider pricing (asOf + source) · 版本化价格数据
tests/ M0/M2/M3a/R0/R1 suites + shared harness (controlled LLM adapter) · 测试套件 + 共享 harness
bench/ paired-baseline harness (Basic vs EF prefix economics) · 配对基线 harness
Design docs · 设计文档
All design documents live in docs/ ·
全部设计文档位于 docs/：
| Document 文档 | Contents 内容 |
|---|
| 00_README_EF.md | Project entry: naming, goals, nine core invariants · 项目入口：命名、目标、九条核心不变量 |
| 01_EF_RFC_001_ARCHITECTURE.md | Architecture spec: truth model, data structures, fold transactions · 架构规范：真相模型、数据结构、折叠事务 |
| 02_EF_IMPLEMENTATION_PLAN_M0_M3.md | Engineering DAG, gates, stop conditions · 工程 DAG、阶段 Gate、停止条件 |
| 03_EF_TEST_BENCHMARK_SPEC.md | Metrics (ALR/SSR/DWR/CR/PMA), test suites, benchmark arms · 指标、测试套件、benchmark 分组 |
| 04_EF_LOCAL_AGENT_WORK_ORDER.md | Execution order and prohibitions for an implementation agent · 实现 Agent 的执行顺序与禁令 |
| 05_EF_DECISIONS_AND_OPEN_QUESTIONS.md | Frozen decisions, hypotheses, open questions · 冻结决定、假设、开放问题 |
| 06_FINAL_REPORT.md | Final implementation report: gates, evidence, deviations, refactor record · 最终实现报告 |
| 07_R0C_EVALUATION_REPORT.md | R0-C evaluation: measurement integrity, corpus, paired continuation, long-horizon economics, next-stage decision · R0-C 评估报告 |
| 08_BOUNDARY_CORPUS_PROTOCOL.md |
profiles/economics/ holds versioned provider pricing data (asOf + source,
caller-overridable) used by the R1 cost model — benchmark input, never
algorithm constants · profiles/economics/ 存放版本化的 provider 价格数据
（带 asOf 与来源、可被调用方覆盖），供 R1 成本模型使用——属于 benchmark
输入，而非算法常量。
License · 许可证
MIT
