The original nine-benchmark report contains single-run summaries. Independent results point both ways, while Standard, self-hosted, and different-task runs cannot directly refute the Minimal protocol.
Automatically summarized from source discussions. Sources and the full record are available below.
Evidence & timeline
Current conclusion
The original results are reported, but generalizability is unverified; counter-results under mismatched protocols do not establish contradiction.
Maintainer response
No verified maintainer response found
Untriaged
Strong original claim vs adjacent counterevidence
The tension comes from outcome direction and protocol comparability; mismatched protocols cannot be collapsed into a single verdict.
Revision criteria
A same-protocol independent reproduction with complete task-level material would upgrade the claim; evidence that differences fall within natural variation or protocol confounding would downgrade it.
Current conclusion
The original table lacks task-level results, trajectories, run IDs, endpoints, commits, and scorer details; adjacent experiments do not show a consistent large gain.
Scope
V4 Flash/Pro and the source author’s single-run project benchmark records.
Verification
Not yet verified
Maintainer response
Untriaged
Author response
Documentation change · Aug 18, 2026, 03:49 PM
Product impact
Affects community Skill selection, benchmark trust, and documentation boundaries.
Suggested next action
Publish task IDs, task-level results, trajectories, run IDs, cost, retries, and exclusions before evaluating the nine-benchmark strong claim.
Lifecycle
Active
Source scope
Public GitHub and Bilibili sources reviewed through 2026-08-22
Attribution confidence · 100%
First found by this site · Aug 22, 2026, 03:30 PM
Status change log
Verification · Unknown → Not yet verified · Aug 16, 2026, 11:39 AM
The strong result table contains single-run summaries without task-level artifacts.
Open evidence ↗Supporting evidence
The author published a nine-benchmark result table, a Minimal-plus-J-Space protocol summary, and several single-run scores.
Conflicting evidence
A custom-task A/B showed no significant difference, four presets converged in the SWE-bench Pro 151 summary, and the author later narrowed the report to engineering observations rather than causal proof.
Key counterargument
The adjacent experiments are not same-protocol reproductions of the nine original benchmarks, so they reduce confidence in generalization but cannot directly refute every reported result.
Missing evidence
Task IDs, task-level results, raw trajectories, run IDs, endpoint, commit, scorer, cost, retries, and exclusion rules are missing.
Verification plan
Publish the original nine-run packet and have an independent party repeat at least one large-gain and one no-gain result under the same Minimal-plus-J-Space protocol.
Revision criteria
A same-protocol independent reproduction with complete task-level material would upgrade the claim; evidence that differences fall within natural variation or protocol confounding would downgrade it.
Scope & boundaries
Public material permits assessment of reporting completeness and adjacent evidence tension, not recalculation of the unpublished nine original runs.
| Source | Contributor | Stance | Independent sources | Exact evidence |
|---|---|---|---|---|
| Replace report with v3.6.1 engineering observations record Aug 18, 2026, 03:49 PM · Version 1 | Tiger3807861189 Resolver | Corrects | Original report and author revision | “它不是研究论文，非学术性质，不提供模型内部机制证明、形式化判定方法、消融设计或因果贡献分解。” |
| Independent A/B verification on DeepSeek V4-Flash Aug 17, 2026, 03:12 AM · Version 1 | HHHEEEWWW Challenger | Challenges | Custom-task A/B | “在最终任务完成度上，A/B 无显著差异” |
| 关于有很大提升，但是不稳定的回复 Aug 17, 2026, 09:59 AM · Version 1 | Tiger3807861189 Synthesizer | Supports | Original report and author revision | “它塑造的是倾向，不是保证” |
| Add DeepSeek V4 × J-Space capability realization report Aug 16, 2026, 11:39 AM · Version 1 | Tiger3807861189 Originator | Originates | Original report and author revision | “J-Space 是一个插件，已通过 benchmark 验证了它对于 Deepseek 的提升巨大，flash基本持平 GLM5.3，pro超越Fable 5。” |
| SWE-bench Pro 151 · DeepSeek Harness vs OpenCode 最终评测报告 Aug 19, 2026, 07:30 AM · Version 1 | suyoumo Challenger | Challenges | SWE-bench Pro 151 summary | “四个 preset 成绩收敛在 109-114（72-76%），差距（2-5 题）小于同配置重跑的自然波动。” |
suyoumo · Not reproduced
| Protocol field | Original run | Reproduction | Comparability |
|---|---|---|---|
| Model provider• | DeepSeek official API claimed; endpoint not published | Unknown | Missing from reproduction run |
| Model• | V4-Flash-0731 and V4-Pro-0813 | DeepSeek V4 Flash | Mismatch |
| Model revision• | Unknown | Unknown | Missing from original run |
| Serving method• | Unknown | Unknown | Missing from original run |
| Harness commit• | Unknown | Unknown | Missing from original run |
| Preset• | DSH Minimal + J-Space | jspace versus minimal and standard | Mismatch |
| Reasoning effort | max | max | Match |
| Benchmark• | nine benchmark summary | SWE-bench Pro | Mismatch |
| Benchmark commit• | Unknown | zh release v30 aligned | Missing from original run |
| Task set• | Unknown | 151 tasks across six repositories | Missing from original run |
| Scorer version• | Unknown | Unknown | Missing from original run |
| Run count | 1 | 1 | Match |
| Temperature | 1 | Unknown | Missing from reproduction run |
| Top P | 0.95 | Unknown | Missing from reproduction run |
HHHEEEWWW · Not reproduced
| Protocol field | Original run | Reproduction | Comparability |
|---|---|---|---|
| Model provider• | DeepSeek official API claimed; endpoint not published | Unknown | Missing from reproduction run |
| Model• | V4-Flash-0731 and V4-Pro-0813 | DeepSeek V4-Flash | Mismatch |
| Model revision• | Unknown | Unknown | Missing from original run |
| Serving method• | Unknown | local DSH environment; endpoint not published | Missing from original run |
| Harness commit• | Unknown | Unknown | Missing from original run |
| Preset• | DSH Minimal + J-Space | control versus J-Space | Mismatch |
| Reasoning effort | max | Unknown | Missing from reproduction run |
| Benchmark• | nine benchmark summary | custom reasoning, repository, and recovery tasks | Mismatch |
| Benchmark commit• | Unknown | Unknown | Missing from original run |
| Task set• | Unknown | three tasks with three runs per arm in round two | Missing from original run |
| Scorer version• | Unknown | Unknown | Missing from original run |
| Run count | 1 | 12 | Mismatch |
| Temperature | 1 | Unknown | Missing from reproduction run |
| Top P | 0.95 | Unknown | Missing from reproduction run |
Discovery
The strong nine-benchmark gain claim was first observed within the monitored sources.
Amplification
The public video combines J-Space with Feng V3Router, so it cannot attribute the outcome to J-Space alone.
Challenge
A custom-task A/B observed no completion gain and reported higher token and time cost.
Author update
The source author described the effect as a tendency rather than a guarantee while retaining a large-improvement claim.
Documentation change
The source author reframed the document as benchmark and engineering observations and disclaimed mechanism proof or causal decomposition.
Challenge
A fixed SWE-bench Pro 151 README summary reports J-Space tied with Minimal and below Standard, without task-level files.
Current conclusion
A near-Miminal run reports a slight decrease and a small-sample run reports a weak positive direction; the self-hosted Standard result has critical protocol differences and cannot directly refute the claim.
Scope
V4 Flash-0731, DSH Minimal, and a Terminal Bench 2.1 on/off experiment.
Verification
Mixed evidence
Maintainer response
Untriaged
Author response
No response recorded
Product impact
Affects Skill recommendations under DSH Minimal and TB2.1 interpretation.
Suggested next action
Freeze Minimal, endpoint, commits, scorer, and task IDs; preregister paired repeats and publish trajectories.
Lifecycle
Active
Source scope
Public GitHub and Bilibili sources reviewed through 2026-08-22
Attribution confidence · 100%
First found by this site · Aug 22, 2026, 03:30 PM
Status change log
Verification · Not yet verified → Mixed evidence · Aug 19, 2026, 05:11 AM
A near-Miminal failed run and a small positive direction coexist, with incomplete artifacts.
Open evidence ↗| Source | Contributor | Stance | Independent sources | Exact evidence |
|---|---|---|---|---|
| GoForceX clarifies the TB2.1 subset and concurrency Aug 17, 2026, 02:03 AM · Version 1 | GoForceX Reproducer | Neutral | Near-Miminal counter-run | “terminal bench 2.1 的一个子集（87 道题，去掉了 2 道）” |
| TB2.1 分数在 DSH Minimal + J-Space Skill 下无法复现 Aug 16, 2026, 02:42 PM · Version 1 | GoForceX Challenger | Challenges | Near-Miminal counter-run | “添加 J-Space Skill 后的 benchmark 成绩略有下降” |
| 无法复现报告中提供的 Terminal Bench 2.1 正确率 Aug 18, 2026, 03:06 PM · Version 1 | Jyleaves Challenger | Challenges | Self-hosted Standard counter-run | “本复现在NVIDIA H20 96G x8 上自部署Deepseek V4 Flash-0731完成。使用dsh standard+jspace。” |
| 独立 A/B 复现：TB2 18 任务的小样本正方向 Aug 19, 2026, 05:11 AM · Version 1 | thl1173653555-create Reproducer | Supports | Small 18-task positive direction | “Control 10/18（55.6%），J-Space 11/18（61.1%），方向上 J-Space 略优但只多对 1 个任务，单次样本达不到统计显著。” |
| Add DeepSeek V4 × J-Space capability realization report Aug 16, 2026, 11:39 AM · Version 1 | Tiger3807861189 Originator | Originates | Original single-run TB2.1 result | “Terminal Bench 2.1：V4-Flash 82.7，V4-Flash + J-Space 87.1，V4-Pro 87.9，V4-Pro + J-Space 90.1。” |
| 无法复现报告中提供的 Terminal Bench 2.1 正确率 Aug 18, 2026, 03:08 PM · Version 1 | Jyleaves Amplifier | Challenges | Self-hosted Standard counter-run | “单次测试正确率69题/89题≈77.5%” |
GoForceX · Not reproduced
| Protocol field | Original run | Reproduction | Comparability |
|---|---|---|---|
| Model provider• | DeepSeek official API claimed; endpoint not published | Unknown | Missing from reproduction run |
| Model• | V4-Flash-0731 | Unknown | Missing from reproduction run |
| Model revision• | 0731 | Unknown | Missing from reproduction run |
| Serving method• | Unknown | Unknown | Missing from original run |
| Harness commit• | Unknown | Unknown | Missing from original run |
| Preset• | DSH Minimal | DSH Minimal | Match |
| Reasoning effort | max | Unknown | Missing from reproduction run |
| Benchmark• | Terminal Bench 2.1 | Terminal Bench 2.1 | Match |
| Benchmark commit• | Unknown | Unknown | Missing from original run |
| Task set• | Unknown | 87-task subset; two tasks removed | Missing from original run |
| Scorer version• | Unknown | Unknown | Missing from original run |
| Concurrency | Unknown | high | Missing from original run |
| Run count | 1 | 1 | Match |
| Temperature | 1 | Unknown | Missing from reproduction run |
| Top P | 0.95 | Unknown | Missing from reproduction run |
Jyleaves · Not reproduced
| Protocol field | Original run | Reproduction | Comparability |
|---|---|---|---|
| Model provider• | DeepSeek official API claimed; endpoint not published | self-hosted | Mismatch |
| Model• | V4-Flash-0731 | V4-Flash-0731 | Match |
| Model revision• | 0731 | 0731 | Match |
| Serving method• | Unknown | NVIDIA H20 96G x8 self-hosted | Missing from original run |
| Harness commit• | Unknown | Unknown | Missing from original run |
| Preset• | DSH Minimal | DSH Standard + J-Space | Mismatch |
| Reasoning effort | max | Unknown | Missing from reproduction run |
| Benchmark• | Terminal Bench 2.1 | Terminal Bench 2.1 | Match |
| Benchmark commit• | Unknown | Unknown | Missing from original run |
| Task set• | Unknown | 89 tasks | Missing from original run |
| Scorer version• | Unknown | Unknown | Missing from original run |
| Run count | 1 | 1 | Match |
| Temperature | 1 | Unknown | Missing from reproduction run |
| Top P | 0.95 | Unknown | Missing from reproduction run |
thl1173653555-create · Partially supports
| Protocol field | Original run | Reproduction | Comparability |
|---|---|---|---|
| Model provider• | DeepSeek official API claimed; endpoint not published | Unknown | Missing from reproduction run |
| Model• | V4-Flash-0731 | DeepSeek V4 Flash; source also says flash 116 | Mismatch |
| Model revision• | 0731 | Unknown | Missing from reproduction run |
| Serving method• | Unknown | Unknown | Missing from original run |
| Harness commit• | Unknown | Unknown | Missing from original run |
| Preset• | DSH Minimal | Harbor + terminus-2 neutral carrier | Mismatch |
| Reasoning effort | max | max | Match |
| Benchmark• | Terminal Bench 2.1 | Terminal-Bench 2 | Mismatch |
| Benchmark commit• | Unknown | Unknown | Missing from original run |
| Task set• | Unknown | 18 tasks | Missing from original run |
| Scorer version• | Unknown | Unknown | Missing from original run |
| Run count | 1 | 1 | Match |
| Temperature | 1 | Unknown | Missing from reproduction run |
| Top P | 0.95 | Unknown | Missing from reproduction run |
Failed reproduction
A near-Miminal 87-task subset reported a slight score decrease with higher tokens and cost; exact scores were not public.
Failed reproduction
The self-hosted H20 DSH Standard result of 69/89 has critical differences from the original Minimal protocol and cannot be marked contradicted.
Corroboration
An 18-task sample showed a weak 10→11 positive direction, with one additional pass and substantially higher cost.
Current conclusion
The public implementation has no J-lens, hidden-activation readout, or causal intervention; evidence supports only an external textual workflow or behavior change.
Scope
The public J-Space Skill implementation and external runtime workflow.
Verification
Not yet verified
Maintainer response
Untriaged
Author response
No response recorded
Earliest source found so far
Unknown
Unknown
Product impact
Affects Skill documentation, architecture language, and user understanding of the mechanism.
Suggested next action
Separate external workflow effects from internal mechanism claims; publish readout and ablation artifacts for any internal-intervention claim.
Lifecycle
Active
Source scope
Public GitHub sources reviewed through 2026-08-22; original originator unknown
Attribution confidence · 0%
First found by this site · Aug 22, 2026, 03:30 PM
Status change log
Verification · Unknown → Not yet verified · Aug 18, 2026, 11:34 AM
Public materials lack internal readout, intervention, and ablation evidence.
Open evidence ↗| Source | Contributor | Stance | Independent sources | Exact evidence |
|---|---|---|---|---|
| Replace report with v3.6.1 engineering observations record Aug 18, 2026, 03:49 PM · Version 1 | Tiger3807861189 Resolver | Corrects | Source-author documentation narrowing | “不提供模型内部机制证明、形式化判定方法、消融设计或因果贡献分解。” |
| 文本 examples 与内部 J-Space 机制的边界 Aug 18, 2026, 11:34 AM · Version 1 | WeZZard Challenger | Challenges | Internal-mechanism boundary challenge | “加载 examples 确实可以塑造模型行为，但这不等于介入了 J-Space。” |
No reproduction recorded.
Challenge
A methodological challenge distinguished textual example-driven behavior shaping from internal J-space intervention.
Add DeepSeek V4 × J-Space capability realization report
Open source Opens in a new tabTB2.1 分数在 DSH Minimal + J-Space Skill 下无法复现
Open source Opens in a new tab【似是故人来】爱.信.等. V4 Flash 双叉臂 使用 Jspace+风V3Router
Open source Opens in a new tabGoForceX clarifies the TB2.1 subset and concurrency
Open source Opens in a new tabIndependent A/B verification on DeepSeek V4-Flash
Open source Opens in a new tab无法复现报告中提供的 Terminal Bench 2.1 正确率
Open source Opens in a new tab无法复现报告中提供的 Terminal Bench 2.1 正确率
Open source Opens in a new tabReplace report with v3.6.1 engineering observations record
Open source Opens in a new tabSWE-bench Pro 151 · DeepSeek Harness vs OpenCode 最终评测报告
Open source Opens in a new tabThe lead source, “benchmark: step5-preview S1-S24 双臂运行归档 + 双判分结果,” says: “StepFun Step 5 Preview（OpenCode Zen 供应，公开发布名；预览端点背后精确 checkpoint 未经独立验证；opencode 编排）在 S 系列基准上的完整 24 题双臂运行、双判分与重跑 follow-up 归档。.” This is currently the only captured source.
The lead source, “feat(benchmark): S18 terminal sprite render trap (half-block SGR leak, trim ghosting, timer pinning),” says: “Summary Adds benchmark task S18-terminal-sprite-render-trap (static, read-only report): a terminal half-block sprite renderer whose three independent rendering bugs (SGR background.” 36 related sources are linked for comparison to determine whether they describe the same event.
The lead source, “feat(benchmark): H22 not-listed trap — the package that installs but never registers,” says: “Distilled from a real top-tier failure: anywhere-labs/dsh-desktop (22k★) failed our 50-plugin live verification on 09-01 with exactly this shape — dsh plugin add succeeds, the entr.” 31 related sources are linked for comparison to determine whether they describe the same event.
