dsh-attention-index

# public evidence (this repo)
open https://perrylink.github.io/dsh-attention-index/

# in-app consumer
dsh plugin --profile web add dsh-community-bench

The same rule, animated.

Why this exists

Plugin discovery is not a popularity contest, and a download count is not a task-outcome claim. dsh-score separates four things that used to be mixed:

1. Discovery — is this plugin maintained and visible?
2. Evidence — was it run with and without on the same task?
3. Grade — how strong is that evidence (E0–E5)?
4. Attention — only a positive, adequately powered result gets a quality slot.

At v0.1 the answer is deliberately uncomfortable: there is no quality #1.

Current state

The first pilot is dsh-memento@0.5.27 on five memory tasks × three runs/arm. Both arms passed every task: it is a null result, capped at E2 because the task pack is unreviewed, the runner is operator-owned, and the sandbox is network-isolated but not seccomp-hardened. rank is null; top3 is empty.

What a grade means

Hard gates: an unreviewed pack or a partial sandbox caps at E2; E5 requires an independent audit in data/audits/index.json. No prior-only row can receive a quality rank. No one can buy a slot.

Full model: docs/SOE-GOE.md · SCORING.md.

What is public

curl https://perrylink.github.io/dsh-attention-index/api/v1/registry.json

Quick start

# 1. Recompute the registry deterministically
node scripts/aggregate.mjs --today 2026-10-18

# 2. Contract checks
node scripts/check-invariants.mjs
node scripts/conformance.mjs
node --test test/*.test.mjs

# 3. Preview the site
python3 -m http.server 8765 --directory site

The pipeline

task pack (pilot, unreviewed)
  → paired with/without runs
  → signed result + environment manifest + fixture hash
  → strict ingest (signature, fixture, duplicate)
  → Newcombe interval + SOE/GOE grade
  → E0/E1 no rank · E2 provisional · E3+ positive quality slot

• P1 measured-evidence runbook
• Sandbox requirements
• P2 integration contract
• P3 audit checklist
• P0–P3 status

Not done yet

• A seccomp/cgroup-hardened sandbox; current bwrap-net has FS/PID/network isolation and a DeepSeek-only egress proxy, but no seccomp.
• Task review and an independent runner for E3+.
• FDR/bootstrap are implemented but not wired; holdouts, version binding and token accounting remain.
• Signed snapshot history and a third-party market pilot.
• E5 (independent audit) — a self-review is published, but it is not E5.
• Repository rename to dsh-score-index; URLs are intentionally stable.

No fabricated results. No prior-only #1. No paid rank.

DSH plugin family

⭐ 如果它帮到了你

这个项目是 DSH 插件家族的一员（40+ 个，全部 Apache-2.0）。如果你在用，给个 star —— 它不会解锁任何功能，但会让下一个人在搜索里更容易找到它。

English: part of a 40+ plugin family for DeepSeek Harness. If it is useful, a star helps the next person find it — nothing is gated behind it.

License

Apache-2.0. See LICENSE.