dsh-learning-bundle: a proof-carrying correction loop for DSH rc.2

[B1lli]

I built an intentionally small, unofficial DSH bundle for a narrow question:

Did an adopted correction reach the model only where allowed, and did it actually change the result?

What it does

1. record creates a non-recallable candidate.
2. adopt explicitly activates it.
3. Recall fails closed across profile, workspace, and session identity.
4. A current explicit instruction overrides the learned default.
5. Every model-visible delivery is a structured user/message receipt that can be reconstructed from the durable session log.

The proof on current DSH

Against official dsh-v0.1.1-rc.2 (b150a55), the same real headless task produced:

Arm	Learning injection	Final response
L− candidate	0	npm install
L+ adopted	1	pnpm install

A second profile received zero injections, five explicit npm/Yarn overrides won, and the logged delivery reconstructed headless + pre-step + workspace identity.

Install the prebuilt release

curl -LO https://github.com/B1lli/dsh-learning-bundle/releases/download/v0.3.0/dsh-learning-bundle-0.3.0.tgz
dsh plugin --profile headless add -w ./dsh-learning-bundle-0.3.0.tgz

Repository and full usage: https://github.com/B1lli/dsh-learning-bundle
Release: https://github.com/B1lli/dsh-learning-bundle/releases/tag/v0.3.0

Evidence and limits

• 34/34 automated tests on Node 22 and 24
• 10/10 release gates
• 8 public scenarios / 72 deterministic contract probes
• historical bounded live DeepSeek sample: 3 L− + 3 L+ decisions
• current rc.2 and historical rc.8 assembled consumers
• independent fresh-context review: PASS in all three dimensions, no blockers

This is not a claim that the DSH memory ecosystem is empty. dsh-mnemon and dsh-continual-evolve are broader products. This project is the dependency-free, proof-focused reference; automatic extraction, Web UI, team sync, aggregate canary, and sealed holdout are not claimed.

I would especially value reports containing one real correction, the intended scope, the next prompt, whether it recalled, and any adjacent prompt where it should not have recalled.