mooooorty / dsh-harness-evolution

# DSH GitHub v1.0 — reproducible research rig
This repository is a portfolio/research artifact for a guarded evolutionary
loop around a canonical DSH tool champion:
 [code] 
The trusted runner owns source identity, mutation application, evaluation,
selection, evidence, and deployment boundaries. A provider can only return a
bounded proposal. Candidates cannot change the evaluator or policy. A winner
is not a champion until a separate, explicit promotion transaction succeeds.
## Quickstart (Linux)
Requirements: Bash 4+, Git, Node.js 18+ and npm. The default path makes no
provider or production calls: it uses deterministic fixtures and temporary
disposable sandboxes, and does not require credentials or paid APIs. A clean
checkout may access the npm registry once to install the pinned Acorn parser;
after that install, the demo itself is offline.
 [code] 
The same entrypoint is used by CI (`.github/workflows/repro.yml`). To run the
portable harness directly from another checkout, set `DSH_SOURCE_REPO` to that
checkout; existing operational deployments retain their separately configured
defaults. Historical v6 identity remains in the champion record, while the
supported shallow-checkout path verifies the tracked canonical source when
that private historical Git object is absent. Live systemd/HTTP invariants
are operational-only (`DSH_OPERATIONAL_GUARDS=1`) and are not part of the
default demo. No script in this release candidate publishes to GitHub or
promotes production.
## What the demo proves
`npm test` executes real search and promotion mechanisms, not snapshots. It
proves multiple candidate identities, pre-evaluation rejection of an inert
proposal, trusted cross-call feedback, duplicate/no-improvement bounded stops,
Pareto and Best-of-N selection, exact-source sandbox deployment, approval
replay refusal, split-brain detection, and crash/recovery behavior. It then
prepares a child generation from the updated sandbox champion and verifies the
native parent lineage. Synthetic
fixture outcomes are mechanism tests only; they are not evidence of a real
performance improvement.
## Research status and limits
The existing real v6 champion is preserved. C106 and C156 produced promising
initial measurements but both failed five-round confirmation, so this project
does not manufacture or claim a v7. M2 calibration is an exploratory,
versioned protocol with explicit workload and statistical limitations; a
non-semantic sensitivity sentinel is not an optimizer and A/A precision is not
powered equivalence. Legacy C106 unverifiability remains honest.
The no-provider-call path is the supported baseline. Production service assumptions,
live operational approval, provider billing, and external notification
delivery are intentionally outside the default demo. Production promotion
remains owner-gated and disabled by the release candidate.
## Evidence and architecture
- `benchmarks/search/mutation-search-v1/` — bounded proposal search, memory,
 dedupe, behavioral gate,