RaD preserves research code, evidence and conclusions. This repository is the durable record; large execution payloads remain in the local working storage. The connected repository is hipotures/rad.
The benchmark archive preserves earlier Qwen phases,
Strata comparisons, hardware characterization, scripts, configurations,
patches and compact results from /srv/ai/benchmarks. These investigations and
the residency laboratory belong to the same RaD project. Find serving scripts
and their configurations in the starter index, including
direct links to every preserved start-128k.sh.
The familiar report shortcuts are now portable links: Qwen first phase, phase 2 and phase 3. See the workspace map for Git and local execution paths.
The existing expert-residency investigation stays in
iq3s-residency-20261004T230051Z.
Its original absolute path remains
/srv/ai/research/iq3s-residency-20261004T230051Z, so measured launchers, frozen
manifests and existing reproduction commands retain their original addressing.
The checkout location is not fixed. Determine the repository root from the current working directory using git rev-parse --show-toplevel; historical absolute paths above are archival references, not current working-directory instructions.
Start with the research index. A prior completed study is Golden Swap Phase 3: 36/36 valid main and 6/6 RFC8259 replay requests. Incoming-query memoization removed about 97% of targeted recomputation, but no main task met the greater- than-3% incremental practical gain criterion. Inventory and Chinook retain conditional gains against current; remaining publication/wait attribution is unresolved. Phase 4's report records six valid v2 CONTROL/TRACE requests plus the retained v1 pair. Replay fidelity passed; the symmetric instrumentation-neutrality gate failed. Structural publication overlap is retained, but quantitative completion-path attribution remains blocked. The Phase 3 post-hoc chronology correction leaves its headline performance conclusion and timed measurements unchanged.
The Golden Swap Experiment Atlas provides a portable interactive retrospective dashboard: startup expert sets, residency generations, churn, current/future-informed comparisons, predictor decision funnels and lease/reuse evidence from 12 completed experiment namespaces. Its Python server needs no GPU service or external CDN; see the review for access and regeneration instructions. Start with the curated visual evidence gallery: twelve exact interactive views, physical slot turnover, and a documented BEFORE/AFTER visual audit.
Reports, scripts, patches, contracts, compact JSON/CSV results, small numeric model exports and research plots belong in Git. Build trees, model weights, source clones, virtual environments, large datasets, tapes, telemetry and large row-level result dumps stay local in their original form. Complete gzip copies of JSON, JSONL, log and text evidence are publishable when each is strictly below 10 MiB compressed. See text evidence and the historical backfill. Git ignores original execution payloads; it does not delete them. See storage and reproduction and the local artifact catalog.
For a fresh task, read AGENTS.md before creating files. Durable
material goes in research/<topic>/: authored code/, small fixtures/,
configs/, isolated run results, reports and recovery manifests. Large inputs, downloaded repositories, environments and compiled output belong in task-owned ignored storage, as specified by the active topic instructions. For topics requiring all work to stay under the current directory, use the topic's ignored work/ directory. Create/check .gitignore before execution.
Before the final commit, package completed text evidence into a fresh namespace:
python3 tools/archive_workspace.py pack-text --source <completed-evidence-directory> --destination <topic>/evidence/<fresh-run-id>Review the compressed manifest and stage those copies with the task artifacts. At completion, stage the actual task-owned paths explicitly, inspect the diff and run the publication check:
# Run from the current research directory; do not change directories.
git status --short --branch
git add -- <task-owned-paths-relative-to-current-directory>
git diff --cached --stat
git diff --cached --check
python3 "$(git rev-parse --show-toplevel)/tools/archive_workspace.py" audit --staged-only
git commit -m "Record the task outcome and reproduction material"
git push origin <actual-branch>Replace placeholders with reviewed paths and the appropriate upstream branch. Verify the remote contains the commit. The default limits are 1 MiB per file, 4 MiB per patch, strictly below 10 MiB per permitted gzip text copy, and 20 MiB of staged content per commit. Larger aggregate imports require a documented policy decision; the historical text backfill has its own scoped exception.
tools/archive_workspace.py plan and stage are bulk-import tools for the
managed workspace, rather than the completion workflow for an individual task.
Review their plan in storage/retention-plan.json before explicitly using them.
The helper never commits, pushes, deletes, rebuilds or runs a benchmark.
Do not use git add -f . to bypass the storage policy.
Read AGENTS.md for the execution protocol. Continue this
investigation in its existing directory. New, unrelated research topics use
research/<short-kebab-case-name>/ as specified there.
The migration record identifies the preserved legacy Git history. The original remote instructions remain in that history; AGENTS.md now defines the fresh-task workflow. LICENSE was retained. Existing source material and dependencies keep their own licensing; see third-party provenance.