experiments: context-tier A/B harness + findings (interface-vs-model reproduction) + ops-context server

[shyan-ai]

Follow-up to #10 and the findings comment there — the second half of the package, raised as a concrete proposal so the placement discussion has something to review. Happy to relocate, split, or trim per maintainer preference; the structure below is my best guess at your lanes, not a claim about them.

What this adds (no data changes, no task changes):

1. experiments/interface-ab/ops-context/ — the context-tier MCP server from the A/B: three layers over the existing mock servers — a declarative linkage map, curated governed views (exclusions applied at the data layer, MSA thresholds pre-joined), and a context fingerprint (per-response provenance: which records/documents an answer consumed). It composes like the stock mocks (own docker-compose.yaml, port 8014) and speaks to the same dataset. I'd have placed it beside crm/pm/file-server, but /mcp-servers/ is a regenerated-archive path (gitignored), so it lives inside the experiment dir — folding it into your archive pipeline instead is a one-word review comment away.

2. experiments/interface-ab/ — the harness + FINDINGS.md from the 75-trial reproduction of the benchmark's interface-vs-model result (same model, same tasks, same judge; only the interface differs). Narrative version: https://shyanming.substack.com/p/same-model-different-floor. The n=1-per-cell caveat travels with every number.

Why experiments/ as a new directory: proposing it as the home for community reproductions — n=1 reruns of each other's results are the cheapest way to harden leaderboard numbers. If you'd rather this live in scripts/, docs/, or stay external and linked, say the word and I'll restructure — this PR exists to turn the "where should this go" question from #10's comment thread into a reviewable diff.

Conformance notes: canary string embedded in every new file (# noqa: E501 on those lines); make validate passes (ruff + repo validation); the runner's setup hints translated from my local substrate to mcp-servers/compose-up.sh (+ the ops-context compose in this dir). The harness references the ops-* tasks from #10 but runs against any task directory with the standard layout.

[nimit2801]

Thanks for the PR. First-touch read: this is an experiments/community reproduction package, not a leaderboard submission, task change, dataset change, or dependency bump. It adds an interface A/B harness, findings writeup, config variants, and an ops-context MCP server under experiments/interface-ab/.

Checks: GitHub reports no checks on this branch yet; the PR body says make validate passes locally. I’ll reassess if CI/checks appear.

Main risk: the new experiment server and findings may be read as benchmark-endorsed results unless placement, reproducibility steps, and caveats stay very explicit.

Maintainer verification points:

1. Confirm experiments/ is the desired home for community reproductions and whether ops-context should instead fold into the generated MCP archive pipeline.
2. Run make validate plus at least one documented harness smoke run from a clean checkout.
3. Check that the findings clearly preserve the n=1-per-cell caveat and do not imply leaderboard status.