Review convergence practice and measure keyless research search - #718

Scope

Review the convergence practice against current upstreams, repair the unsupported inferred Alpaca SDK pin, and supply the research-skill round-2 input pack. Complete the preregistered 30-query keyless comparison: retain current DDGS(auto) BY_DESIGN within this corpus; SearXNG fails the quality gates.

• Base commit: cb339488e3e004b04e8a4e5b4fd246442a275e3c.
• Lane: lane:foundation; this PR stays draft for the command center's Claude read and landing.
• Owned paths: convergence docs/recipes, landscape source projection and its existing tests, and bounded research-skill/web-search artifacts. Shared evidence registration is exclusively in the final commit. Shared templates, install-plan paths and provider configuration remain with their owners.
• Frozen search preregistration source: 05481fc85e6dfcdf46f8e45a893886250ea0a0a3, published before live calls; the rebased pack preserves every original input byte.

SOTA sources

• native-agent-stack@4c897418:AGENTS.md, trading north star and landscape README review-queue item 11: the user pins an Alpaca adapter path, without pinning the SDK repository. These sources govern the projection repair, discriminating regression assertion and anti-pattern record.

• U11 revision 5 at the review base and source-only converter contract: sources for the recipe/architecture/runner documentation corrections.

• Anthropic agent evaluation guidance, GPT Researcher v3.7.0, DeerFlow v2.1.0, and DeepResearch Bench: current primary-source anchors for the practice review. The decision contains pinned release/source/doc citations for every inventoried research runtime, evaluation harness, SDK, native orchestration route and alternative, rather than transferring historical scores to current routes.

• Native agent-lab reference@e070125d:layer-verdict-lane.js: supported orchestration reference, not acceptance of the complete V2 policy.

• native-agent-stack@ec0b8fd8:docs/lanes.md, hot-file protocol: source for taking main's registry, native file re-registration, generator checks and registry-last commit.

• GPT Researcher@0957c301:.mcp.json:4, pyproject.toml:23, published package metadata, and native uv tool semantics: sources for the isolated executable probe and source/distribution distinction.

• gptr-mcp@63884773:server.py:277, release history, DeerFlow@345f08be:HTTP skill, and chat mode flags:55: maintenance, lifecycle and HTTP-gate input sources.

• promptfoo@34f74d34:test-agent-skills guide: report/entry-route evaluation design; native skill-creator remains a separate skill-adoption gate.

• promptfoo@34f74d34 Python provider, native assertions, native CLI filters, unchanged DeerFlow DDGS tool and SearXNG tool: native keyless comparison and only necessary provider/scorer glue.

• SearXNG@d48c4b55 supported container, JSON format, native upstream container CI and rootless Docker: isolated published-image measurement, no rebuild or shared installation.

• SciPy one-sided binomial-test documentation: reference for the preregistered exact greater-tail sign-test calculation; no new evaluation runner or statistical dependency.

• DeerFlow@345f08be documented sandbox provider: correction of the required-but-unused selector in the frozen adapter. Direct search calls instantiate no sandbox; the pack establishes no sandbox/full-runtime acceptance.

Evidence-class table

Local commands run

$ rtk proxy env TMPDIR=<private-lane-cache>/tmp python3 -m unittest discover -s tests -p test_landscape_sweep_harness.py
exit 0; 289 tests, 3 skipped
$ rtk proxy env TMPDIR=<private-lane-cache>/tmp python3 scripts/validate.py
exit 0; 69 components, 10178 hashed files, 4 profiles, 204 receipts
$ rtk proxy python3 scripts/validate_convergence.py evidence/artifacts/web-search-comparison-20261005/preregistered-experiment.json evidence/artifacts/web-search-comparison-20261005/experiment.json
exit 0; both records valid
$ rtk proxy python3 scripts/component_matrix.py --write
exit 0; 32 rows, 0 flip-rule violations; reports unchanged
$ rtk proxy python3 scripts/new_host_grand_list.py --write
exit 0; 32 layers, 66 winners; reports unchanged
$ promptfoo eval --config <frozen-config> --repeat 1 --filter-pattern <frozen-10-query-filter> --max-concurrency 1 --delay 1000 --no-cache --no-share --no-write --no-table --no-progress-bar --output <private-native-export>
nine exact commands retained in experiment.json; gradedexit100 each, 180 calls retained
$ rtk git diff --check
exit 0

Rebase onto captured main cb33948 completed with no source conflict. Only the old final registry conflicted; it was skipped and rebuilt from main through the native protocol. Forty owned files are re-registered through native host_receipts.register_file from main's exact pinned-base registry. Only the observed scoped experiment is added to convergence_records[]; receipt-array entries, component pins, install-plan rows and host status are preserved. The corrected registry passes validation; the initial moving-origin registry mismatch and literal-UUID publication failure are retained. Public identifiers are exact UTF-8 SHA256 digests; actual output/export hash byte scopes are explicit.

Decision record

• docs/decisions/2026-10-05-convergence-practice-upstream-review.md: inventory, alternatives, native capability limits, U11revision5 source-only boundary and overturn conditions.
• docs/decisions/2026-10-05-research-skill-round-2-input-pack.md: maintenance, isolated executable probes, HTTP facts and conditional round-2 design.
• docs/decisions/2026-10-05-keyless-web-search-comparison.md: immutable preregistration, method, full native results, cleanup and completeness critic.

Candidate paired wins/losses/ties=3/13/14; exact predeclared greater-tail p=65399/65536=0.9979095459 and primary-hit gain=-7/30. Completeness and latency gates pass; gain, significance and usable-rate gates fail. The candidate returned72 valid empty lists. Retain BY_DESIGN within this fixed corpus/deployment, with no SearXNG READY row proposal or general-equivalence claim.

Software-only source-family URL targets do not validate answer truth or versions; q27 is ambiguous. The strict paper-window gap, shared-engine/index/cache correlation, omitted engine diagnostics, exact-string duplicate scope, and excluded keyed/PDF/table/non-English/financial/regulatory modalities feed the next sweep. Log CAPTCHA/rate-limit/timeout records do not establish per-query causes. The frozen unused sandbox selector remains unchanged and establishes no sandbox acceptance.

U11GateA script review/trading-owner acknowledgement and the command center's cross-family read remain landing conditions. Rebase under docs/lanes.md immediately before merge; this lane never merges or pushes main.

Host evidence

No evidence/hosts/ files changed. Ruling10 reserves shared-host plan apply to co-op/readiness-runner; this lane ran no install.sh. The isolated published rootless container and owned config/data are removed. Native usage/internal retries remain unknown; no model, GPU, full research or token-savings claim is made.

Checklist

• No GitHub Actions changed; no workflow pins or permissions introduced.
• No secrets printed, logged or committed; no required secret added.
• No new paid hosting, subscription or billing surface introduced.
• Peer-owned files/worktrees preserved.
• Evidence registry committed last; draft maintained for the command center's read.