Skip to content

chore(eval): promote OpenRouter DeepSeek baselines from run 42993062842 - #3

Open
manykarim wants to merge 1 commit into
change/openrouter-ci-eval-tierfrom
baselines/openrouter-deepseek-flash
Open

manykarim wants to merge 1 commit into
change/openrouter-ci-eval-tierfrom
baselines/openrouter-deepseek-flash

Conversation

@manykarim

Copy link
Copy Markdown
Collaborator

Promotes the first OpenRouter baselines (deepseek/deepseek-v4.1-flash) for rf-api, rf-core, rf-mobile and rf-web, under plugins/<plugin>/evals/baselines/openrouter/deepseek-v4.1-flash/ (openrouter-ci-eval-tier tasks 6.1–6.2). Stacked on #2; merge #2 first.

Source: run 42993062842: all tiers, both arms, N=6, turn-limit factor 1.5, all trigger sets on both splits, on 42be259. Files are the candidate-baseline-* artifacts, unchanged.

rf-desktop is not in this PR. Its PlatynUI tasks stopped at the 1.5 limit in 3–5 of 6 sessions and swung between 0.33 and 0.83 across two runs, so #2 now gives it factor 2.0 (007a7ab); its baseline is being recorded in run 43096210688 and will be added here.

Review

Pass rate with the skill (T) vs. without (B), N=6:

Plugin Notable tasks
rf-core skill lift on adv-language-force-tags (T 1.00 / B 0.00), adv-language-rf71-typed (0.83 / 0.17), adv-setup-system-pip (0.83 / 0.33)
rf-web lift on adv-browser-deprecated-wait (0.83 / 0.33)
rf-mobile adv-appium-deprecated-visibility 0.67 / 0.50
rf-api every task 1.00 in both arms (no measurable lift on DeepSeek)
  • Flaky: adv-legacy-syntax and adv-selenium-nonexistent-kw have T 0.83 (one failed session of 6) against B 1.00; expect noise in the gate there.
  • Turn-limit stops: common in the treatment arm (for example 6/6 on narrow-language-env-varfiles-01, 4/6 on adv-restinstance-expectation-leak), but those sessions still pass, so 1.5 stays for these four plugins.
  • Triggers: recall about 1.0, precision about 0.5–0.8; DeepSeek loads skills on many negative queries. The baseline records that level.
  • No incomplete runs.

Spend (key usage before/after each job)

Plugin Key spend Harness estimate (list prices)
rf-api $1.41 $2.23
rf-core $3.47 $5.25
rf-desktop $0.78 $1.18
rf-mobile $0.94 $1.49
rf-web $1.73 $2.43
Total $8.34 $12.58

The list-price estimate is 37–60% above the charged amount, so the cost caps err on the safe side.

Verification

After #2 merges, the next PR eval run should report the gate as pass (not rebaseline-needed) for these four plugins.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FHEmzL3ErAhGQykKWDx2wP

Candidate baselines for rf-api, rf-core, rf-mobile and rf-web from the N=6
run on deepseek/deepseek-v4.1-flash (all tiers, both arms, turn-limit
factor 1.5, trigger sets on both splits), copied unchanged from the
candidate-baseline artifacts. rf-desktop follows from a run at factor 2.0.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FHEmzL3ErAhGQykKWDx2wP
@manykarim manykarim closed this Nov 12, 2026
@manykarim manykarim reopened this Nov 12, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant