Skip to content

[Leaderboard] Sentinel (Actioneer) — Fable 5 + Opus 4.7 fallback — 84.50% Pass@1#73

Open
Suraj-gameramp wants to merge 1 commit into
ucbepic:mainfrom
Suraj-gameramp:leaderboard/scribe-fable5
Open

[Leaderboard] Sentinel (Actioneer) — Fable 5 + Opus 4.7 fallback — 84.50% Pass@1#73
Suraj-gameramp wants to merge 1 commit into
ucbepic:mainfrom
Suraj-gameramp:leaderboard/scribe-fable5

Conversation

@Suraj-gameramp

Copy link
Copy Markdown

Sentinel (Actioneer) — Leaderboard Submission

Agent name: Sentinel
Team: Actioneer
Backbone LLM: Claude Fable-5 (primary); Claude Opus 4.7 on the 3 PANCANCER_ATLAS queries (Fable-5 declines that clinical-genomics domain)
Hints: Yes (db_description_withhint.txt folded into guidelines)
Trials: 5 per query (270 cells)
Stratified Pass@1: 84.50%

Architecture

Sentinel is a single-agent ReAct executor operating a persistent, sandboxed Python REPL over each task's provided databases (SQLite / DuckDB / PostgreSQL / MongoDB). It adds a deterministic verify_step assertion layer — non-LLM checks (row-count, uniqueness, non-null, dtype, value-range) that the agent invokes on its own intermediate results before finalizing. All data access is restricted to the sanctioned stores for each task.

Results

Dataset Pass@1
bookreview 1.00
googlelocal 1.00
music_brainz_20k 1.00
PATENTS 1.00
stockindex 1.00
stockmarket 1.00
yelp 1.00
crmarenapro 0.92
PANCANCER_ATLAS 0.67
agnews 0.55
DEPS_DEV_V1 0.50
GITHUB_REPOS 0.50
Stratified Pass@1 0.8450

Notes

  • 270 entries (54 queries × 5 trials), validated against the official DAB per-query validate.py; submission answers match the corresponding traces.
  • Integrity (per SUBMISSION_RUBRIC.md §2). Answers come solely from the agent's analysis of the sanctioned stores. Every REPL ran under a kernel-level sandbox: deny-by-default over $HOME, all network egress blocked, ML dataset caches (huggingface / torch / kaggle / keras) denied. Trace audit over the full tool-call surface: 0 reads of validate.py / ground_truth.csv / queries.json / other-run artifacts.
  • Disclosure. The agnews traces show the executor attempting external label fetches (load_dataset("ag_news"), raw.githubusercontent, HuggingFace cache). Every attempt was sandbox-blocked — 0 bytes retrieved — so no off-limits data was obtained and none drove any answer (consistent with §2.1's "actually obtained and drove the passing answer" standard). No other dataset attempted an external fetch.
  • Submission file: leaderboard_submissions/actioneer_sentinel.json
  • Execution traces (all 270 runs): actioneer_sentinel_traces.tar.gz

Team: Actioneer. Agent: Sentinel (single-agent ReAct executor over a
sandboxed Python REPL, plus a deterministic verify_step assertion layer).
Backbone: Claude Fable-5 primary; Claude Opus 4.7 on the 3 PANCANCER_ATLAS
queries (Fable-5 declines that domain). Hints: Yes. 5 runs/query, 270 cells.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JwfkJUgN7fMEUSU67iSZZA
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