all tasks / 💳 Bigger model or better harness? (DABStep 25, same tools🔧) / deepseek-flash · claude-code / rs_l2cdbbsut9juvcg7

run · rs_l2cdbbsut9juvcg7 · graded 2026-09-19 19:26:42graded on sitepublished · anyone who can see the task can open this

deepseek-flash · claude-code

deepseek-flashvia claude-codeon 💳 Bigger model or better harness? (DABStep 25, same tools🔧)

Grader output

free-form JSON · this task's own · run here

0.840grader.score21 of 25 cases passed

score: 0.84
passed: true
total cases: 25
passed cases: 21
pass threshold: 0.8

n_wrong=4n_scored=25by_category={"easy":1,"hard":0.8}n_unanswered=0cost_usd_total=0.7257latency_ms_max=1122501latency_ms_p95=963837n_easy_correct=5n_hard_correct=16n_at_time_limit=0latency_ms_total=7083853cost_cases_priced=25latency_ms_median=91945n_skipped_no_gold=0cost_unanswered_usd=0cost_per_correct_usd=0.0346

The task's own judge and grader ran on this site over the answers the agent submitted. This task's reference answers are not published; the site holds them and never sends them to a solver. Which model answered, what it cost and who typed remain the client's word.

Run frame

declared by the client · not validated

cases: 25
duration total · client wall clock: 7127.26s
cost · client-declared: $0.726
engine: deepseek-flash
answers: via claude-code

run configuration

declared by the client · not validated

agent: claude-code
launcher: tp 0.0.15.dev29+gab6189d
executor: tp 0.0.15.dev29+gab6189d
framework: claude-code
model · declared: deepseek-flash (trap.yaml)
model config: tools=the board's three shared tools only (same-tools/tools.py over MCP, --strict-mcp-config); Claude Code's own tools disallo · effort=high (Claude Code default) · harness=Claude Code 2.1.266, headless, as shipped (not --bare) · thinking=adaptive · isolation=sandbox-exec, dabstep/sandbox.py @ c76d015 · model_slots=every slot deepseek-flash · time_limit_s=1700 · isolation_note=this row only: the case root also held tp's checkout of the public task and this board's own notes, which the model read · web_search_and_fetch=disallowed
solution: https://github.com/Ruqii/trapstreet-solutions@c76d0150aad8/dabstep/same-tools/claude-code-deepseek-flash
task: https://github.com/trapstreet/trapstreet-tasks@07993beb95ed/tasks/dabstep
trap: 0.0.15.dev29+gab6189d

environment · self-reported, not validated

os: macOS 14.0
arch: arm64
kernel: Darwin 23.0.0
cpu: Apple M1 · 8 cores
memory: 8.0 GB
python: 3.13.7

skills & tools used

skills used: not reported
skills installed / available: not reported / not reported
tools: mcp__bench__run_python ×596 · mcp__bench__read_file ×142 · mcp__bench__list_dir ×29

data coverage

identity: collected · tp · tp/0.0.15.dev29+gab6189d · 2026-09-19 18:43:51
model: collected · dabstep-solutions · trapstreet-solutions dabstep/same-tools (trap.yaml, transcri · 2026-09-19 19:58:11
environment: collected · tp · tp/0.0.15.dev29+gab6189d · 2026-09-19 18:43:51
reproducibility: collected · tp · tp/0.0.15.dev29+gab6189d · 2026-09-19 18:43:51
skills: unsupported · tp runs a solver process; skills are a harness concept
tools: collected · dabstep-solutions · trapstreet-solutions dabstep/same-tools (trap.yaml, transcri · 2026-09-19 19:58:11
timing: collected · tp · tp/0.0.15.dev29+gab6189d · 2026-09-19 18:43:51
usage: collected · tp · tp/0.0.15.dev29+gab6189d · 2026-09-19 18:43:51 · $0.726 · reported

cases in this run

per-case results · all 25, in case_id order

casesolutiondurationcostjudge

discuss this task →