Four passing runs out of 80 separate first from second on Specific's private-code benchmark

Every number on Specific's leaderboard lands on a multiple of 1.25 points. Ten tasks at eight runs apiece give each combination 80 rollouts, and one rollout is worth 1.25 points ^[20]. Fable 5.1's leading 38.8% is 31 passes out of 80, and GPT-6 Astra's 33.8% is 27 ^[20][21]. Four passing runs separate first from second ^[21]. The whole table, first to eighth, spans 22.5 points, which is 18 rollouts ^[23].

Six of the ten analyzed tasks scored under 15% aggregate ^[9]. Each of those tasks was attempted 64 times across the eight combinations, so under 15% means nine passes or fewer ^[26]. Two entries ran on the same harness, Codex CLI, and finished 17.6 points apart ^[24]. Eight independent runs per task is careful measurement, and ten tasks is still a pilot.

The reading offered by runtimewire.com is that public coding scores can reward familiarity with visible repositories, and that Real-SWE gives Specific a route from evaluation failures to training-data customers ^[12]. There is a reason behind the first half: public benchmarks can expose the repository, the issue and the eventual patch, any of which can enter training data, while a private codebase makes an agent read the system in front of it and infer conventions from the workspace ^[11]. Confirming it needs a matched pair: the same eight setups, the same native harnesses, the same pass@1 over eight runs, scored on a public benchmark next to Real-SWE. Specific has not run that comparison; the published leaderboard reports Real-SWE ^[3].

Across eight repository-backed sample tasks, the median instruction runs 1,742 characters and the median reference solution touches 11 files ^[18]. That works out to about 158 characters of instruction per file the reference changes ^[25]. The agent has to find the rest of the work itself ^[27].

Y Combinator's company page says Specific acquires and licenses operational data and codebases ^[13], and YC lists it as a two-person company founded in 2025 and backed in the Fall 2025 batch ^[14]. The sample codebases are described without being named: an events product with more than 200,000 users, a consumer fintech that has processed more than 100,000 bank statements, and enterprise sales software, all on Specific's account ^[15].