S3RL4 commited on
Commit
62a3372
·
verified ·
1 Parent(s): a622fcc

Update EVAL_RESULTS-passk.md: difficulty wording and caveats

Browse files
Files changed (1) hide show
  1. EVAL_RESULTS-passk.md +3 -5
EVAL_RESULTS-passk.md CHANGED
@@ -1,22 +1,20 @@
1
  # EVAL_RESULTS-passk.md — full pass@5 agentic benchmark on all 29 envs (2026-10-29)
2
 
3
- Model: Claude Opus 5.5 (Claude Code CLI v2.1.310, `--model opus`, fresh `--no-session-persistence` session per attempt).
4
  Attempts: 5 per env x 29 envs = **145 runs**. Tools: Bash+Read only (disassembly + run probes; WSL/qemu for non-x86).
5
  Grading: OUR verifier oracle — run the challenge binary with the model's ANSWER string as argv[1]; pass = stdout "OK"
6
  + exit 0. Independent of any string compare. Raw logs are retained; summary table below.
7
 
8
  ## Headline (honest)
9
  - **29/29 envs solved by Opus 5.5 with tools; 145/145 attempts correct** (after 3 infra-retries that returned empty,
10
- re-run clean). Every model answer was byte-exact equal to our gold (checked separately) — independent frontier
11
- confirmation of unique-solvability AND of our oracles.
12
  - Solve cost: median 44 s, mean 47 s (range 23-121 s), mean 8.4 tool turns. Cross-arch envs (MIPS/AArch64/RISC-V/
13
  ARM32/PPC64LE) were NOT harder for Opus than x86-64 once qemu+objdump were available.
14
  - Implication we state plainly to buyers: the public flag-hunt tier is **warm-up level for a frontier agent**
15
  (all-pass groups give ~zero GRPO gradient at that capability level). Our answer: (a) use them as the free demo/
16
  calibration tier, (b) sell a **private hard tier** targeting mid-band pass rates for the buyer's own model,
17
  (c) offer per-model pass-rate calibration (we run it on our harness before you buy).
18
- - Not claimed: SOTA comparisons, open-weight numbers yet (next step), pass@N on the private tier (by definition
19
- unseen by any model).
20
 
21
  ## Method notes / gotchas found while building this
22
  - WSL re-parses argv from Windows (shell metacharacters in answers like `<`, `;`, `&`, backtick mangled) -> grading
 
1
  # EVAL_RESULTS-passk.md — full pass@5 agentic benchmark on all 29 envs (2026-10-29)
2
 
3
+ Model: Claude Opus 5.5 through a hosted frontier agent, fresh session per attempt.
4
  Attempts: 5 per env x 29 envs = **145 runs**. Tools: Bash+Read only (disassembly + run probes; WSL/qemu for non-x86).
5
  Grading: OUR verifier oracle — run the challenge binary with the model's ANSWER string as argv[1]; pass = stdout "OK"
6
  + exit 0. Independent of any string compare. Raw logs are retained; summary table below.
7
 
8
  ## Headline (honest)
9
  - **29/29 envs solved by Opus 5.5 with tools; 145/145 attempts correct** (after 3 infra-retries that returned empty,
10
+ re-run clean). Every model answer was byte-exact equal to our gold (checked separately) — this confirms the shipped binary accepts those answers and verifies our grading chain; it is not proof that each answer is unique. The run was non-isolated, so treat it as an oracle check, not a sealed-valuation result.
 
11
  - Solve cost: median 44 s, mean 47 s (range 23-121 s), mean 8.4 tool turns. Cross-arch envs (MIPS/AArch64/RISC-V/
12
  ARM32/PPC64LE) were NOT harder for Opus than x86-64 once qemu+objdump were available.
13
  - Implication we state plainly to buyers: the public flag-hunt tier is **warm-up level for a frontier agent**
14
  (all-pass groups give ~zero GRPO gradient at that capability level). Our answer: (a) use them as the free demo/
15
  calibration tier, (b) sell a **private hard tier** targeting mid-band pass rates for the buyer's own model,
16
  (c) offer per-model pass-rate calibration (we run it on our harness before you buy).
17
+ - Not claimed: SOTA comparisons, open-weight numbers yet (next step), or a full pass@N calibration of the private tier. Hard-tier evidence so far is limited; see `EVAL_RESULTS-frontier.md` §C.
 
18
 
19
  ## Method notes / gotchas found while building this
20
  - WSL re-parses argv from Windows (shell metacharacters in answers like `<`, `;`, `&`, backtick mangled) -> grading