benchmark final wording

DATASHEET.md

@@ -37,7 +37,7 @@ expat 2.9.0 — platform-tagged in provenance)**, curated PE samples, cross-comp
 - Scrubbing: every symbol equal to the answer or a callee replaced with `<...>`; independent audit: **0/10,251 answer leaks**.
 - Ground truth derived from artifact itself (symbol tables, header fields, decompiler output).
 - Auto-grader included (`grading.py`): deterministic graders for exact/MC/crossarch/flag/library families (exact-match is alias-tolerant); rubric judge for open-ended. Test harness `test_grading.py`: 10,172 deterministic items scored 1.0 on gold answers, 0.0 on abstention; 79 routed to judge.
-- **Frontier reference scores measured** (`EVAL_RESULTS-frontier.md`, 2026-10-29, Claude Code Opus 5.5; small samples n=12/3, full pass@5 leaderboard in progress): single-shot no-tools ID sample 7/12 (~58%, CI wide — still unreliable); agentic flag-hunt pass@1 3/3 correct with tools in ~70–80 s → public envs are solvable by frontier agents; hard private tier + per-model pass-rate calibration is where RL training signal lives.
+- **Full agentic benchmark measured** (`EVAL_RESULTS-passk.md`, 2026-10-29): Claude Opus 5.5 solved **29/29 envs, 145/145 attempts** with tools (median 44 s) — every answer byte-exact vs our gold: independent confirmation of unique-solvability and of our oracles. The public tier is frontier-warm-up level; the paid product is the private difficulty-calibrated tier (target mid-band pass rates for YOUR model) + per-model pass-rate calibration service. Single-shot no-tools ID reference stays 7/12 small-n (`EVAL_RESULTS-frontier.md`).
 - Agentic challenges: round 1 (8) + round 2 (6, stripped) + **round 3 (6: self-decrypting .text via .init_array, ptrace/TracerPid/timing anti-instrumentation, chained-seed stages, masked-blob hash + explicit_bzero, control-flow flattening with decoy states, encrypted-bytecode VM with permuted opcodes)** + **round 4 (5: two-stage ARX+FNV memory-only chain, MIPS-BE branch-delay-slot trap, Windows PE TLS-callback packer, AArch64 computed-`blr` chain with PAC noise, RISC-V PC-relative jump-table `jalr`)** + **round 5 (4: x86-64 static-pie relocation-table/GOT-indirect hunt, ARM32 Thumb/ARM interworking trap, PPC64LE condition-register juggling gate, PE PEB-walk API hiding with export-name hashing)** — technique superset of published binary-RE agent benchmarks; every binary verified OK-on-flag / NO-on-mutation (24 single-char mutations each, zero false OKs); exit status = first failing position enables black-box probing; `strings` shows no flag fragments.
 - Provenance per item: binary name, sha256_16, address, size, arch.