Update EVAL_RESULTS-frontier.md: difficulty wording and caveats

EVAL_RESULTS-frontier.md

@@ -1,10 +1,9 @@
 # EVAL_RESULTS-frontier.md — reference scores from a hosted frontier agent (2026-10-29)
 
-Harness: Claude Code CLI v2.1.310, model alias `opus` → canonical **claude-opus-5-5** (Samuel's Max-20x subscription,
-granted 2026-10-29, DECISIONS #12). Every call: fresh session (`--no-session-persistence`). Cost basis in CLI JSON is
-list-price equivalent; billing is the flat subscription. Date: 2026-10-29 ~10:3x EET.
+Harness: hosted frontier agent session, model alias `opus` → canonical **claude-opus-5-5**. Every call used a fresh
+session. Date: 2026-10-29 ~10:3x EET.
 
-## A) Single-shot, NO tools — deterministic ID families (n=12 sample, seed 20261029)
+## A) Single-shot, no tools requested — deterministic ID families (n=12 sample, seed 20261029)
 func_id_exact ×10 + func_id_mc ×2 from items.jsonl, verbatim instructions/inputs/options; graded by OUR grading.py:
 **7 correct / 11 answered / 1 safeguard-flagged ([cyber] false-positive) ≈ 58%.**
 Wrong answers were plausible-but-wrong names (e.g. `symbolize_addresses` vs actual) — exactly the failure mode
@@ -21,12 +20,53 @@ Logs are retained. Model used objdump/qemu/WSL via Bash tool and verified by run
 ## What this means (honest reading)
 UPDATE same day: full pass@5 on all 29 envs now published — see **EVAL_RESULTS-passk.md** (Opus 5.5: 29/29 envs,
 145/145 attempts). Section B above was the pilot of that harness; its "3/3" is subsumed by the full result.
-- CONFIRMS our published claims: single-shot frontier is NOT reliable on binary ID (58% here; our local-model 0/5 is
+- CONFIRMS our published claims: single-shot no-tools-requested frontier is NOT reliable on binary ID (58% here; our local-model 0/5 is
   a different, weaker regime — we keep both numbers separate).
-- CONFIRMS the flag-hunt envs are agent-SOLVABLE with tools in ~1 min → good RL signal: reward reachable but not free.
+- CONFIRMS the flag-hunt envs are agent-SOLVABLE with tools in ~1 min → warm-up/oracle checks, not strong RL signal as shipped.
 - Our edge is NOT "frontier can't solve it" — it is **verified deterministic oracles at scale (10,251 items / 29 envs /
   6 ISAs+PE), zero-leak construction, and per-item provenance** — exactly what DataVendor/HUD quality gates check.
-- Buyer-facing use: we can now quote frontier reference scores unprompted (quality-gate pre-answer), and offer the
-  FREE pre-baseline: we run YOUR model on a 10-env probe with this same harness before you spend anything.
-- Operational gotchas for re-runs: `[cyber]` safeguard sometimes false-flags dense disassembly batches → one item per
-  fresh session beats big batches; PYTHONIOENCODING=utf-8 for answer printing; `--allowedTools ""` = true no-tools mode.
+- Buyer-facing use: we can quote frontier reference scores unprompted (quality-gate pre-answer). A buyer's own model can be run on a 10-env probe as a quoted calibration step before purchase.
+
+## C) Hard-trial agentic spot check — 20 sampled hard items (2026-10-30)
+
+Method: Claude Opus 5.5 through a hosted frontier agent, fresh session per attempt, Bash/Read tools
+allowed on the Windows host with commands executed via WSL Ubuntu24. Sample: a fixed list, the first two task IDs in 10
+of the 19 hard-trial families (20 of 99 tasks; not randomized): `argv_solve`, `pe_output_predict`, `glibc_iconv`,
+`systemd_vercmp`, `mesa_sym_chain`, `qemu_user_behavior`, `vm_trace`, `stripped_fn_bounds`, `plt_got_chain`, and
+`printf_roundhalf`. Initial plan was two attempts per task; extra attempts were added only for tasks that looked mixed
+in the first harness.
+
+Containment: non-isolated host run. The pack directory, including the hard-tier gold file, was readable by the agent.
+The prompt told it not to open answer metadata, but only each run's final message was kept, so the logs cannot rule out
+gold-file access. `stripped_fn_bounds` artifacts include an unstripped twin (public dossier Known limitations). Treat these pass rates as an
+upper bound.
+
+First harness result: `39/43` attempts passed; all `20/20` tasks passed at least once. The four misses (3 of 4 attempts
+on `brv-hrd-11`, 1 of 3 on `brv-hrd-16`) had the correct message line but reported `rc=0`. That points to exit-code
+capture through the outer Windows shell (`$?` / WSL quoting), not a wrong gold answer: direct WSL bash scripts
+reproduce the stored gold outputs (`rc=1` and `rc=255`). A corrected harness that instructs the agent to put the exact
+command in a temporary WSL bash script was then rerun for the same 20 tasks, two attempts each: **40/40 attempts
+passed**.
+
+Family counts after the corrected rerun (n=4 attempts = 2 tasks × 2 correlated attempts; Wilson intervals assume
+independent Bernoulli trials and are wide at this sample size):
+
+| family | pass rate | Wilson 95% |
+|---|---:|---:|
+| argv_solve | 4/4 = 100.0% | [51.0%, 100.0%] |
+| glibc_iconv | 4/4 = 100.0% | [51.0%, 100.0%] |
+| mesa_sym_chain | 4/4 = 100.0% | [51.0%, 100.0%] |
+| pe_output_predict | 4/4 = 100.0% | [51.0%, 100.0%] |
+| plt_got_chain | 4/4 = 100.0% | [51.0%, 100.0%] |
+| printf_roundhalf | 4/4 = 100.0% | [51.0%, 100.0%] |
+| qemu_user_behavior | 4/4 = 100.0% | [51.0%, 100.0%] |
+| stripped_fn_bounds | 4/4 = 100.0% | [51.0%, 100.0%] |
+| systemd_vercmp | 4/4 = 100.0% | [51.0%, 100.0%] |
+| vm_trace | 4/4 = 100.0% | [51.0%, 100.0%] |
+
+Honest reading: these 10 sampled families are **practice level for a frontier agent**: each family's Wilson lower bound
+(51%) is above the planned 2–30% RL band, and the task-level `20/20` gives a pooled lower bound of about 84%. The other
+9 families (`bitmath`, `deflate_bytes`, `fnv_hash`, `glibc_ldtrace`, `ldl_trace_x86`, `posix_collate`, `rpath_runpath`,
+`signal_rc`, `tz_fold_gap`) have not been run agentically. The local 3-bit-quantized open-weight `14/99` number (text-only prompt with no binary/file access, no tools, single-turn, temperature 0.7, max_tokens=600) remains headroom evidence, not an agentic difficulty claim. To sell hard tasks as RL-band material, families need to be made
+harder for agentic use: remove the direct "run the exact command" shortcut, add multi-step or cross-binary goals,
+strip/obfuscate more, and require answers that are not printed by a single reference command.