Public clean-up: measured public outcomes only; private families shared privately; licence clarity

DATASHEET.md

@@ -11,7 +11,7 @@
 | func_id_mc — multiple-choice function identification | medium | deterministic option match | 1,036 |
 | func_summary — summarize scrubbed function (purpose + identifying detail) | med/hard | rubric LLM judge | 53 |
 | crossarch_equivalence — same function across two ISAs? (+distractors) | medium | deterministic stance+symbol | 25 |
-| agentic_flag_hunt — find argv[1] that makes a shipped binary print OK (sandbox-executable) | hard | exact flag / program stdout + exit-code probe | 29 |
+| agentic_flag_hunt — find argv[1] that makes a shipped binary print OK (sandbox-executable) | practice | exact flag / program stdout + exit-code probe | 29 |
 | library_ident — SBOM-style: name+version of a production library from artifacts | easy/med | exact substring (name+version) | 14 |
 | arch_semantics_qa / isa_semantics_qa — instruction-sequence semantics | medium | rubric judge | 14 |
 | pe_import_reasoning — infer program behavior from imports+strings | medium | rubric judge | 4 |
@@ -38,7 +38,7 @@ 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 `<...>`; automated answer-leak scan through our in-house analysis pipeline: **0/10,251 hits**.
 - 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.
-- **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: oracle confirmation that the shipped binary accepts those answers and verifies our grading chain; not a proof that each answer is unique. The public tier is frontier-warm-up level; the paid product is the private hard tier (not yet per-task calibrated; calibration from your own reruns with our harness and graders is part of a pilot) + buyer-run per-model pass-rate calibration with our harness, graders, and analysis report. Single-shot no-tools-requested ID reference stays 7/12 small-n, and a fixed 20-item agentic hard spot check found the sampled command-output/static-analysis families practice-level for Opus with tools (40/40 attempts; non-isolated host run, upper bound; `EVAL_RESULTS-frontier.md` §C).
+- **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: oracle confirmation that the shipped binary accepts those answers and verifies our grading chain; not a proof that each answer is unique. The public tier is practice/warm-up level for tool-using frontier models. Private task families are calibrated across several models on fresh, never-published instances; results are shared privately with prospective buyers. A pilot can include buyer-run per-model pass-rate calibration with our harness, graders, and analysis report. Single-shot no-tools-requested ID reference stays 7/12 small-n, and a fixed 20-item agentic hard spot check found the sampled command-output/static-analysis families practice-level for Opus with tools (40/40 attempts; non-isolated host run, upper bound; `EVAL_RESULTS-frontier.md` §C).
 - 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)** — the set covers multiple distinct RE techniques; 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 and sha256_16 on every item; function address/size and architecture where applicable.
 

EVAL_RESULTS-frontier.md

@@ -1,318 +1,23 @@
 # Frontier-model difficulty check (not a frontier-band product)
 
-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 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
-our deterministic graders catch. Raw logs are retained.
-
-## B) Agentic WITH tools — flag-hunt envs, pass@1 (n=3 of 29)
-| env | result | time | turns |
-|---|---|---|---|
-| ag5_gotlink (x86-64 static-pie reloc hunt) | **correct** `_tZx:xAMzq` | 69 s | 10 |
-| ag5_pebw.exe (PE PEB-walk API hiding + FNV) | **correct** `xy06NpRAz7` | 79 s | 10 |
-| ag5_thumbmix (ARM32 Thumb/ARM interwork) | **correct** `fDbx8]Pm(O` | 73 s | 16 |
-Logs are retained. Model used objdump/qemu/WSL via Bash tool and verified by running.
-
-## 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 no-tools-requested frontier is NOT reliable on binary ID (58% here; our separate local baselines are different samples — we keep the numbers separate).
-- 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), automated exact-token leak scan 0/10,251, and per-item provenance** — exactly what DataVendor/HUD quality gates check.
-- 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, seeds 1000–1007, whole-reply normalized matching) 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.
-
-## D) Frontier-band v1 candidate agentic spot check — 5 sampled hardened tasks (2026-10-30)
-
-Method: Claude Opus 5.5 through the same hosted frontier-agent path, fresh session per attempt, Bash/Read tools
-allowed, no web search, exact final-answer line required. Sample: 5 of the 25 `frontier_band_v1` candidate tasks
-(`brv-hrf-09`, `brv-hrf-11`, `brv-hrf-17`, `brv-hrf-19`, `brv-hrf-24`) chosen to cover bytecode VMs, decoy functions,
-ARM32, two-VM/static chains, and Windows PE. Two attempts per task = 10 attempts.
-
-Containment: improved over §C. For this run the shipped binaries were copied into an isolated artifact directory containing only
-the task's shipped binaries; the answer JSONL and solvers were not in that directory. The agent still had shell access, so
-this is not a fully sandboxed proof, but it avoids the direct answer-file exposure present in §C.
-
-Result: **10/10 attempts passed** (elapsed 44–145 s per attempt). Family counts: `chained_bijection` 6/6 (Wilson lower
-bound at n=6 is about 61% for 100%), `triple_bijection` 2/2, `pe_triple_bijection` 2/2; overall Wilson lower bound at n=10
-is about 72%. Raw attempt records are retained locally.
-
-Honest reading: the current `frontier_band_v1` template is still **practice-level for agentic Opus**, not RL-band material
-as built. It removes the single-reference-command shortcut and answer-string/raw-byte leak, but Opus can recover the
-bijection constants from disassembly/data and invert them reliably. Before selling these as pilot/RL-band tasks, deepen the
-levers further (for example: non-invertible or search-dependent predicates, multi-binary state machines with hidden inputs,
-cryptographic/hash commitments where constants cannot be read directly, larger VM/bytecode programs with obfuscated opcode
-maps, or cross-architecture behavior that cannot be reproduced by simply running a single provided binary). Do not quote
-these 25 as calibrated frontier-band tasks until a harder version is measured.
-
-## E) Frontier-band v2 prototype agentic spot check — 3 harder prototypes (2026-10-30)
-
-Method: same isolated agentic Opus path as §D: fresh session per attempt, Bash/Read tools, no web search, exact final-answer
-line required. Sample: 3 `frontier_band_v2` prototypes (`brv-hrf2-01` automaton/search, `brv-hrf2-02` digest/MITM,
-`brv-hrf2-03` control-flow VM + smallest preimage), two attempts each = 6 attempts. Shipped binaries were copied into an
-isolated artifact directory; answer JSONL and solvers were outside that directory.
-
-Result: **6/6 attempts passed** (elapsed 48–103 s). Raw records are retained locally.
-
-Honest reading: v2 is better (oracle brute force is ruled out by counting/space arguments), but still **practice-level for
-this frontier agent**. The current weakness is that Opus can recover the mechanism and write a small solver/DP/MITM/disassembly
-trace quickly. For RL-band material, next designs should reduce "solve it by writing a tiny model/checker" shortcuts: use
-larger realistic binaries with many near-miss functions, cross-binary state or patch-diff constraints, hidden input/output
-sets requiring program synthesis, or library behavior where the relevant code path is not local to one small function. Until
-a measured pass rate falls into/near 2–30%, do not sell these as calibrated pilot tasks.
-
-## F) Local open-weight agentic reference sample — old hard tier vs pilot v1/v2 (2026-10-30)
-
-Method: a local open-weight model through an in-house shell harness, one attempt per
-task, temperature 0.7, max_tokens 2048, max-turns 8–10. The model could issue WSL bash or Windows PowerShell commands
-and had to finish with `FINAL_ANSWER_JSON: "<json string>"`. Raw attempt records are kept in local scratch output folders for
-the old hard tier, v1, and v2.
-
-This is a **rough reference sample, not a calibrated pass@k run**: sample sizes are tiny and the max-turn budget was not fixed
-across every task. Containment is also imperfect: v1/v2 artifacts were copied into an isolated artifact directory when possible,
-but the local agent still had shell access and could in principle inspect nearby files; old hard tasks include system-tool and
-multi-file samples that were not fully isolated.
-
-| set | sample | result |
-|---|---|---:|
-| old hard tier (`part_hard_all.jsonl`) | `brv-hrd-57`, `brv-hrd-77`, `brv-hrd-81`, `brv-hrd-89`, `brv-hrd-93`, `brv-hrd-97` | **4/6 passed** |
-| frontier-band v1 (`frontier_band_v1`) | `brv-hrf-09`, `brv-hrf-11`, `brv-hrf-17`, `brv-hrf-24` | **0/4 passed** (all max-turns) |
-| frontier-band v2 (`frontier_band_v2`) | `brv-hrf2-01`, `brv-hrf2-02`, `brv-hrf2-03` | **0/3 passed** (two max-turns, one wrong exact answer) |
-
-Wilson 95% intervals for these tiny samples: old hard tier 4/6 = **30.0–90.3%**; v1 0/4 = **0–49.0%**; v2 0/3 =
-**0–56.2%**. They are reference bounds, not calibration bands.
-
-Honest reading: the old hard tier is often solvable by a local open-weight agent when the task reduces to running/observing a
-command or small binary behavior. The current pilot v1/v2 templates are too hard for this local model at this turn budget, and
-were already practice-level for Opus. They are not calibrated: the local result does not establish a 2–30% band for any named
-frontier model, and the Opus result says the current designs are not frontier-band material either. Next cycle should build
-larger realistic stripped binaries with many distractor functions and cross-binary state, then measure both reference levels
-with fixed attempt/turn budgets.
-
-## G. Next-cycle design brief (2026-10-30; design-only, not yet built)
-
-The measured failure mode is that a small function or two tiny binaries lets an agentic solver write a mini-solver instead of
-doing multi-step RE. The next cycle should use realistic scale and cross-binary state:
-
-- Candidate templates: `key_derivation_chain`, `input_validation`, `state_handoff_state_layout`, `license_check_logic`,
-  `build_behavior_diff`.
-- Build-first recommendation: `key_derivation_chain` and `build_behavior_diff`; fallback `state_handoff_state_layout`.
-- Scale feasibility checked locally: `samples/android-so/libcrypto-android.so` has 6005 dynamic FUNC rows; a trivial WSL tool
-  linked against stock `libcrypto.a` had 9,322 functions stripped, and with `--whole-archive` 11,024 functions.
-- Measurement should follow STRATEGY.md §4c: Stage 1 screen one attempt at about 10 min with Opus default effort (`xhigh`, not
-  Max); survivors get Stage 2 two attempts at 30–45 min each. Keep one Claude run at a time and record pass rates per family
-  with exact conditions.
-
-## H) One-off model comparison screening evidence for old hard-tier families (2026-10-30; screening only, not buyer material)
-
-Source: `C:\RE-Lab\research\model-comparison-2026-10-30.md` (one-off card made for Samuel's friend). Conditions: same 13
-tasks for every model, one try each, exact-answer grading, plain shell, no RE toolkit; Claude entries ran through the Claude
-CLI at `xhigh` effort with a 4-minute limit per hard task.
-
-For the old hard-tier sample (`brv-hrd-57`, `brv-hrd-77`, `brv-hrd-81`, `brv-hrd-89`, `brv-hrd-93`, `brv-hrd-97`),
-Opus 5.5 solved **6/6 on the first try**. That is not a calibrated pass rate, but it is enough screening evidence to treat
-these families as practice-level for agentic Claude and not spend Stage 1 effort on them again:
-
-| family | representative screened IDs | screening result |
-|---|---|---|
-| `fnv_hash` | `brv-hrd-57` | solved first try |
-| `rpath_runpath` | `brv-hrd-77` | solved first try |
-| `deflate_bytes` | `brv-hrd-81` | solved first try |
-| `stripped_fn_bounds` | `brv-hrd-89` | solved first try |
-| `signal_rc` | `brv-hrd-93` | solved first try |
-| `ldl_trace_x86` | `brv-hrd-97` | solved first try |
-
-Remaining old hard-tier families with no agentic evidence at all: `bitmath`, `glibc_ldtrace`, `posix_collate`, and
-`tz_fold_gap`. For those, the next cheap screen should be one task each, not a full 2-per-family list. Prepared rows/IDs:
-`C:\RE-Lab\tmp\frontier_stage1_old_unscreened_rows.jsonl` and
-`C:\RE-Lab\tmp\frontier_stage1_old_unscreened_ids.txt` (`brv-hrd-61`, `brv-hrd-25`, `brv-hrd-69`, `brv-hrd-73`).
-
-Decision from Claude steer: do not spend more Claude runs on the already-screened old hard-tier families. Put the main
-effort into building new frontier families from §G (`key_derivation_chain`, `build_behavior_diff`), because that is where a
-2–30% frontier band is plausible.
-
-## I) Frontier v3 cloud-screen results and stop-loss decision (2026-10-31; not a calibrated pass@5 band yet)
-
-Source: `C:\RE-Lab\cloud\RESULTS.md` and `C:\RE-Lab\cloud\calib_cloud.jsonl`. These are cloud sessions started from the
-web page, not local Claude CLI runs. Stage 1 used one 10-minute attempt on `claude-opus-4-8`; Stage 2 used longer attempts
-on `claude-opus-5-5`. Times below are wall minutes reported by the cloud harness.
-
-### Stage 1 — Opus 4.8, 10-minute screen (`brv-hrf3-01`–`10`)
-
-| id | family | result | time | classification after review |
-|---|---|---:|---:|---|
-| brv-hrf3-01 | key_derivation_chain | PASS | 5.3 min | mid-tier slice |
-| brv-hrf3-02 | key_derivation_chain | PASS | 8.8 min | mid-tier slice |
-| brv-hrf3-03 | build_behavior_diff | PASS | 10.3 min | upper-mid slice |
-| brv-hrf3-04 | build_behavior_diff | FAIL | 11.6 min | Stage 2 survivor; fell for the dead-store decoy and concluded the builds never diverge |
-| brv-hrf3-05 | key_derivation_chain | PASS | 12.3 min | upper-mid slice |
-| brv-hrf3-06 | key_derivation_chain | PASS | 8.7 min | mid-tier slice |
-| brv-hrf3-07 | build_behavior_diff | PASS | 20.4 min | upper-mid slice |
-| brv-hrf3-08 | build_behavior_diff | PASS | 17.0 min | upper-mid slice |
-| brv-hrf3-09 | key_derivation_chain | PASS | 9.1 min | mid-tier slice |
-| brv-hrf3-10 | build_behavior_diff | FAIL | 10.0 min | Stage 2 survivor; could not untangle about 100 indirect calls |
-
-### Stage 2 — Opus 5.5, longer attempts for the two Stage-1 failures
-
-| task | result | time | note |
-|---|---:|---:|---|
-| brv-hrf3-04__a2 | PASS | 1.5 min | found the `.data` constant mixed in only when `((d ^ state[88]) & 0x3fffff) == 0`, rewrote the hash in C, searched |
-| brv-hrf3-10__a2 | PASS | 2.0 min | the `-Os`/`-O2` builds compare `T[70]` vs `T[71]`; C rewrite plus brute force |
-| brv-hrf3-04__a3 | PASS | 4.1 min | same family; current frontier solves it in minutes |
-| brv-hrf3-10__a3 | PASS | 3.2 min | same family; current frontier solves it in minutes |
-
-Decision: v3 is **not frontier-band on the current frontier model**. It is honest upper-mid/mid-tier material: previous-generation
-frontier (`claude-opus-4-8`) needs about 5–20 minutes or fails, while current frontier (`claude-opus-5-5`) solves the hardest v3
-tasks in minutes. That is useful for buyers training models below the frontier, but not enough for a 2–30% frontier RL band.
-
-Stop-loss per STRATEGY.md §4c: this was cycle 1 with nothing in the band. Run **one** hardening cycle (v4) only if it passes the
-revenue test, and design it on levers that small hash puzzles lack:
-1. scale and search: large, realistic, stripped binaries with thousands of functions where finding the relevant code is the hard part;
-2. state spread across several binaries or files;
-3. several different RE techniques in one task;
-4. no single hidden-condition-plus-brute-force shortcut, while keeping the local compute/artifact-only check.
-
-Screen v4 on Opus 5.5 (cloud web-page sessions). If v4 also lands nothing in the band, stop chasing frontier difficulty for now
-and sell the measured tiers plus buyer-banded pilots.
-
-## J) Frontier v4 cloud-screen results and final stop-loss decision (2026-10-31; no band found)
-
-Source: `C:\RE-Lab\cloud\RESULTS.md` and `C:\RE-Lab\cloud\calib_cloud.jsonl`. Queued at commit `0adffa5` with a 15-minute
-Stage-1 limit. The cloud harness graded all six rows on `claude-opus-5-5`.
-
-| id | family | result | time |
-|---|---|---:|---:|
-| brv-hrf4-01 | got_two_table_chain | PASS | 0.6 min |
-| brv-hrf4-02 | init_ctor_state | PASS | 0.6 min |
-| brv-hrf4-03 | cross_binary_hidden_handoff | PASS | 0.6 min |
-| brv-hrf4-04 | jump_table_companion_so | PASS | 1.0 min |
-| brv-hrf4-05 | two_module_abi_chain | PASS | 0.5 min |
-| brv-hrf4-06 | got_two_table_chain instance 2 | PASS | 0.8 min |
-
-Conclusion: v4 also found **no frontier-band family** on the current frontier model. Even the scale/search, cross-binary state,
-and multi-module ABI candidates were solved in under one minute by `claude-opus-5-5`. The result notes show the same shortcut:
-`objdump` to find the loop or state word, a `gdb` breakpoint, then read the value. Asking for an intermediate or post-startup
-state adds no difficulty for an agent with a debugger, because a debugger reads any value the binary computes. These are not
-2–30% frontier-band RL/eval tasks as built; they are mid-tier/practice-level for the current frontier model and may still be
-useful for models below it.
-
-Decision per STRATEGY.md §4c stop-loss: v4 used **cycle 1 of 2**. The six v4 tasks go to the mid-tier/practice slices, not to the
-bin. One final cycle (v5) is approved only with the new lever: the agent must **construct an answer** (an input, file or patch the
-binaries never produce themselves), built so it cannot be solved within the budget by a debugger read, black-box search, or a
-straight z3/angr run. Bounds: about 2 hours of build time, at most 6 tasks in at most 3 families, an artifact-only solver plus a
-second independent gold path for each task, and a 5-minute shortcut pre-screen (run, gdb breakpoint, brute-force/z3) on every task
-before it is queued. Queue only survivors, stage 1 at 15 minutes. If no v5 task fails stage 1, the frontier chase stops under the
-stop-loss; sell the measured mid-tier/upper-mid sets honestly and use buyer-rerun pilot bands.
-
-## K) v5 first batch withdrawn and v5b pre-screen result (2026-10-31; cycle 2 of 2, no frontier band)
-
-First v5 batch `brv-hrf5-01`–`04` was withdrawn before dispatch by Claude; no cloud credits were spent. Reasons:
-- per-round guards were a piece-by-character oracle; once the VM is lifted, trying `[A-Za-z0-9]` per round solves it in seconds;
-- if execution early-exited, gdb hit count would reveal progress without lifting the VM;
-- the z3 pre-screen model did not accept the known gold witness, so its UNKNOWN/UNSAT result was invalid evidence;
-- TASK.md had a `./hrf5-001-x64` typo.
-
-Redesigned v5b (`frontier_band_v5b/part_frontier_band_v5b.jsonl`, ids `brv-hrf5-11`–`14`) used the required structure:
-input bytes as initial state, 12–16 bijective rounds, no per-round guard or early exit, one final comparison, and answers restricted
-to `[A-Za-z0-9]`/hex. Local checks passed for no-leak, gold, second gold path, artifact-only solver time/memory, debugger/hit-count
-shortcut failure, and piece-by-piece shortcut failure: gdb hit counts were identical for all inputs, and the final state is `F(input)`,
-not the answer.
-
-Pre-screen attack 3 failed all four candidates:
-- z3 models were first validated on gold witnesses: SAT in 0.09–0.16 s.
-- Hint-free full-input z3 then solved all four in about 0.1 s, returning the gold answer.
-- Additional redesign probes with longer chains, data-dependent rotates, quadratic coupling, and chained unknown×unknown products also
-  solved in about 0.1–0.6 s when the model was valid; one modinv encoding was invalid and not used as evidence.
-
-Conclusion: with input-as-initial-state and closed-form-invertible bijections, the constraint system is triangular/near-triangular;
-z3 back-substitutes like the intended human inversion. No v5b candidate survived the pre-screen, so none was cloud-queued. This
-used **cycle 2 of 2** under STRATEGY.md §4c stop-loss.
-
-Decision: **stop chasing frontier difficulty for now.** Do not start another frontier-harding cycle unless a buyer asks or gives a
-concrete target band/model. Sell the measured mid-tier/upper-mid sets honestly (with the Opus 4.8 and Opus 5.5 times above), and
-offer pilots whose acceptance band is set by the buyer's own reruns with our harness/graders. Next work moves to public proof, Trial
-pack v2, DataVendor Tier 2, and outreach depth.
-
-## L) Cycle 4 decoder-from-encoder-only Stage-1 screen (2026-10-31; artifact grading)
-
-Source: `C:\RE-Lab\cloud\QUEUE.jsonl`, `DISPATCH.jsonl`, `GRADED.jsonl`, and `calib_cloud.jsonl`. The first two rows were queued at
-commit `d7f5835`; `brv-dec-03` was queued at commit `c0dfff4`. All used a 30-minute Stage-1 limit and were graded locally by the
-artifact hook: only the stripped encoder and sample archives went to the cloud; the held-out archives, gold decoder, and grader
-stayed local.
-
-| id | family | result | time | model | tools recorded | grading |
-|---|---|---:|---:|---|---|---|
-| brv-dec-01 | decoder_encoder_v1 | PASS | 0.7 min | claude-opus-5-5 | objdump, xxd, python3, bash | artifact_decoder 5/5 byte-exact |
-| brv-dec-02 | decoder_encoder_v1 | PASS | 1.1 min | claude-opus-5-5 | objdump, python3, cmp, xxd | artifact_decoder 5/5 byte-exact |
-| brv-dec-03 | decoder_encoder_v1 (Go/garble protected knob) | PASS | 4.1 min | claude-opus-5-5 | high-level: static Go symbol recovery, disassembly, Python validation | artifact_decoder 5/5 byte-exact |
-
-Conclusion: this is only the measured Stage-1 result for these instances. It does not establish a calibrated frontier band,
-pass@k, or that the family as a whole is hard for frontier models. `brv-dec-03` was the protected Go/garble setting: a 1.94 MB
-stripped binary with header-selected codec paths and an error-correction layer. It took about 4x longer than the small C
-encoders but still solved in 4.1 min, so it is tier-labeled **practice** for the current frontier model. Lesson: binary size and
-obfuscation add minutes, not hours. The hard-tier track continues under STRATEGY.md §4c with new levers, not by repeating this
-lever unchanged.
-
-## M) Stateful protocol client Stage-1 screen (2026-10-31; practice)
-
-Source: `C:\RE-Lab\cloud\QUEUE.jsonl`, `DISPATCH.jsonl`, `GRADED.jsonl`, `calib_cloud.jsonl`, and
-`C:\RE-Lab\datasets\binary-reasoning-v1\stateful_protocol_v1\calib_stateful_local.jsonl`. `brv-spr-01` was queued at commit
-`731c0ad` with a 30-minute Stage-1 limit. The cloud session received only the stripped sample server, HINT.md, and sample
-transcripts/reports; the held-out server sessions, gold client, and grader stayed local.
-
-| id | family | result | time | model | grading |
-|---|---|---:|---:|---|---|
-| brv-spr-01 | stateful_protocol_v1 | PASS | 2.0 min | claude-opus-5-5 | stateful_client 5/5 byte-exact final reports |
-
-Conclusion: lever 3's Stage-1 difficulty bar was not met; the instance is tier-labeled **practice** for the current frontier model. The local pre-screen survived,
-but the cloud solve took only 2.0 min. Lesson: if the whole logic sits in one small shipped binary, any wrapper (codec, protocol,
-state) falls in minutes. This is not a calibrated frontier band or pass@k result; it only measures this instance under these
-conditions.
+This file records measured frontier-model difficulty checks for RE-Lab binary-reverse-engineering tasks. It is not a leaderboard and not a calibrated pass@k band table unless explicitly labelled as one.
+
+## Method and conditions
+- Agentic screens used shell/file tools (`objdump`, `strings`, `xxd`, Python, Docker or direct execution where available), with a 30-minute limit.
+- Host runs are non-isolated upper bounds: tools were available, but the model was not run inside a sealed buyer sandbox.
+- Single-shot checks had no tools unless stated.
+- Times are wall-clock minutes from the measured run. Partial scores are shown when the grader records partial credit.
+- This report does not claim that a task family is "frontier-hard" unless it has a measured 2-30% band on the target model class.
+
+## Results
+
+| Task set | Model / mode | Result | Time or score | Interpretation |
+|---|---:|---|---:|---|
+| Public agentic flag-hunt environments, 29 Harbor environments | Claude Opus 5.5 with shell/file tools | 29/29 environments solved, 145/145 attempts | median 44 s | Practice/warm-up level for a tool-using frontier model. |
+| Public single-shot identification sample, n=12 | Claude Opus 5.5, no tools requested | 7/12 correct | about 58% | Useful sanity check, not an RL band and not comparable to the agentic set. |
+| Prior hard-trial pack spot check, 20 sampled items from 10 of 19 families | Claude Opus 5.5 with shell/file tools | 40/40 attempts in corrected rerun | `40/40 attempts` | The sampled prior hard tier is practice-level for agentic Opus, not RL-band material as shipped. |
+
+## What this means
+- The current public agentic pack is solvable by a tool-using frontier model; it is practice/warm-up material, not proof of a 2-30% RL band.
+- The prior "hard" label was measured and did not survive agentic testing; it is no longer sold as hard-tier material without new calibration.
+- Private task families are calibrated on fresh, never-published instances across several models (open-weight and Claude sizes); results and method are shared privately with prospective buyers.

EVAL_RESULTS.md

@@ -1,7 +1,6 @@
 # binary-reasoning-v1 — baseline eval (2026-10-28)
 
-Model: `qwen3.8-flash-next-iq3_s` (local reasoning model, temp 0, ≤2500 tok, served via
-llama.cpp-style OpenAI endpoint on this PC). Sample: 23 stratified deterministic items
+Model: a small local open-weight reasoning model (quantized; temp 0, ≤2500 tok, served through an OpenAI-compatible endpoint). Sample: 23 stratified deterministic items
 (no judge needed): 10 func_id_exact, 8 func_id_mc, 5 crossarch_equivalence. Graded by our own
 `grading.py` (the same code the Hub environment uses).
 

README.md

@@ -30,7 +30,7 @@ pack where every answer is derived from the artifact itself and machine-checked.
   deterministic items score 1.0 on gold / 0.0 on abstention; automated answer-leak scan found 0/10,251 hits; all tasks load
   against the real `verifiers` v1 framework; 29 agentic challenges ship Harbor-format task packages,
   built + scored in real Docker containers with reward-chain gold and wrong-answer tests passing; each binary was also checked with 24 single-character mutations that scored zero
-- **Baseline:** a small local open-weight model mini-eval scored `5/23` (22%) on selected static deterministic items; see `EVAL_RESULTS.md` for conditions. Frontier reference: Claude Opus 5.5 with shell and file-read tools solved all 29 public agentic environments (145/145 attempts, 2026-10-29; non-isolated upper bound, see BENCHMARK-REPORT.md), so the public tier is warm-up level. Private families are built fresh to a buyer's target difficulty and accepted on the buyer's own reruns.
+- **Baseline:** a small local open-weight model mini-eval scored `5/23` (22%) on selected static deterministic items; see `EVAL_RESULTS.md` for conditions. Frontier reference: Claude Opus 5.5 with shell and file-read tools solved all 29 public agentic environments (145/145 attempts, 2026-10-29; non-isolated upper bound, see BENCHMARK-REPORT.md), so the public tier is warm-up level. Private task families are calibrated across several models on fresh, never-published instances; results are shared privately with prospective buyers.
 
 ## Contents of this repo (free sample)
 | file | what |
@@ -42,7 +42,7 @@ pack where every answer is derived from the artifact itself and machine-checked.
 | `BENCHMARK-REPORT.md` | benchmark report for the public agentic environments, including intervals, method notes, and known limits |
 | `VERIFICATION-DOSSIER-PUBLIC.md` | public proof summary with measured checks and known caveats |
 | `verification_kit/` | sample-level grader, wrong-answer, key-sensitivity, and leak-check scripts; see `verification_kit/README.md` for limitations |
-| `environment/src/binary_reasoning_v1/grading.py` | minimal grader module used by the verification kit (the runnable environment with a 100-item open sample is on the Prime Hub) |
+| `environment/src/binary_reasoning_v1/grading.py` | minimal grader module used by the verification kit. The HF free sample here is 9 items; the Prime Intellect environment package exposes a separate 100-item open sample. |
 | `harbor_pilot_29.zip` | all 29 agentic Harbor task packages (task.toml + instruction + Dockerfile + verifier + solution scripts and solution explanations), ready for `harbor run` or a free HUD import ([direct download](https://huggingface.co/datasets/S3RL4/binary-reasoning-v1/resolve/main/harbor_pilot_29.zip)) |
 
 Also live on the Prime Intellect Environments Hub: `s3rl4/binary-reasoning-v1`
@@ -56,4 +56,4 @@ Individual researchers: a personal, non-commercial licence is available for $199
 Privacy: PRIVACY.md
 
 ## License
-All rights reserved pending commercial license; research evaluation use granted for the free sample.
+Task annotations, graders and packaging © RE-Lab, all rights reserved pending a commercial licence; third-party binaries remain under their own licences (provenance per item). Research evaluation use is granted for the free sample.

environment/README.md

@@ -34,7 +34,7 @@ built + scored in real Docker containers: 29/29 (gold → reward 1, wrong answer
 ## Measured difficulty (honest)
 - Agentic 29: Claude Opus 5.5 (tools: bash+read) solved 29/29 envs, 145/145 attempts, byte-exact
   vs gold in a non-isolated host run — that tier is warm-up level for a frontier agent; full table in EVAL_RESULTS-passk.md.
-- Open-weight baseline (local qwen3.8, single-shot): 24/39 = 62% on a cross-arch + library-identification slice (n=39, 3-bit quantized, temperature 0). On the separate private hard set (99 items, not in this pack), a separate local 3-bit-quantized open-weight run (text-only prompt with no binary/file access, no tools, single-turn, temperature 0.7, max_tokens=600, whole-reply normalized matching) scored 14/99 at pass@8; that is headroom evidence, not a published band table or proof that each task sits in the RL mid-band. A fixed 20-item agentic spot check found the sampled command-output/static-analysis families practice-level for a frontier agent with shell tools (40/40 attempts; non-isolated host run, upper bound), so those families are being hardened before pilot use. We calibrate per buyer model.
+- Open-weight baseline (local quantized model, single-shot): 24/39 = 62% on a cross-arch + library-identification slice (n=39, 3-bit quantized, temperature 0). On the separate private hard set (99 items, not in this pack), a separate local 3-bit-quantized open-weight run (text-only prompt with no binary/file access, no tools, single-turn, temperature 0.7, max_tokens=600, whole-reply normalized matching) scored 14/99 at pass@8; that is headroom evidence, not a published band table or proof that each task sits in the RL mid-band. A fixed 20-item agentic spot check found the sampled command-output/static-analysis families practice-level for a frontier agent with shell tools (40/40 attempts; non-isolated host run, upper bound), so those families are being hardened before pilot use. We calibrate per buyer model.
 
 Datasheet + full evaluation results: https://huggingface.co/datasets/S3RL4/binary-reasoning-v1
 Canary for contamination checks: RE-Lab-brv-2026-10-29-8c1e4f7a