File size: 2,212 Bytes
1c4be39
 
e67a78d
1c4be39
 
 
 
 
 
 
 
 
 
 
45dd2b3
 
1c4be39
 
 
 
 
 
 
 
62079a7
45dd2b3
 
62079a7
 
45dd2b3
62079a7
45dd2b3
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
# binary-reasoning-v1 — baseline eval (2026-10-28)

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).

| family | score | note |
|---|---|---|
| func_id_exact | 1/10 = 10% | guessing internal symbol names from ~608-lib universe |
| func_id_mc | 2/8 = 25% | at random-guess baseline (4 options) |
| crossarch_equivalence | 2/5 = 40% | includes easy frim/frip pairs + hard distractors |
| **TOTAL** | **5/23 = 22%** | |

## Interpretation (honest)
- A small local quantized model is at or near chance level on this pack. Treat it as a weak-model sanity baseline,
  not evidence that the pack is frontier-hard or calibrated for RL.
- Sample n=23 is small; treat as a sanity baseline, not a leaderboard number.
- Caveats: one local model only; some items graded from reasoning text when final content
  was truncated (we note it rather than hide it). Reproduce with `mini_eval_win.py`
  (harness in this folder; raw per-item output in `EVAL_raw.txt`).

## What we will NOT claim
No "SOTA" or cross-model leaderboard until run on 2-3 hosted models (Prime Inference or
any OpenAI-compatible key makes this a 10-minute job — same harness, swap URL/model).

## UPDATE 2026-10-29: first hosted frontier numbers measured (Claude Opus 5.5)
See **EVAL_RESULTS-frontier.md**: single-shot no-tools ID sample 7/12 (10 `func_id_exact` + 2 `func_id_mc`; ~58%, incl. 1 safeguard-flag),
agentic flag-hunt pass@1 3/3 correct with tools in ~70-80 s. Honest reading: ID-from-stripped-binaries is
still unreliable for frontier single-shot (our local 22% vs their ~58% on different samples — both far from
reliable); the flag-hunts were tool-solvable in a non-isolated run, so they are treated as warm-up/oracle checks rather than strong RL signal as shipped.
The "no cross-model leaderboard" line above is now partially unblocked (one hosted model measured);
still no SOTA claim. Re-runs used a fresh session per item ([cyber] safeguard false-flags batches).