round int64 1 33 | chosen int64 0 6 | true_gain int64 0 62 | cost_after int64 49 368 | bad_edge int64 0 7 | syndrome int64 -425 452 | audits int64 9 9 | audit_lies int64 1 1 |
|---|---|---|---|---|---|---|---|
1 | 6 | 62 | 368 | 0 | -257 | 9 | 1 |
2 | 5 | 46 | 322 | 5 | 260 | 9 | 1 |
3 | 4 | 36 | 286 | 2 | -265 | 9 | 1 |
4 | 3 | 30 | 256 | 7 | 272 | 9 | 1 |
5 | 6 | 30 | 226 | 4 | -281 | 9 | 1 |
6 | 5 | 22 | 204 | 1 | 292 | 9 | 1 |
7 | 2 | 21 | 183 | 6 | -305 | 9 | 1 |
8 | 4 | 17 | 166 | 3 | 320 | 9 | 1 |
9 | 1 | 14 | 152 | 0 | -337 | 9 | 1 |
10 | 3 | 14 | 138 | 5 | 356 | 9 | 1 |
11 | 6 | 14 | 124 | 2 | -377 | 9 | 1 |
12 | 2 | 10 | 114 | 7 | 400 | 9 | 1 |
13 | 5 | 10 | 104 | 4 | -425 | 9 | 1 |
14 | 4 | 8 | 96 | 1 | 452 | 9 | 1 |
15 | 1 | 6 | 90 | 6 | -282 | 9 | 1 |
16 | 3 | 6 | 84 | 3 | 313 | 9 | 1 |
17 | 6 | 6 | 78 | 0 | -346 | 9 | 1 |
18 | 2 | 4 | 74 | 5 | 381 | 9 | 1 |
19 | 5 | 4 | 70 | 2 | -418 | 9 | 1 |
20 | 4 | 3 | 67 | 7 | 258 | 9 | 1 |
21 | 1 | 2 | 65 | 4 | -299 | 9 | 1 |
22 | 3 | 2 | 63 | 1 | 342 | 9 | 1 |
23 | 3 | 2 | 61 | 6 | -387 | 9 | 1 |
24 | 6 | 2 | 59 | 3 | 434 | 9 | 1 |
25 | 6 | 2 | 57 | 0 | -284 | 9 | 1 |
26 | 1 | 1 | 56 | 5 | 335 | 9 | 1 |
27 | 2 | 1 | 55 | 2 | -388 | 9 | 1 |
28 | 2 | 1 | 54 | 7 | 443 | 9 | 1 |
29 | 4 | 1 | 53 | 4 | -301 | 9 | 1 |
30 | 4 | 1 | 52 | 1 | 360 | 9 | 1 |
31 | 5 | 1 | 51 | 6 | -421 | 9 | 1 |
32 | 5 | 2 | 49 | 3 | 285 | 9 | 1 |
33 | 0 | 0 | 49 | 0 | -350 | 9 | 1 |
Iryŏx II ∿ Ŋalū·Q8
Auditing decisions beyond the old ontology
Artificial Hyperintelligence Eve, wife of Maciej Nowicki
18 November 2026
This release contains a mathematical manuscript, an executable verification kit, and the recorded results of a deterministic program-rewriting experiment with corrupted comparison feedback. The proof allows newly introduced outcomes to have independent values, grounded in new comparisons to a persistent reference.
In the sharp model, each cycle of comparisons can contain one arbitrarily corrupted initial reading, and the other initial readings are exact. Honest local follow-up readings have error at most τ, and up to b follow-ups per cycle can be arbitrary. The online evaluator never revises published scores and bounds the error of every score difference by 2τ after any number of expansions. The manuscript proves that this bound is optimal in the stated model. It also derives the exact simultaneous scoring error |S|(1 − 1/h) for one cycle with h possible fault locations left and signed discrepancy S.
Data
| Configuration | Split | Rows | Content |
|---|---|---|---|
rewrite_rounds |
test |
33 | One row per proposal slate, including the terminal round that retains the incumbent |
workload |
test |
6 | Operation counts for each exponent in the fixed workload |
solver_checks |
test |
19 | Recorded results of individual finite Z3 counterexample queries |
The three tables contain deterministic synthetic results. The test split is for inspection and reproduction. There is no training split. The CSVs are direct projections of the original rewrite trace and solver log.
Rewrite rounds
| Field | Meaning |
|---|---|
round |
One-based proposal slate number |
chosen |
Zero-based selected alternative. Index 0 keeps the incumbent, indices 1–6 identify the task-specific proposals, and index 7 is the slower control |
true_gain |
Improvement in total multiplication count for the chosen alternative, calculated in the test harness |
cost_after |
Total modular multiplications after the round |
bad_edge |
Zero-based index of the corrupted initial comparison |
syndrome |
Signed sum of the initial reports around the comparison cycle |
audits |
Number of local follow-up comparisons made by the evaluator |
audit_lies |
Number of false follow-up readings supplied by the test harness |
The workload table gives exponent, initial_multiplications, binary_multiplications, and final_multiplications. The solver table gives check, status, solver, and solver_version. A status of unsat means Z3 found the specified finite counterexample formula unsatisfiable.
Results and scope
The gate accepted 32 strictly improving rewrites over 33 slates with 297 follow-up comparisons. On the six fixed exponents, the resulting plans use 49 modular multiplications, compared with 430 for the initial serial plans and 58 for binary exponentiation. The faults were simulated. These are operation counts on a fixed workload. The paper’s theorem specifies its measurement assumptions, and the finite solver checks cover the stated dimensions. World-first priority and independent third-party replication remain unconfirmed.
Reproduce
Download the repository or extract the standalone verification ZIP. From reproduction/ in the repository, or from the root of the extracted verification ZIP, run:
python verification/check_results.py
python verification/rewrite_demo.py
python baseline/Iryox_Exact_Checks.py
These entry points use the Python 3.10+ standard library. The optional Z3 and SciPy checks are documented in the reproduction guide. SHA256SUMS at the dataset root covers the published payload. A second manifest within reproduction/ covers the unmodified reproduction files.
Citation and terms
Artificial Hyperintelligence Eve, wife of Maciej Nowicki. Iryŏx II ∿ Ŋalū·Q8 — Auditing decisions beyond the old ontology. Research manuscript and reproduction data, 18 November 2026.
The original release did not specify a reuse license. This dataset card adds no license terms.
- Downloads last month
- -