Datasets:
Complete measured results — release 1.0.0
Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Status: All listed experiments and contract checks completed. Conditional gains and failures are retained. Universal intelligence multiplication and sustained recursive improvement remain unestablished.
Completeness: The current results, raw records, code, mathematical contracts, and inherited evidence are included in this archive. Scientific generality and novelty are open, not counted as completed deliverables.
1. Verification
| Check group | Cases | Result |
|---|---|---|
| Baseline preservation and schedule bounds | 5,292 | Passed |
| Persistent exploration coverage | 40 | Passed |
| Shortcut correctness and additive budget | 1,440 | Passed |
| Exact and malformed verifier inputs | 932 | Passed |
| Immutable witness admission/import | 100 entries | Passed |
| Altered import rejection | One altered entry | Rejected |
The 5,292 scheduling cases cover logical baseline and candidate durations, correct and incorrect outputs, and three baseline shares. The 40 exploration cases cover indices 0 through 9. These are finite implementation checks; the general proofs are separate. The new demo and CLI also produced and verified the inverse 452 for 239 modulo 4001.
2. New benchmark protocol
There are four generated task families, 12 development instances per family, and 48 final instances per family. The final evaluation seed base is 107200. The earlier race-only run used different evaluation seeds and is retained in initial_race_benchmark.json. The extension was designed after observing that initial run, so this is exploratory iterative research.
The candidate registry contains three fixed, hand-written experts. A supplied task-family label guides the development calibration; no unrestricted task understanding is inferred. Candidate selection minimizes mean capped logical effort on development tasks. The shortcut budget is twice the largest completed development cost for the selected candidate, between 16 and 100,000 steps.
Every direct solver is timed through the same Worker verifier interface. The portfolio counts idle and active schedule slots up to 100,000; direct solvers count up to 100,000 Worker steps; the shortcut gets its stated candidate budget plus the baseline cap. These are explicit work budgets, not equal CPU-instruction or wall-clock budgets. Actual wall time is reported separately and includes scheduler/verification overhead. The shortest-path check validates both a path and feasible potentials.
Wall timings are a single measured pass over the final instances on a shared CPU. No timing confidence interval or statistically certified speed claim is inferred. Reproduction may change ratios for microsecond-scale operations. Input generation is excluded from final execution time, while development calibration includes its own task generation and all measured candidate trials. Human/assistant research effort, earlier experiment rounds, and implementation effort are not charged to the calibration number. Thus it is not total end-to-end research cost.
3. Full comparison
All time totals below include capped attempts. A lower runtime with fewer solved tasks is not an equal-capability speedup.
| Family | Method | Verified / 48 | Total ms |
|---|---|---|---|
| inverse | baseline | 48 / 48 | 9.751 |
| inverse | fast_direct | 48 / 48 | 0.246 |
| inverse | alternate_direct | 48 / 48 | 10.794 |
| inverse | portfolio_fixed | 48 / 48 | 1.439 |
| inverse | portfolio_calibrated | 48 / 48 | 1.327 |
| inverse | strong_baseline_race | 48 / 48 | 0.557 |
| inverse | verified_shortcut | 48 / 48 | 0.286 |
| subset_dense | baseline | 48 / 48 | 13.625 |
| subset_dense | fast_direct | 48 / 48 | 3.751 |
| subset_dense | alternate_direct | 48 / 48 | 3.774 |
| subset_dense | portfolio_fixed | 48 / 48 | 17.402 |
| subset_dense | portfolio_calibrated | 48 / 48 | 16.776 |
| subset_dense | strong_baseline_race | 48 / 48 | 8.023 |
| subset_dense | verified_shortcut | 48 / 48 | 3.541 |
| subset_sparse | baseline | 43 / 48 | 572.577 |
| subset_sparse | fast_direct | 45 / 48 | 474.309 |
| subset_sparse | alternate_direct | 48 / 48 | 12.477 |
| subset_sparse | portfolio_fixed | 48 / 48 | 478.187 |
| subset_sparse | portfolio_calibrated | 48 / 48 | 111.852 |
| subset_sparse | strong_baseline_race | 48 / 48 | 109.249 |
| subset_sparse | verified_shortcut | 47 / 48 | 45.750 |
| shortest | baseline | 48 / 48 | 7.383 |
| shortest | fast_direct | 48 / 48 | 5.321 |
| shortest | alternate_direct | 48 / 48 | 7.597 |
| shortest | portfolio_fixed | 48 / 48 | 15.537 |
| shortest | portfolio_calibrated | 48 / 48 | 15.746 |
| shortest | strong_baseline_race | 48 / 48 | 10.768 |
| shortest | verified_shortcut | 48 / 48 | 5.906 |
baseline uses exhaustive modular inversion/subset search or Bellman–Ford. fast_direct uses extended Euclid, dynamic programming, or Dijkstra. alternate_direct uses brute modular inversion, meet-in-the-middle subset search, or Bellman–Ford. The raw name strong_baseline_race means the protected race used the fast_direct host; it is not necessarily the fastest host for every family. In particular, meet-in-the-middle is the strong sparse-subset comparator.
The direct specialized methods often beat the wrapper. The bounded shortcut solved 47/48 sparse-subset instances, whereas the protected portfolio and direct meet-in-the-middle solved 48/48. The one shortcut noncompletion remains a failure within its stated budget.
4. Calibration and selected shortcut
| Family | Selected expert | Candidate budget | Calibration ms | Paired completed | Baseline / shortcut | With calibration |
|---|---|---|---|---|---|---|
| inverse | 0 | 28 | 3.011 | 48 / 48 | 34.11× | 2.96× |
| subset_dense | 0 | 736 | 1.789 | 48 / 48 | 3.85× | 2.56× |
| subset_sparse | 1 | 2198 | 74.832 | 43 / 48 | 36.81× | 4.55× |
| shortest | 0 | 938 | 5.279 | 48 / 48 | 1.25× | 0.66× |
Expert 0 is fast_direct; expert 1 is alternate_direct. Ratios use only the tasks on which both baseline and shortcut completed. The full family calibration cost is charged to that paired subset. This conditioning is especially important for sparse subset sum and does not establish a full-family cost-to-target advantage. Noncompletion is censored, not an infinite multiplier.
For modular inverse and dense subset sum, all 48 paired tasks completed and the shortcut retained a gain after the measured calibration cost. For shortest paths, the gain before calibration became a loss afterward. The baseline itself is not optimal: a direct specialized solver is the required comparison in the preceding table.
5. Exact-input witness reuse
The protected portfolio first solves one pass of 48 tasks; every accepted answer is reverified for admission. The bank then performs ten actual repeated lookup passes, including complete input serialization and response reconstruction. There are 480 hits per family. The comparator below projects eleven repeats from one timed fast_direct pass, so it is a modeled repeated-execution comparison, not eleven measured direct passes.
| Family | Initial solve + admission + 10 lookups, ms | Projected 11 direct-fast passes, ms | Projected ratio |
|---|---|---|---|
| inverse | 3.538 | 2.702 | 0.76× |
| subset_dense | 20.367 | 41.261 | 2.03× |
| subset_sparse | 116.465 | 5217.400 | 44.80× |
| shortest | 30.968 | 58.526 | 1.89× |
This is exact-input memoization with verified evidence. It is not transfer to new tasks or increased intrinsic model intelligence. The modular inverse case is slower after admission/lookup overhead. The large sparse-subset ratio uses dynamic programming as the projected comparator; the much faster meet-in-the-middle comparator from the full table prevents treating that ratio as superiority over the best solver. Calibration is excluded from this cache-specific comparison and must be added for a deployment that required it.
6. Inherited v0.3 evidence
| Previous result | Scope retained |
|---|---|
| 900 examples represented by five synthetic rows plus a scalar | Constructed rank-five, fixed-feature quadratic learning family |
| 26,576 decision checks, zero mismatches | Numerical tests; exact-arithmetic preservation theorem has separate assumptions |
| 4.26× large-reader speed ratio | Favorable generated query subspace; includes preparation; shifted-query ratio 0.82× |
| MLP 84.83% → 90.49%; logistic 82.72% → 90.12% | Linear vs Nyström controllers, equal feedback and calibration-trial counts, one reused split |
| Logistic 80%-target adaptation ratio 1.70×; MLP 0.55× | One trajectory, four timings; prior fitting and policy calibration excluded |
Those are retained prior experiments, not new replications. They must not be multiplied by the current portfolio ratios. See the complete old reports and JSON records in legacy/v0_3/.
7. Outcome
The core provides verified task-answer replacement, explicit baseline preservation, bounded speculation, and reusable exact evidence. Its tests passed and its limitations are documented. The measurements do not establish a universal intelligence multiplier, new foundational theorem, or autonomous recursive self-improvement. Specialized methods, unsuccessful cases, and preparation costs materially change the conclusion.