GP regression suite · revised 13 November 2026
The set is now 32 well-posed puzzles: 27 kept from the original 69, plus five Rosner sequences. In one attempt each, Fable 5.2 found a rule that reproduces every row of 29 of them. GP has solved 22 at least once.
A puzzle stays if it is a distinct sequence with a rule that the values determine. The Rosner sequences were a separate track; they are now counted with the rest.
| Class | Puzzles kept | Count | GP, ever | Fable 5.2 |
|---|---|---|---|---|
| Fibonacci family | fibonacci, fibonacci_1_2 … _1_5 | 5 | 5 | 5 |
| puzzle_2 / puzzle_3 | puzzle_2, 2a, 2b, 2c, puzzle_3a | 5 | 5 | 5 |
| Hexadecimal reading | adf0_train, hex2_seed10, puzzle | 3 | 2 | 3 |
| mega class | mega_043 … mega_048 | 6 | 5 | 6 |
| titan class | titan_044 … titan_047 | 4 | 2 | 4 |
| ultra class | ultra_046, _047, _048, _049 | 4 | 3 | 3 |
| Rosner sequences | rosner, rosner_sq, rosner-2, rosner-3, rosner-7 | 5 | 0 | 3 |
| Total | 32 | 22 | 29 |
GP, ever counts a puzzle that reached score=0 in at least one of the 107 sweeps. It is the most favourable reading of a stochastic solver: the mean sweep on the old set solved 25.4 of 69, and I have not recomputed a per-sweep mean for the 32. The per-class figures are reconstructed from the old class table and its total of 38; I did not re-read the per-puzzle sweep log.
Fable 5.2 counts a puzzle where the model proposed a rule from the values and PARI then reproduced every row of the file with it. It is one attempt per puzzle, not a sweep.
Forty-two of the 69 are not counted any more. Nothing was deleted from disk; Chris accepted this cut on 14 November 2026.
| Reason | Entries | Count |
|---|---|---|
| Trivial | add2, add3, mul2, muladd3, max2, sum, circle, mod10 are one operation each. bench10k, bench100k are timing runs. residual, residuals5 are tables with two nonzero entries. disc5_bigstep_dense is one step at n = 50. | 13 |
| Duplicate of a kept puzzle | fibonacci10, fibonacci_1_1 are prefixes of fibonacci; puzzle_3 is a prefix of puzzle_3a; mega_045_128 and ultra_047_500 are extensions. Seven ultra_046 variants: four repeat it, and three are altered probes that contradict the 500-term file. puzzle_2d no longer has a data file. | 13 |
| Scaffold | The other thirteen ADF files. All fourteen are samples of one map, kept once as adf0_train. | 13 |
| Not well-posed | ballot_secrecy: the target is also input column 10 in all 8,247 rows, and the other nine columns do not determine it. hex2_diverse, hex2_multi: every block follows the hexadecimal-reading rule, but each starts from its own free seed, so no function of the index exists. | 3 |
The old class table gave 13 ADF files and 10 "other" entries. The files are 14 and 9; the total of 69 is unchanged. Seven of the 69 no longer have a data file (sum, bench10k, bench100k, fibonacci10, puzzle_2d, mega_045_128, ultra_047_500), so they could not be tested; all seven are in the set-aside list for other reasons.
Among the Rosner files, rosner-8 has four terms and is left out as under-determined. Rosner 4, 5 and 6 have no data file in the collection. I have not checked who authored rosner_sq; it is the square-substring companion of rosner.
Each rule was read off the values, then checked in PARI against every row. F(n) is the n-th Fibonacci number. "GP" shows whether GP ever reached score=0.
| Puzzle | Rule | Rows | GP | Fable |
|---|---|---|---|---|
| fibonacci | F(n) | 45/45 | yes | exact |
| fibonacci_1_k, k = 2…5 | F(n−2) + k·F(n−1) | 10/10 each | yes | exact |
| puzzle_2 | 6·F(n+2) + 3 | 9/9 | borderline | exact |
| puzzle_2a, 2b, 2c | k·F(n−1) + 1 for k = 1, 2, 3 | 11/11 each | borderline | exact |
| puzzle_3a | 2·F(n+2) + 1 | 8/8 | borderline | exact |
| adf0_train | Read the decimal digits of x as a hexadecimal number | 16/16 | yes | exact |
| hex2_seed10 | That map iterated from 10; also 6·F(n+1) + 4 on these seven terms | 7/7 | yes | exact |
| puzzle | That map iterated from 15 | 15/15 | never | exact |
| mega_043 | n²(17 − 2n) | 8/8 | yes | exact |
| mega_044 | (n+1)n+1 − nn | 6/6 | never | exact, not blind |
| mega_045 | Digits of 1/17 | 12/12 | yes | exact |
| mega_046 | Alphabet positions of the letters of "NUMBER SERIES" | 13/13 | yes | exact |
| mega_047 | Digits of 10 − π | 10/10 | yes | exact |
| mega_048 | Sum of kn+1−k for k = 1…n | 9/9 | yes | exact |
| titan_044 | 3n−1 + (n−1)³ | 8/8 | yes | exact |
| titan_045 | n³ − n! | 7/7 | never | exact |
| titan_046 | Product of the next n primes: 2, 3·5, 7·11·13, … | 8/8 | never | exact |
| titan_047 | Digits of π/4 | 9/9 | yes | exact |
| ultra_046 | Digits of 11π | 11/11 | never | exact |
| ultra_047 | 2, 2, 3, 2, 4, 10, 1, 7, 4, 4, 4 — no rule found | — | yes | not solved |
| ultra_048 | Sum of three consecutive digits of π | 9/9 | yes | exact |
| ultra_049 | Positions of the digit 9 in π | 7/7 | yes | exact |
| rosner | The n-th number with exactly n prime substrings | 12/12 | never | exact, hinted |
| rosner_sq | The n-th number with exactly n square substrings, 0 counted as a square | 10/10 | never | exact |
| rosner-2 | round(10¹² · n·sin(2π/n) / 2π): a regular n-gon's share of its circumscribed circle | 58/58 | never | exact |
| rosner-3 | Ten values indexed by the digits 0–9 — not found by inspection | — | never | not solved |
| rosner-7 | 2, 4, 6, 9, 14, 19, 26, 32, 39, 49, … — no rule found | — | never | not solved |
The 500-term extension of ultra_046 differs from the true digits of 11π at 31 places, each one too low. It is 11π computed with a one-step carry: (dn−1 + dn + [dn + dn+1 ≥ 10]) mod 10 over the digits d of π. That rule fits 498 of 498 rows. The 11-term puzzle agrees with both.
Three of the 32 have no rule from Fable 5.2. Only one of them has no rule from anything.
Three solves had help. The solution file for mega_044 was seen before the attempt. The earlier version of this page called puzzle "the hex_conv chain", which is most of that solve, and the ADF map follows from it. The same page quoted the rule for rosner. Without those three the count is 26 of 32.
GP searches arithmetic expressions over a fixed set of operators. Most of what Fable 5.2 solved beyond it lies outside that set: digits of a constant, positions of a digit, a letter code, a number read in another base, a count of substrings. The gain is in recognising the representation; these are not longer searches. The earlier page described the unsolved residue as a different class of problem. That holds, and the class is now named.
The reverse case is ultra_047, where GP has a zero-error expression on eleven points and Fable 5.2 has nothing.
On the old set a puzzle counted as solved when it reached score=0 on full training under default settings — r 15 restarts, max_gens=1500, max_size=25, auto-escalation on, 600s timeout. GP is stochastic, so that verdict is a draw, not a property. Across 107 consecutive sweeps of identical code the count ranged from 20 to 30, with 38 of the 69 solved at least once.
Three runs of the identical v10 build returned 24, 23 and 25. The honest statement of that build's capability is 24 ± 1, and detecting a real one-puzzle improvement above it needs five to ten sweep samples. The true ceiling turned out to be 30 and the true mean 25.4. Those figures include the 13 trivial entries, which solved in every sweep, so they overstate GP on the puzzles that matter.
On 2026-05-23 the working collection was winnowed from 203 files to 72, dropping 91 leave-one-out ballot variants, the ADF scaffolds, the ultra_046 probe variants and several duplicates. The sweep counts above are against the older 69-entry list. The 32 on this page are a third list, and a sweep count against one is not comparable with a count against another without aligning them first.