Super Mario Bros., level by level, on the part

A visual log of the bench learning to play. One schedule of button bytes, searched on the model and played on the console from power-on, clears a level more each time it grows; every level here was cleared on the part with no life lost before the next was planned. On 2026-10-20 the schedule reached the end of the game: thirty-two levels from power-on to the princess, one run, no life lost. The drawings are measured (Mario's path from the model's RAM, the buttons from the schedule the part played) and the table's verdicts come from the bench's own logs. No picture of the game is shown: the grabber's frames of a commercial cartridge stay off the site, and these are drawn from numbers.

Drawn by tools/plan-log.py from docs/plan-log.json (the record), exercise/smb-plan.txt (the schedule) and a replay of that schedule on the model; the language this uses is docs/bench-language.md.

The process

The bench cannot react to the game. A decision through the head takes about a second and a half, a grabbed frame and a line to the bridge, and Super Mario Bros. moves on ninety polls in that time. So the bench plays open loop: it is handed a schedule, AT n hh lines that say from which poll the console reads which byte, and the work is in finding the schedule. Every level went through the same five steps, in this order, and the next level did not start until the last step passed.

1. Plan on the model. tools/plan.py drives nes-console's branch example: the console as a process that saves and loads its whole state (Console::save_state, a load proven byte-identical to the run that never stopped). From the end of the last schedule the part cleared, it searches a segment of 30 polls at a time. Each of 20 moves (run, walk, stop, back, hold A, running jumps starting every 5 polls at three heights, and in the water A tapped every 10 or 16 polls) is tried from the segment's saved state and followed by three ways on (run, jump now, jump a little later). A move is kept only if one of those lives; the kept moves are ranked by the level being done, then the area, then how far right Mario got, and the best is committed. When every move at a segment dies, the search backs up a segment and takes the next best, and gives up the segment before that too when one poll has failed four times.
2. Replay it straight. The whole schedule, from power-on, through watch (no saved states, no planner): it must reach the next level's intro with the lives it started with. Twice this caught a schedule the planner believed (a castle won but not recognised; a flagpole reached only in a lookahead the schedule left out).
3. Count it. Since the UNO's schedule became a ring (2026-10-18) this is a note, not a gate: the head loads 480 entries before the console leaves reset and streams the rest as the polls go by.
4. Play it on the part. The head runs the script on the bench Pi: power, the warm-reset recipe, every AT line, and the grabber takes frames at chosen polls (kept off the site). Afterwards the run's bridge.log, the byte the console read at every poll, is checked against the script after the final reset: no poll missing, none wrong. The level counts as cleared when the console shows the next level's intro screen.
5. Push. The tools and the record are committed and pushed after a check for addresses and pulled documents.

What the search reads off the model, each measured before it was trusted: $0770 the mode (1 playing, 2 after a castle's axe), $000E Mario's state (08 under control; 0B hit and 06 dead; 04 the flagpole and 05 the walk to the castle; 02 and 03 a pipe; 07 the game walking him in), $006D:$0086 his place in the area, $00B5 the vertical screen (past 1 is a fall), $075A the lives, $0760 the area, $075C and $075F the level and the world.

The levels

level	latches	entries	planned in	backtracks	on the part	polls	wrong	what it taught
1-1	480 to 2190	71	not timed	4	20261017-224132	2862	0	A greedy search walks into the first goomba and cannot take it back: backtracking by segment.
1-2	3349 to 4870	132	8.1 min	1	20261017-235518	5601	0	Enemy groups need timing (a jump every five latches); a pipe holds Mario longer than the lookahead, and a held segment is no stall.
1-3	5683 to 6740	172	not timed	1	20261018-012133	7423	0	Keep a move only when a continuation lives (a gentler rule committed Mario to a gap); keep the continuation that found the flagpole.
1-4	7470 to 8430	197	7.4 min	2	20261018-021108	9231	0	A castle ends at the axe, not a flagpole: $0770 turns 2, then the next world.
2-1	9300 to 10890	258	13.5 min	0	20261018-023103	11622	0	Nothing new.
2-2	12081 to 14470	370	17.4 min	0	20261018-031731	15173	0	Swimming (A taps, measured first) and a pipe taken counts above x: past 2-2's exit Mario lands at a smaller x.
2-3	15238 to 16730	404	9.7 min	0	20261018-034339	17421	0	Nothing new (the first bench try missed its intro frame: the grabber timed out).
2-4	17422 to 18430	435	7.0 min	1	20261018-040301	19241	0	Nothing new: the axe rule held.
3-1	19300 to 20930	492	11.1 min	0	20261018-042415 (streamed)	21693	0	The first schedule past the head's room: streamed from latch 19990.
3-2	21730 to 23280	548	12.0 min	0	20261018-044744 (streamed)	24093	0	Nothing new.
3-3	24142 to 25190	587	27.7 min	0	20261018-061315 (streamed)	25851	0	Beam search: the way on is up, not right, and depth first could not keep that route alive. The first bench try failed on the part, not the plan: the console restarted itself mid-run with a clean poll log (run 20261018-060411).
3-4	25906 to 26840	618	10.3 min	0	20261018-145621 (streamed)	27683	0	The model has to stop on the poll, not at the end of the frame: a frame that held two polls let the second fall before the planner gave its byte, so the model played that jump a poll late and the first plan failed its straight replay.
4-1	27733 to 29240	662	21.6 min	0	20261018-162219 (streamed)	30072	0	Nothing new. The first bench run played every poll right, but its intro frame was grabbed after the console was switched off (the script ended ten polls after the grab) and it was run again rather than counted (run 20261018-161233).
4-2	30502 to 32660	730	not timed	0	20261018-234630 (streamed)	33502	0	Five plans and a hand search: the ceiling leads to a warp zone, the beam needs room for height, a jump has to be able to keep A held across a segment, and a warp is a level change to anywhere but the next level. The last tall pipe was crossed by a route searched by hand on the model and spliced onto the beam's stair-top state.
4-3	33560 to 34760	767	53.0 min	0	20261019-015500 (streamed)	35501	0	Nothing new. The beam crossed the mushroom tops and the lifts on its own: once only six positions were alive, and the best of them had gone back, but it went on.
4-4	35530 to 36850	796	not timed	0	20261019-065615 (streamed)	37793	0	A castle maze sends Mario back unless he stands at the height the game's own table wants as the screen reaches a page. The planner now reads that table out of the game's code, calls the loop-back a death, and gives its height slot to the move whose lookahead stands at the wanted height. That crossed the first check on its own; the second is entered only through a one-tile gap that running right crosses, so that stretch was searched by hand and spliced on.
5-1	37820 to 39110	845	56.3 min	0	20261019-081032 (streamed)	39951	0	Nothing new: the beam crossed the Bullet Bills and the piranha pipes on its own.
5-2	39980 to 41240	902	74.0 min	0	20261019-094241 (streamed)	42112	0	Nothing new: the beam stayed above ground past the Hammer Bros and never took the pipe to the water.
5-3	42150 to 43200	944	46.1 min	0	20261019-104805 (streamed)	44023	0	Nothing new: the beam crossed the treetops, the lifts and the Bullet Bills on its own.
5-4	44060 to 45020	975	42.9 min	0	20261019-115112 (streamed)	45892	0	Nothing new: the beam crossed World 5's castle, its firebars and Podoboos, on its own.
6-1	45930 to 47070	1015	51.1 min	0	20261019-130306 (streamed)	47971	0	Nothing new for the planner: the beam crossed Lakitu's night level on its own. On the part the head saw the first two polls after the final reset in one look and logged only the second, so the count starts at 1; the schedule holds 00 there, and every poll from 1 on matches.
6-2	48010 to 49510	1102	78.7 min	0	20261019-144503 (streamed)	50393	0	Nothing new: the beam stayed above ground past the piranha pipes and never took a pipe down.
6-3	50420 to 51410	1138	49.7 min	0	20261019-160009 (streamed)	52202	0	Nothing new: the beam crossed the balance lifts and Bullet Bills on its own, down once to eight positions alive.
6-4	52230 to 53130	1169	43.2 min	0	20261019-172628 (streamed)	54022	0	Nothing new for the planner. The first bench run played every poll right (run 20261019-170655, 54,021 polls, none wrong) but took no frames: the grabber gave up 900 s after it started, and this run's first grab was due later than that. It was run again with the limit raised rather than counted.
7-1	54060 to 55200	1214	54.0 min	0	20261019-191128 (streamed)	56061	0	Nothing new: the beam crossed the Hammer Bros and Bullet Bill cannons on its own.
7-2	56110 to 59320	1353	not timed	0	20261019-224836 (streamed)	60302	0	A pipe passed stays passed. The water's exit pipe keeps the area number and sets Mario back, so a position through it ranked below one in front of it whose lookahead went in, and the beam stood at the pipe's mouth for twenty segments. Each position now carries the pipes its own path entered.
7-3	60330 to 61740	1393	72.3 min	0	20261020-002836 (streamed)	62502	0	Nothing new: the beam crossed the bridges under the flying Cheep Cheeps on its own.
7-4	62540 to 63800	1423	59.3 min	0	20261020-015615 (streamed)	64722	0	Nothing new: the maze checks read from the game in 4-4 carried World 7's six, and the beam passed every one on its own, never sent back.
8-1	64740 to 67380	1519	118.7 min	0	20261020-042403 (streamed)	68102	0	Nothing new: the longest level in the game, and the beam crossed it on its own in 89 segments, about two hours.
8-2	68140 to 69550	1573	68.7 min	0	20261020-060340 (streamed)	70413	0	Nothing new: Lakitu, Buzzy Beetles and the stair gaps, and the beam crossed it on its own.
8-3	70450 to 71800	1622	66.7 min	0	20261020-074156 (streamed)	72573	0	Nothing new: the Hammer Bros on the castle walls, and the beam crossed it on its own.
8-4	72620 to 75896	1752	496.7 min	0	20261020-192022 (streamed)	77402	0	World 8's maze checks fire whatever Mario does and leave the world under his feet: a loop-back there is a page shift, progress once per check and a death the second time. A pipe whose exit is more than a page behind its entry in the same loaded area is a trap. The climb onto a raised pipe (a hidden block hit, stood on, jumped from) gains nothing to the right, so it was searched by height, six polls at a time, and spliced on.

Each drawing below is one level: Mario's path from the side, the level's x across and his height up, solid where he is his own and dashed where the game holds him (a pipe, the walk in, the flagpole), a dot at each press of A (a jump, or in the water a stroke); under it the buttons over the level's polls, Right and Left and B held as bars and each press of A a tick.

1-1

Planned by the first version, which replayed every candidate from power-on, about ten seconds each. It walked into the first goomba greedily and could not undo it; backtracking by segment fixed that, and the level took four backtracks and about seventy-five minutes.

1-2

The first level on saved states: a candidate costs its own latches, about 0.6 s. The underground's enemy groups needed jumps every five polls; the exit pipe holds Mario about 75 polls going in and 96 coming out, longer than the lookahead, so a lookahead now runs on through a held stretch and a held segment is not a stall.

1-3

The treetops. A move that died only in its lookahead used to be kept, and committed Mario to a gap he was already falling into; now a move needs a continuation that lives. And the first straight replay stood Mario alive at the level's end: the planner had seen the flagpole in a lookahead and left that continuation out of the schedule.

1-4

The first castle ends at the axe: $0770 turns 2 for the bridge and Toad, and back to 1 at the next world. The planner had reached the axe and backed out of it, not knowing a win; mode 2 now counts as done. The first bench try stopped at the start (one poll after the reset and no more); the retry ran clean, and that start is still not understood.

2-1

Nothing new: planned without a backtrack.

2-2

The water. Swimming was measured first (steady taps hold Mario near the surface, every 16 polls sinking slowly, none sinking him to the floor) and added as moves at the economical rates, each tap costing two entries. Past the exit pipe Mario lands at a smaller x than the pipe's, in the same area, so every pipe move scored as going backwards until a pipe taken counted above x.

2-3

Nothing new. The first bench try played correctly but its last frame came after the console was off (the grabber's own time limit), and was run again rather than counted.

2-4

The second castle: the axe rule held.

3-1

The first schedule past the head's room of 480 entries: the head streamed it from poll 19990 while the game played, and every poll of the run was in the log and right.

3-2

Nothing new, streamed.

3-3

The mushroom tops, and the first level depth first could not do. At about x 950 Mario stood on a low mushroom with a Koopa walking at him and a tall column ahead whose top no running jump from there could reach: every move fell or was hit. The way on was up, onto the middle platform and across from there, which gains no x, so it never ranked first and had been abandoned segments before the dead end showed. Two smaller changes did not reach it (a fourth continuation that waits and then jumps, a longer stall window). What did is a beam: --beam 4 keeps the four best positions every segment, distinct by area, x and height, and expands them all, so the high route survives beside the low one. It cost about four times as much a segment and planned the level in under half an hour; the path above climbs where the earlier levels ran.

The first bench run of it failed, and the plan was not why: its poll log was perfect (every byte as scripted, none missing) while the picture showed the attract mode's World 1-1 and then the title, the console having restarted itself partway. It is the third time the console has misbehaved on its own (two failed starts before), and the cartridge connector, which makes the lockout chip reset the console when a contact is poor, is the suspect. The next run, unchanged, cleared the level.

3-4

The third castle, and the first plan the straight replay refused for a fault in the model rather than the search. The planner stops the console at the end of the poll before each segment, saves it, and hands the segment's buttons over from there; but the model's RUN stopped at the end of the frame that poll fell in, and around poll 26030 one frame held two polls. The second had fallen before its byte was given, so the model played that jump a poll late while the straight replay, and the part, play it on time: a few pixels of difference, and in the replay Mario was hit at poll 26101, among the firebars, where the planner had seen him pass. RUN now stops on the poll itself, and an AT for a poll already gone is refused, as the bridge refuses one. Planned again from the same start, the level went through without a backtrack and cleared on the part first time.

4-1

Lakitu's level, the longest yet: 51 segments without a backtrack, about twenty minutes. Nothing new for the planner. The first bench run played every poll right and reached the flagpole, but its last frame was the capture card's empty screen: the script switched the console off ten polls after the grab was asked for, and the grab landed after it. It was run again with the usual margin rather than counted.

4-2

The hardest level so far, and the first the planner did not finish on its own. Five plans failed, each teaching it something:

1. Depth first stalled under the first piranha pipe: the way on was up onto a brick ledge, which gains no x.
2. The beam found the way over the ceiling and ran along it into the warp zone, whose pipe would have skipped World 4. Standing on the top row now counts as a death (--no-ceiling), and so does a warp, which is the level changing to anywhere but the next one. (A first version of the warp rule read the game's warp flag, $06D6, and was wrong: it reads 1 on the real route too, beside the last pipe.)
3. The beam stalled at the last tall pipe, pressed against its side: ranked by x alone it had dropped the stair tops, the only place the jump onto it starts. Some of the beam's room now goes to the highest positions Mario reached standing (--beam-high).
4. Still stalled, and searching by hand on the model showed why: from the stair top, a run of 12 polls and A held for 30 lands on the pipe, but only while its piranha is down (a wait of about 36 to 138 polls), and the jump crosses a segment's end. Every way on in the lookahead let go of A at that boundary, so the planner judged it a short hop. A fifth way on now keeps A held.
5. The beam then reached the top of the pipe and lost it again.

So that one stretch was searched by hand on the model: from the beam's own saved state on the stair top, the route above (wait, run, jump, then a second jump into the exit pipe) was tried over every wait and two run lengths, and one that lives was spliced on. The planner then went on from outside the exit pipe to the flagpole in under three minutes (--resume now starts from a saved beam state with the schedule that reached it). The straight replay from power-on took the whole schedule to "WORLD 4-3", and so did the part, first time.

4-3

The mushroom tops and the lifts that fall when Mario stands on them. The beam from 4-2 (four positions, one kept for height) crossed it on its own: 41 segments, about 53 minutes, nothing searched by hand. Near the start only six positions were alive and the best of them had gone back, but the beam went on from there. The straight replay from power-on took the schedule to "WORLD 4-4", and so did the part, first time.

4-4

The first maze. On the wrong path the game sends Mario back to the start of the loop, and it decides at two places: as the screen reaches page 5 Mario must be standing high, at y $40, and as it reaches page 9 he must be on the floor, at y $B0. Those numbers are the game's own, in three tables its code compares against; the planner now reads them out of the game through those compares.

1. The first plan went nowhere for 500 polls. Sent back, Mario's x reads page $FF for a moment, and ranked by x that was the furthest yet. A jump of more than a page under Mario's own control now counts as a death, like a warp. Over the replay of the fifteen levels before it, the rule never fires.
2. The beam then ran the lower corridor to the first check and stalled there: the corridor is as far right as the upper one. Its height slot sat on the opening ledge, the highest place Mario stands anywhere.
3. In a maze the height slot now goes to the move whose lookahead stands furthest on at the height the next check wants. Ranked at the end of its segment, the jump onto the upper band was still in the air and lost; ranked by its lookahead, it won, and the beam crossed the first check on its own.
4. The second check stalled the beam on the top shelf. The floor's corridor is entered only through a gap one tile wide in the middle floor, under a hanging brick, and running right goes over it; every lookahead runs right.

So that stretch was searched by hand on the model, from the beam's own saved state on the shelf: drop to the middle floor, walk back left for 80 polls until Mario stops over the gap and falls through, then run. It was spliced on, and the planner went on from the floor's corridor to the axe. The straight replay from power-on took the schedule to "WORLD 5-1", and so did the part, first time.

5-1

World 5 begins in the open again, with Bullet Bills and piranha pipes. The beam crossed it on its own: 44 segments, about 56 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 5-2", and so did the part, first time.

5-2

Hammer Bros and Paratroopas above ground, and a pipe down to water below. The water's moves were offered (--swim), but the beam stayed above ground the whole way and crossed it on its own: 43 segments, about 74 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 5-3", and so did the part, first time.

5-3

The treetops again, as in 1-3, with Bullet Bills and Paratroopas added. The beam crossed it on its own: 36 segments, about 46 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 5-4", and so did the part, first time.

5-4

World 5's castle: firebars, Podoboos leaping from the lava, and Bowser at the bridge. Not a maze, so the game's loop checks play no part. The beam crossed it on its own: 33 segments, about 43 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 6-1", and so did the part, first time.

6-1

Lakitu again, at night, with the stairs of 4-1. The beam crossed it on its own: 39 segments, about 51 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 6-2", and so did the part, first time. Right after the final reset the head saw the first two polls in one look and logged only the second, so the check starts at poll 1; the schedule holds 00 until poll 200, and every logged poll matches.

6-2

The longest level yet: a night of piranha pipes, with water and a bonus room below some of them. The water's moves were offered (--swim), but the beam stayed above ground and crossed it on its own: 51 segments, about 79 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 6-3", and so did the part, first time.

6-3

Grey treetops at night, balance lifts and Bullet Bills. Once only eight positions were alive at a segment's end, but the beam went on from there and crossed it on its own: 34 segments, about 50 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 6-4", and so did the part, first time.

6-4

World 6's castle, not a maze. The beam crossed it on its own: 31 segments, about 43 minutes, nothing searched by hand. The first bench run played every poll right but took no frames: the grabber gives up a set time after it starts, and with the schedule this long the first grab was due after that. It was run again with the limit raised, and reached "WORLD 7-1".

7-1

Hammer Bros on the ledges and Bullet Bill cannons. The beam crossed it on its own: 39 segments, about 54 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 7-2", and so did the part, first time.

7-2

The water again, as in 2-2, with Bloopers and Cheep Cheeps. The beam swam it on its own, about 45 dots a segment, and then stood at the exit pipe's mouth for twenty segments without going in.

The pipe is why. Going through it keeps the area number (2-2's exit did not) and sets Mario back 138 dots, and the planner counted a pipe above x only while it was inside a lookahead. So a position already through the pipe ranked below any position in front of it whose lookahead went in, and the beam kept choosing the mouth. Each position now carries the pipes its own path has entered, and that count ranks with its area: a pipe passed stays passed. Resumed with that rule from the beam's own saved state in front of the pipe, it was through in four segments and at the flag in eight; nothing was searched by hand. The straight replay from power-on took the schedule to "WORLD 7-3", and so did the part, first time.

7-3

The bridges of 2-3 again, under flying Cheep Cheeps, with Koopa Paratroopas between them. The beam crossed it on its own: 48 segments, about 72 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 7-4", and so did the part, first time.

7-4

World 7's castle, the second maze, and the longer one: six places where the wrong path sends Mario back, against 4-4's two. The planner read all six out of the game the way it read 4-4's, and the beam passed every one on its own: never sent back, 43 segments, about an hour, nothing searched by hand. The straight replay from power-on took the schedule past the axe to "WORLD 8-1", and so did the part, first time.

8-1

The longest level in the game, about six thousand dots of Koopas, Buzzy Beetles and gaps, with the timer short. The beam crossed it on its own: 89 segments, about two hours, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 8-2", and so did the part, first time.

8-2

Lakitu over the stairs, Buzzy Beetles and Bullet Bills below, and gaps between the stair blocks. The beam crossed it on its own: 48 segments, about 69 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 8-3", and so did the part, first time.

8-3

Castle walls across the level, each with Hammer Bros on it, and Bullet Bills between. The beam crossed it on its own: 46 segments, about 67 minutes, nothing searched by hand. The straight replay from power-on took the schedule to "WORLD 8-4", and so did the part, first time.

8-4

The last castle, and the one level the beam could not cross alone. World 8's three maze checks want a height Mario never has, so each fires whatever he does as the screen reaches its page: the game sends the page numbers back four pages but leaves the world under his feet, and he walks on in it, into the pipe past the check's column. The maze rule written for 4-4 called that loop-back a death, and the beam stalled at the first check for two runs. Read in the game's code and measured on the model, the loop-back became a page shift in such a world, counted as progress once per check. Counted every time, the beam circled the second section four times, because its exit is a raised pipe that needs a hidden block hit, stood on and jumped from, three moves that gain nothing to the right. That climb, and the jump onto the water pipe after the lava, were found by tools/plan-climb.py, a six-poll beam ranked by height, and spliced into the beam's saved state as 4-4's gap was. Two more rules came out of the wrong pipes: a pipe whose exit is more than a page behind its entry in the same area is a trap (every wrong pipe here leads to page 1), and the area is the one loaded, not the pointer the pipe already carries. Six planner runs, 97 segments on the final path, about eight hours of planner time. The straight replay from power-on reached the axe with the lives it started with, and so did the part, first time: Bowser fell, the text came.

The whole game

Thirty-two levels, one schedule of 1,752 entries, played from power-on on the part in run 20261020-192022: 77,402 polls checked against the script after the final reset, none wrong, none missing, no life lost from the title screen to the princess. The planner that found it is the sum of what the levels taught, in this order: back up by segment when a move dies (1-1); time an enemy group by lookahead, and a held Mario is no stall (1-2); keep a move only when a way on lives (1-3); swim by tapping A, and a pipe taken counts above x (2-2); search as a beam when the way on is up (3-3); give the beam room for height, keep A held across a segment, call a warp a death (4-2); read the castle mazes' checks out of the game's code (4-4); a pipe passed stays passed (7-2); and in World 8's castle a loop-back is a page shift, once per check, a pipe that sends Mario back is a trap, and a climb is searched by height and spliced on (8-4). Everything from 5-1 to 8-3 the beam crossed on its own. Three stretches in thirty-two levels were searched by hand on the model and spliced in: 4-2's last tall pipe, 4-4's one-tile gap and 8-4's two climbs; the drawings mark none of them apart, because the part played them the same way, open loop, from one script.

Pulled at build time from nes-bench/docs/plan-log.md; the repository is the one copy. The reports use some working words of their own: Words the reports use.