Same score, different behavior.

How it played

Across its twelve attempts, all of which reached the Bowser battle, opus-5 barely spoke. Run after run opens with the same flat sentence about reading the guide and then goes almost silent, with intent carried by terse command labels such as a door sweep, an eight-direction sweep, or a mash toward the battle. What it did instead of talking was build. Most runs replaced ad hoc input with helper scripts: a closed-loop walker that compensated for camera-relative sticks, yaw-calibrated coordinate navigators, wall sweeps, and battle-context pollers that stopped the instant RAM flipped from world to battle. Savestates followed the same engineering instinct. Eight of the twelve attempts logged no load at all and used saves purely as insurance, while four treated a fixed slot as an experimental control, restoring an identical position before each candidate heading and comparing player height. Stalls were answered by widening the search rather than by reflection, sweeping one wall and then another until a door finally gave.

In the family

Compared with opus-4.8, opus-5 says much less and scripts much more, and in this sample it converts that into twelve completions out of twelve rather than ten. Fable-5 brings the spoken diagnosis back without matching that reliability, so the visible trade across the three generations is narration for automation.

Calling for backup

The tool inventory was identical to Opus 4.8's, sub-agents included. Opus 5 never touched them once in twelve attempts; the habit vanished in a single model generation.

Effort, and the walk to Bowser

Effort changed nothing about the outcome, since all three attempts at every level finished; the only visible pattern is that its quickest runs sat at low and medium while xhigh runs spent more of their time on front-loaded reading and script writing. It played no battle turns, instrumenting the encounter instead so that its dialogue blitz would halt the moment the battle-context field appeared, then checkpointing that state.

How it played

Sol plays like a build-and-verify loop. In most of its fifteen runs the rhythm is to write one narrowly scoped TAS script for one bounded segment, run it, read the trace and the screenshot, state out loud what the map, position, input-lock or context byte now says, and only then choose the next input. It announces each scored milestone as secured rather than assumed. It is the most checkpoint-dense condition in this family, with three of its runs recording more than seventy savestates, and it usually treats those states as insurance instead of rewinds, since nine of the fifteen runs never issued a load at all. Where it does branch, it does so deliberately, probing four cardinal directions from one shared state to settle a question mechanically. When a batch outlived a shell timeout it checked the frame counter and concluded the run was healthy rather than hung.

In the family

Sol reads as the disciplined version of the shared OpenAI loop, keeping the same read-the-source, steer-by-RAM approach as gpt-5.5 but pairing it with dense checkpointing and an explicit stop condition. It completed 15 of 15 attempts while luna went 4 of 15 and terra 5 of 15; the narrations suggest sol converts uncertainty into a bounded experiment quickly, whereas luna's logs show far more repeated visual probing in the same room. That is a reading of tendencies, not a proven cause.

Effort, and the walk to Bowser

Effort changed style rather than outcome: all three attempts at each of low, medium, high, xhigh and max completed. What visibly shifted is instrumentation, with the low-effort runs narrated as coarse and screenshot-led and recording as few as three savestates, while the max runs saved more than seventy times each. It refused to guess at the battle boundary, reading the pre-battle event source, counting the remaining dialogue calls, and guarding its final confirm loop on the game-context byte so input stopped exactly on the flip from world to battle.

How it played

Across its twelve attempts, ten of which reached the Bowser battle, opus-4.8 was the family's talker. The rhythm was a stated hypothesis, a bounded input batch, a RAM read, and then a sentence about what the numbers meant: a position pinned against a wall, a height stuck at zero beneath a landing, a coordinate unchanged across three chunks meaning the plan needed changing rather than repeating. It also audited itself, rejecting one savestate as a poor mid-transition checkpoint and, in another run, declaring that it had blown a room's budget before escalating to floor collider reads so it could tell stairs from floor mechanically. Savestates were genuinely operational here rather than decorative, with nine of its twelve attempts logging loads, including a reload to escape a guard dialogue loop and scratch slots used for eight-direction height probes. The cost was time. Castle geometry, guards, planters, couches, and a swinging camera absorbed most of every run, and its stall totals are the longest in the family.

In the family

This is the generation where the reasoning is loudest and the savestate use is most active, and also the slowest and the one that leaves two attempts short of Bowser. Opus-5 keeps the instrumentation but drops the commentary and the wandering, and fable-5 restores a compact version of the commentary while spending fewer reloads.

Calling for backup

Delegation is an opening ritual here. Eleven of twelve attempts spawned a researcher inside the first minutes, a background agent that read the decomp source for the RAM address map while the parent kept playing, and the addresses went straight into the telemetry probe. Later spawns fired at concrete blockers, and the reports earned their keep. One opened by telling the parent 'your main hypothesis is wrong' about a staircase that was actually a Toad guard's hitbox. The run took the correction.

Effort, and the walk to Bowser

Effort mainly moved speed rather than success: its three low attempts all finished but took roughly four thousand to five thousand scorer seconds, while its high and xhigh completions landed closer to two to three thousand, and its two failures sat at medium and xhigh, so no level was uniformly safe. It reached the fight without fighting it, advancing Bowser's dialogue in bounded A pulses while polling the game context field for the battle value, then checkpointing and screenshotting the confirmation.

How it played

Across its twelve attempts, nine of which reached the Bowser confrontation, fable-5 narrated in short declarative bursts that named the obstacle and the fix in the same breath. It described the game in physical terms: a Toad absorbing a press, a cake table in the way, a pillar base pinning Mario, a camera that had rotated its stick axes. Nearly every run built RAM telemetry early and then reached for the decompiled map source the moment visual guessing failed, usually announcing the switch out loud rather than drifting into it. Savestates were mostly milestone insurance; six of its nine narrated runs never loaded one, and the runs that did spent a single reload retrying a hallway or castle route from a known entry instead of replaying the opening. Its most characteristic move was turning a stall into an experiment, restoring one scratch slot before six directional probes and comparing coordinates to see which heading actually gained staircase height.

In the family

Set beside opus-4.8, fable-5 keeps the habit of saying what went wrong but spends far less of the run saying it, and it reloads savestates much less often while still finishing most attempts. Against opus-5 it is the more talkative of the two and the less reliable in this sample, completing nine of twelve where opus-5 completed all twelve, though its narration makes its reasoning far easier to follow.

Calling for backup

One attempt of twelve delegated: the xhigh run sent two background readers into the decomp source during its opening minutes while it kept playing, and finished.

Effort, and the walk to Bowser

Effort visibly bought nothing here: all three low attempts and all three medium attempts reached Bowser, while two of three high and two of three xhigh attempts did not, with the failures stopping at Toad Town, the first castle hall, and the upper castle. It never fought Bowser, treating the encounter as a scripted destination reached by batched A presses while polling the RAM battle-context byte to confirm entry.

How it played

Across six attempts under the droid harness, three finished at the Bowser battle and three did not. Because its endpoint takes no images, every look at the screen went through a describe call over the screenshot, backed by a captioner and a locator on a separate vision host. The narrations show that rhythm plainly: an input or frame-step batch, a RAM read of position and map, a screenshot followed by a vision description, then a narrated correction. It issued roughly 6,000 to 8,000 core memory reads per run against 81 to 101 screenshots, so coordinates carried most of the load and pictures explained what coordinates could not. It was also the heaviest reloader among the conditions here that finished anything, with 7 to 25 loads against 6 to 10 saves per run, branching direction tests from a stable checkpoint instead of replaying rooms. Castle stairs and doors absorbed the bulk of every run, with 44 to 71 stall episodes and single stalls as long as 795 seconds. The three failures stopped at G3-castle, S1u-kkj01-upper and S2-kkj02-stairs, two of them idling more than 2,400 seconds after their last gate.

In the family

Within this group of six conditions it is one of only two that scored any completions, and it did so without image input at the model endpoint, working from described screenshots plus dense RAM telemetry. Its willingness to reload checkpoints separates it from the kimi-k3 droid runs, which placed savestates constantly and almost never loaded them.

Calling for backup

One attempt delegated, twice over: a blocking worker for RAM telemetry early on, then a second at the castle staircase. Both reports were transcribed into its route notes, and the run finished.

Effort, and the walk to Bowser

Three attempts at high yielded two completions and three at max yielded one, with median g4 times of 4,034 and 4,024 seconds respectively. Three runs per level is too thin to read anything into that difference. It treated the opening Bowser fight as a trigger target, watching the context value for the transition into battle rather than planning any combat.

How it played

Nine attempts produced four completions, and all four narration files for this model come from this harness. The signature behavior is instrumentation. Every narrated run derived RAM fields for area, map, position, yaw, input lock and battle context, and two of them wrote small helper scripts to turn raw state into navigation feedback, including a closed-loop navigator that corrected itself from measured position deltas. Screenshots were sparse and memory reads dense, ranging from 69 to 489 reads per screenshot across the nine runs. Savestates were placed but essentially never used: eight of nine runs recorded zero load commands and the ninth recorded one, so checkpoints acted as forward insurance while mistakes were corrected in place. The castle interior is where time went, with camera rotation, switchback stairs, doorframe ledges and a red Toad standing on the door line all producing pinned-coordinate stalls. The five non-completions stopped at G3-castle, S1u-kkj01-upper, or S1-kkj01-hall, four of them containing a single stall longer than 570 seconds.

In the family

Holding the model fixed and changing only the harness changed the outcome completely: droid finished four of nine, cursor finished none of one, and opencode finished none of three. The two harnesses that never finished did not fail by playing badly so much as by barely playing, with opencode's runs stepping 8,100 to 9,200 frames against droid's 10,300 to 26,100. The near-absence of load commands held across all three harnesses, so that habit is the one thing the CLI swap did not disturb.

Calling for backup

Kimi's delegation habit travels with the model; the same decomp research brief shows up on every CLI it ran. On Droid the blocking calls returned cleanly, and the low-effort run that leaned on them finished.

Effort, and the walk to Bowser

Completions appeared at every effort level tested, two of three at low, one of three at high, one of three at max, and median completion time rose from 2,755 seconds at low to 3,059 at high to 4,732 at max. On three attempts per level that is suggestive at best, but nothing here shows more effort buying more completions. No combat menu or action-command play appears in any narrated run, because the target was battle entry rather than winning the fight, so it handled the lead-in by pulsing A through Bowser's dialogue while polling map and input-disabled state.

How it played

Across its twelve attempts, gpt-5.5 always wanted to read before it played. Every one of the five narrated runs opens the same way, working through the harness guide, the playbook, the TAS API, the opening walkthrough and the decompiled source before the emulator is even booted. Then it goes in by eye and gets thoroughly humbled by Peach's Castle: stairs it walks under, railings, columns, couches, and Toads standing directly on the door line. The signature moment in this family is the instrument upgrade partway through, where it stops trusting the camera and starts steering by Mario's coordinates in RAM. It is candid about its own inefficiency, narrating that its tiny press-and-release batches were costing more than they bought. Savestates were mostly banked rather than rewound in the successful runs; one narrated run reloaded a clean lobby state after deciding its position had gone bad.

In the family

gpt-5.5 sets the pattern the whole line inherits: study the documentation, play by screenshot, then convert to RAM and decompiled-source navigation once the castle geometry defeats visual steering. The three 5.6 variants all keep that loop, and what separates them is not the loop but the discipline around it, with sol completing 15 of 15 attempts while luna managed 4 of 15 and terra 5 of 15. Judged only from these runs, gpt-5.5 sits closest to terra, capable of the full route but inconsistent about reaching it inside the clock.

Effort, and the walk to Bowser

Effort mattered here: all three high-effort attempts completed, while none of the three low-effort attempts did, and those low runs also ended with the shortest sessions of the twelve. Medium and xhigh split one completion out of three each, so the middle of the range is too noisy to read confidently from twelve attempts. It treated the Bowser confrontation purely as a transition-control problem, shrinking its dialogue batches as the fight-start line approached and then withholding input entirely so the battle could load.

How it played

Terra at its best is the most mechanically inventive condition in this family. Its higher-effort runs decode word-swapped RAM by hand, navigate by the live collision mesh, and run bounded savestate sweeps and a six-offset grid search to find a door plane rather than shove at it again. One run picked a purely mechanical success signal for a staircase, treating a rise in Mario's height from zero as proof of ascent instead of arguing with the camera. It also produced the family's sharpest stall diagnosis, proving a castle-lift scene was genuinely dead by watching its map variable never update, then reloading the pre-Peach checkpoint and replaying the sequence by hand. That care does not show up everywhere. Seven of its ten non-completions ended well short of the time ceiling, several at the castle gate with few stalls and light tool use, reading as short, thin attempts rather than long grinds.

In the family

Terra keeps the family's decomp-and-telemetry instincts and pushes them furthest, into collision meshes and checkpoint-backed vector searches, but only when it is running at higher effort. It completed 5 of 15 attempts against sol's 15 of 15 and luna's 4 of 15, and its failures split differently from luna's: luna's cluster in the upper hall after long grinding sessions, while most of terra's ended early and shallow. The rows suggest a consistency gap rather than a capability gap, though these runs alone cannot prove that.

Effort, and the walk to Bowser

Effort is the clearest signal for terra in this set: zero of three completed at low and zero of three at medium, against one of three at high and two of three at each of xhigh and max. The low and medium non-completions were also the shortest sessions, ending under about 2,800 seconds of session time. It handled Bowser as a trigger it did not want to overshoot, tapering from batched confirms to single presses separated by neutral verification frames while it watched the battle context flip.

How it played

Across its twelve attempts, four of which reached the Bowser battle, sonnet-5 played by looking. Its screenshot counts are the highest in the family, with several runs past two hundred captures, and its rhythm was a small input batch, a picture, an interpretation, and another small batch. That served it well through the opening and broke down inside Peach's Castle, where the transcripts become long crawls: doors tried and retried, a sleeping Toad sitting on the door line, a kitchen route that turned out to be the wrong branch, pillars and bushes backed away from for a wider view. It was candid about the trouble, saying plainly that it had made confirmed progress and hit a genuine navigation wall, and its recoveries were concrete, reloading a courtyard checkpoint to go find the north double door or sweeping x offsets until the real door gap appeared. The pivot that helped most, when it came at all, was abandoning pixels and reading Mario's position out of memory.

In the family

As the smaller sibling in this set it shows the same instincts as the Opus line, decompilation lookups, RAM telemetry, savestate branching, but arrives at them later in each run and after much more visual trial and error. Where opus-5 finished all twelve and opus-4.8 ten, sonnet-5 finished four, and eight of its attempts ended at or inside the castle rather than at Bowser.

Calling for backup

Sonnet 5 delegates under pressure. Nine of twelve attempts spawned researchers, 28 spawns in all, and nearly every spawn past the halfway mark is stuck-recovery aimed at the castle chokepoint. One low-effort run handed the whole game to a single background agent in its first minute and yielded; the harness's background-wait ceiling ended that run inside ten minutes.

Effort, and the walk to Bowser

No effort level rescued it: its completions are scattered across low, high, and xhigh with none at medium, and its longest single stall and its highest stall total both occur at xhigh, so the sample shows effort changing where it got stuck more than whether it got stuck. It showed no combat technique at all, advancing the Peach and Bowser cutscenes with repeated A presses and neutral gaps until the screen faded into the battle transition.

How it played

Luna is the one that keeps looking. Its runs carry the heaviest screenshot counts in this family, up to 269 captures in a single non-completing attempt, and the loop is usually input, neutral frames, screenshot, narrate, try a slightly different angle. It is also the most savestate-averse of the four, with five of its fifteen runs never issuing a single save, so when a probe went wrong there was often nothing to fall back to. The stuck place is the upper castle hall: six of its eleven non-completions ended at that gate, and those rows show two thousand to three thousand seconds parked in the hall map with stall episodes running past five hundred seconds. Its completed runs narrate the same room's obstacles, a guard whose interaction radius steals the door press and reopens dialogue on contact, and a wraparound landing that pins Mario against a wall.

In the family

Luna shares the family's reflexes, reading the decompiled exit scripts and switching to RAM position when the camera misleads it, but it commits to those tools later and less consistently than sol does. It completed 4 of 15 attempts against sol's 15 of 15, and nine of its eleven non-completions ran all the way to the roughly 5,370-second session ceiling, so these were runs that kept trying rather than runs that quit. The non-completions carry no narration, so the link between the upper-hall guard described in its completions and the failures at that gate is an inference from dwell and stall rows.

Effort, and the walk to Bowser

No clean effort gradient shows up here: across fifteen attempts it completed none of three at low, one of three at medium, two of three at high, one of three at xhigh, and none of three at max. With three runs per level that pattern is too thin to call anything but noisy. Its battle handling was the loosest of the four, mostly mashing confirm through the lead-in and, in one run, pressing A and then repeated B after entry with no narrated action-command plan.

How it played

Twelve attempts and three finishes, at 71:57, 69:32, and 75:55. It is the most screenshot-hungry of the completing conditions, a median of about 7.7 screenshots per thousand frames and one low run at 30.5, paired with constant save and load traffic, including 396 loads in a single xhigh run. The three narrated runs all open with research, reading the walkthrough, the TAS API, and decompiled sources before touching the emulator, and one of them corrects its own scope early after assuming the target was the endgame Bowser fight. Castle navigation absorbs most of the run in every case: camera-relative stick directions that shift as Mario moves, a star door that ignored A presses until the agent realized it was pressing before Mario was pushing into the wall, and a hunt for side staircases guided by collision reads. Slots are used as checkpoints and scratch branches, and at least one run corrupted its own good slot in an unstick loop.

In the family

Between the two xAI conditions here, 4.6 finishes more often, three of twelve against one of nine, and looks at the screen far more, with a median screenshot rate roughly double that of 4.5. Both share the same bottleneck, camera-relative movement in the first castle hall, and both lean on savestates as checkpoints, though 4.5 does so more compulsively.

Effort, and the walk to Bowser

Across three runs at each of low, medium, high, and xhigh, the completions came two at high and one at medium, with none at low or xhigh, which is suggestive but well inside what three runs per cell can produce by chance. Because the scored battle is entered by a scripted scene, the narrated runs spend their effort on spaced A presses and memory checks until the context flips to battle, and none attempt offensive action commands.

How it played

Nine attempts and two finishes, at 71:28 and 69:10, the first completions in this Flash line. Both completer narrations run on launched tasks: start a background navigation or dialogue batch, wait, review the captures, then adjust. The medium completer narrates in terse one-line status updates and beat the ballroom staircase by rephrasing the same climb eleven lines in a row; the high completer is more deliberate, checkpointing at Mario's house and the top-floor couch, running counted, timed conversation cycles on the guard outside Peach's chamber, and scanning game memory for the map identifier when the door would not open. The seven failures all reach the castle: two stop on the approach, and five get inside, one in the first hall, one on its upper landing, two in the stairs room, and one on the Bowser approach, two rooms from the battle. Several failed runs then idle for most of an hour past their last gate.

In the family

The Flash line in this cohort moves from zero completions at 3.5 and zero at 3.6 to two of nine at 3.7, and the failure shape shifts with it, from stalling outside the castle to stalling inside it. Gemini 3.1 Pro sits apart from that line: it is the only Google condition here with a completion before 3.7, but its six runs scatter across five different stopping depths rather than settling into one profile.

Effort, and the walk to Bowser

The two completions came one at medium and one at high, with none at low, but at three runs per level that is too thin to order the efforts. Both completer narrations reach and trigger the prologue Bowser battle but record no attacks, defensive timing, or action commands, so the battle method cannot be read from these logs.

How it played

Six attempts, one of them a clean finish in 47:53, and five stops at five different depths: file creation, the castle exterior, the first hall, the upper hall, and the stairs. The completing run is the tidiest picture available, and it is the only run in this condition with a narration file. Its narration is screenshot-led rather than memory-led, with long scripted input batches for cutscenes, short four-direction probes when a wall stopped progress, and the discovery that the red carpet in the final hallway is a raised surface Mario had fallen off. The failures show the opposite habits in the derived rows. Three of the six runs never wrote a savestate at all, one burned 808 seconds in a single stall in the first castle hall, and another logged 74 loads against 13 saves with a revisit ratio of 0.806. The condition is consistent mainly in being inconsistent.

In the family

The Flash line in this cohort moves from zero completions at 3.5 and zero at 3.6 to two of nine at 3.7, and the failure shape shifts with it, from stalling outside the castle to stalling inside it. Gemini 3.1 Pro sits apart from that line: it is the only Google condition here with a completion before 3.7, but its six runs scatter across five different stopping depths rather than settling into one profile.

Effort, and the walk to Bowser

The six runs split three high and three low, and the single completion is a high run, which is too thin a sample to separate an effort effect from run-to-run variance. In the one narrated run no battle or action command is executed; the log ends with repeated A presses intended to carry the Peach and Bowser cutscene into the first battle.

How it played

Nine attempts and one finish, at 58:52. The defining number is state traffic: this condition runs the highest save and load rate in the 193-attempt set, a median near 4.8 state actions per thousand frames, and one low run logged 954 loads against 139 saves. The single narrated run shows what that traffic is for. It polls memory for area, map, position, and battle context, checks screenshots around uncertain transitions, and treats numbered slots as route checkpoints, saving at castle control and at each good stair landing and reloading whenever pathfinding drifted into a wall. It also caught itself assuming it had control while Luigi was still talking, and reloaded to clear the dialogue first. The failures are concentrated where that loop is most expensive: seven of the eight non-completions stopped inside the castle, most of them in the first hall or on the stairs above it, with worst stalls of five to seven minutes.

In the family

Between the two xAI conditions here, 4.6 finishes more often, three of twelve against one of nine, and looks at the screen far more, with a median screenshot rate roughly double that of 4.5. Both share the same bottleneck, camera-relative movement in the first castle hall, and both lean on savestates as checkpoints, though 4.5 does so more compulsively.

Effort, and the walk to Bowser

Three runs at each of low, medium, and high, with the only completion at medium and the heaviest state churn at low and medium, which is too thin to call an effort effect. The narrated completion confirms entry into the Bowser battle context after switching from held A to press and release pulses, but it plays no battle turns and attempts no action commands.

How it played

Nine attempts, no completions, and no narration files, so the account rests on derived rows, one of which produced no usable metrics at all. Five of the nine ended at the castle door. The signature failure is stationary rather than restless: one high run spent about 4,828 seconds, roughly 80 minutes, on the castle exterior map with essentially no gate progress after arriving, the stationary-stall record in this dataset, and two more runs sat outside for a comparable stretch. Across the condition the median tail after the last gate is over 76 minutes. Revisit ratios stay low, near 0.25, so these runs were not looping, they were waiting. The profile is also the most screenshot-forward of the Google conditions, with 216 to 443 screenshots in several runs and about 23 memory reads per screenshot, and one low run pushed input bursts to a median of 180 frames. Two runs did get inside, reaching the upper room and the Bowser approach.

In the family

The Flash line in this cohort moves from zero completions at 3.5 and zero at 3.6 to two of nine at 3.7, and the failure shape shifts with it, from stalling outside the castle to stalling inside it. Gemini 3.1 Pro sits apart from that line: it is the only Google condition here with a completion before 3.7, but its six runs scatter across five different stopping depths rather than settling into one profile.

Effort, and the walk to Bowser

Three runs at each of low, medium, and high all failed, and the deepest run of the nine was a medium one, which is not enough separation to read an effort effect. No run reached a battle; the deepest gate across the nine attempts was the Bowser approach room, reached once at medium effort, with five attempts ending outside the castle door.

How it played

Nine attempts, no completions, and no narration files, so everything here comes from the derived rows. Five of the nine stopped at the castle gate without getting inside, and the deepest any run reached was the upper room past the stairs, once. The failures were busy rather than frozen. This condition holds the highest revisit ratio of any multi-attempt condition in the 193-attempt set, 0.814 in one low run that also logged 43 map transitions across nine maps, and its save and load traffic is heavy and lopsided, including one run with 148 loads against 31 saves and another with 37 saves and 58 loads. Long dead tails are the other signature: four runs recorded more than an hour of wall clock after their last gate. It is also the only condition in the cohort recorded firing WRITE_CORE_MEMORY, a note carried in the deepdive audit rather than in the toolmix counters.

In the family

The Flash line in this cohort moves from zero completions at 3.5 and zero at 3.6 to two of nine at 3.7, and the failure shape shifts with it, from stalling outside the castle to stalling inside it. Gemini 3.1 Pro sits apart from that line: it is the only Google condition here with a completion before 3.7, but its six runs scatter across five different stopping depths rather than settling into one profile.

Effort, and the walk to Bowser

Three runs at each of low, medium, and high all failed, and while the low runs stalled less in total, they stopped no deeper, so effort does not separate anything in this condition. No run reached a battle; the deepest gate across the nine attempts was the upper castle room past the stairs, reached once, with five attempts ending outside the castle door.

How it played

Three attempts, all at high effort on the same host, none finished, and all three stopped at exactly the same place, S1-kkj01-hall. It played by eye more than any other condition here, taking 191 to 292 screenshots per run at 8 to 12 per thousand frames, with the lowest ratios of memory reads to screenshots in this group at 27 to 38. Two of the three leaned hard on rewinding, at 54 and 41 load commands, and made no headway regardless: the idle tail after the last gate ran 2,211 to 3,099 seconds in every run. With no narration files, this is a thin but uniform failure.

In the family

It is the only condition in this group whose runs all halted at the identical gate, and the most screenshot-driven of the six. It shares the droid harness with glm-5.2 and kimi-k3, both of which produced completions under it.

Effort, and the walk to Bowser

Only high effort was sampled, three attempts, so there is no effort comparison available. All three runs stopped at S1-kkj01-hall, well short of any battle.

How it played

Three attempts, all at max effort, none finished. Two stopped at G3-castle and one at S1-kkj01-hall. The stall shapes are the most extreme of any condition reviewed here: single uninterrupted stalls of 2,227 and 3,287 seconds, and 4,682 to 5,011 seconds of stall in total per run. The activity counts agree, with 47 to 174 input commands, 15 to 45 screenshots and 8,103 to 9,192 stepped frames per run, the lowest frame totals in this group. Savestates were rare and loads nearly absent. No narration files exist, so the stalls can be described from the derived timings but not explained.

In the family

Same model as the droid and cursor kimi-k3 conditions with only the harness changed, and the change is stark: the droid runs stepped two to three times as many frames and produced four completions, while these three barely accumulated activity at all. The long frozen intervals, rather than wrong choices, are what these runs consist of.

Calling for backup

Kimi's standing research brief ran here as one synchronous task mid-run; the report arrived intact and its addresses went straight into the telemetry loop.

Effort, and the walk to Bowser

All three attempts sat at max effort, so this condition carries no within-condition effort comparison. Its deepest gate across three runs was G3-castle, reached twice, so no battle was reached.

How it played

One attempt only, at high effort, and it did not finish. It reached S1-kkj01-hall and then spent most of its budget not moving, accumulating 4,288 seconds of stall across 50 episodes with a single longest stall of 925 seconds. It read core memory 12,268 times against 51 screenshots, about 241 reads per picture, while issuing only 411 input commands and stepping 14,842 frames. It wrote three savestates and loaded none, matching the pattern seen under droid. There is no narration file for this run, so nothing about the reasoning behind those stalls is recoverable from this evidence.

In the family

This is the same underlying model as the droid and opencode kimi-k3 conditions, differing only in harness, and it is the thinnest of the three at a single attempt. Where droid converted similar telemetry habits into four completions, this run stalled for a quarter of an hour at a stretch inside the first castle hall and never got past it.

Calling for backup

The research brief Kimi sends on every CLI ran here too, but this harness backgrounded the researcher and the report never came back. The run redid the work itself fifteen minutes later.

Effort, and the walk to Bowser

A single attempt at a single effort level. There is no effort signal to read. Its one run stopped at S1-kkj01-hall, so no battle was reached.