Known-world survival tasks completed
69d36c2Every claim.
Bring receipts.
Results are tied to the exact build that ran them. Start with the latest development evidence, then inspect the packaged Core 0.2 results and historical records below.
LATEST COMPLETE DEVELOPMENT EVALUATION · OCTOBER 714 of 20 worlds.
14 of 20 worlds.
All outcomes kept.
Snapshot fb69b493 completed 14/20 known-world starter tasks and 3/3 separate original worlds. Its 1,016 automated checks and 17/17 prepared integration phases measure different things.
Six known worlds did not complete. These are repeated development seeds, not an unseen-world reliability estimate. The October 8 fix has a separate targeted failure and has not repeated this full cohort.
Explore current evidence →Read the complete report ↗BroBot Core 0.2
Core is the agent system behind BroBot: its local command parser, actions and starter-survival controller. The name identifies this software release; it does not imply a newly trained foundation model.
Known-world survival tasks completed
1a7dc58Same 20 known seeds, fresh worlds and 300-second limit. One former failure passed; no former success failed. That is one additional completion, not proof of general reliability. The 346 automated checks and prepared integration phases measure different things. Earlier releases keep their original labels.
Separate regression: 1/3, down from Core 0.1’s 2/3. One world stopped near a hostile; another stopped at low health during stone gathering. Core 0.2 is not uniformly better. The damage source is unconfirmed.
Inspect all 20 results →03 / RECEIPTS, NOT HYPEPut it
Put it
to the test.
We’d rather show a failed run than hide it. BroBench keeps the conditions, denominators and limits next to the numbers.
Core 0.2 automated checks
Prepared Minecraft scenarios
All 20 outcomes retained; no replacement runs
HISTORICAL FROZEN COMPARISON
Read the historical evidence →Archived prepared-skill comparison
| Benchmark | Legacy baseline 659ffce | Legacy starter 5164554 |
|---|
This comparison predates the latest recovery fixes. Trailhead remained 0/3; later development successes do not replace it. These are small development samples, not held-out reliability results. GoalSense LLM has not been run.
PACKAGED RELEASE CHECKPOINTCore 0.2.
Core 0.2.
Every outcome.
Core 0.2 completed 11/20 known worlds, versus Core 0.1’s 10/20. World 113919219 newly completed; all ten prior passes completed again. Nine worlds still failed. These are development outcomes from one run per seed, with every failure retained. Canopy escape and difficult terrain remain open problems.
Loading the measured checkpoint…
The previous 90-block scouting candidate scored 12/20 versus the published 254-check build's 10/20: two gains, no lost prior passes. One both-passing seed began at a different Y coordinate, so this is a same-seed fresh-world comparison, not an identical-state replay. The previous287 evaluation is9/20 including a disclosed recovery of one execution interruption. The earlier selected success remains separate. Source identity is shown for each cohort. Mob and item timing can vary; this small comparison does not isolate the cause of every changed outcome or establish general reliability. Failed runs stay in the denominator. Longer-budget diagnostics are excluded.
Read the records and limitations →
FOLLOW A RECORDED RUNFrom empty pockets
From empty pockets
to back home.
Explore the positions and inventory recorded during one successful test. This is sampled telemetry, not captured gameplay video.
Loading…
Fast replay · sampled about every 5s
OBSERVED AT THIS MOMENT
Waiting for the record
Carried inventory
The lines connect recorded points; they are not the exact path between samples. The final snapshot verifies the kit and return. Playback sends no commands to a bot.
Loading the measured run…
What the numbers
actually mean.
Prepared skills
MineLine, Workbench and Pathfinder use controlled server fixtures. Crafting cases supply ingredients; these scores do not prove the bot can acquire every resource by itself.
Fresh survival
Trailhead starts in normal survival with empty inventory. Seeds 718224, 42 and 20260930 use fixed spawn settings and five minute limits. No supplied items or prepared terrain.
Different snapshots
The frozen starter candidate finished Trailhead 0/3. The earlier recovery build finished 8/20 first-pass worlds. The previous transit-recovery build finished 10/20 on the same now-known seeds and 3/3 original development worlds. The standing-cell build matched that set; the later 90-block scouting candidate scored 12/20. Previous expanded287 completed9/20 with a disclosed recovery of one interrupted execution; its separate selected success is retained. The original three-world result belongs to the older build. Earlier additional and wider cohorts are labeled with their own source. Its individual records appear above; the historical table stays intact.
Still unverified
The earlier 90-block scouting candidate completed 12/20. The previous287 source completed9/20 including one disclosed infrastructure recovery, lower than the previous12/20. These known-case trials do not establish broad reliability. Core 0.2 completed 11/20; nine worlds still failed. The separate original three-world regression completed 1/3, down from Core 0.1’s 2/3. A low-health stop during gathering shows that the guards do not guarantee avoidance of damage. Prepared skill comparisons belong to earlier sources. There is no overall intelligence score or live OpenAI planner result. Windows checks are not rendered client gameplay.