5,050 Points — Automated Search Milestone
A real-time Automated Policy Search milestone with zero model calls, exact replay and pending video publication.
Read the updateExplore the laboratory
Meaningful milestones and experiment explanations, with honest provenance and publication status.
We publish significant milestones, algorithm changes, perception comparisons and useful failures. Ordinary runs remain metrics and records rather than thousands of near-identical public videos.
YouTube is preferred once an editorial candidate is approved. The site also supports approved HTML5 video. Gameplay captures can contain third-party graphics and audio; creating a recording does not automatically clear the underlying rights.
Pengo M01–M03:21,210 /31,040 /100,400 native verified trajectory points. M03 review v2.1 is the current editorial version; the original is preserved. Project Pengo.
Historical200k video: M03B / 216,480 verified, 36 rounds completed, ACT 5 of the third cycle unfinished. A complete new native reset replay.
Historical 300k video: M04 /301,030 verified,52 rounds completed, ACT5 of cycle4 unfinished.0 MODEL CALLS. Complete new native reset replay.
Latest Pengo video: M05 / 500,350 verified. 500,350 alive / 500,350 at tape end; 4 lives remain, zero detected deaths. 86 chronological rounds completed; ACT 7 of cycle 6 is unfinished with 2 objectives. 0 MODEL CALLS. Complete new native reset replay.
14 of 36 pieces, 0 of 9 password letters. The game is still incomplete.
E135 independently verifies a new fragment from Middle112/146 in SECURITY TERMINAL22. All thirteen earlier native shapes and RGB colors are preserved. Ordinary accepted DISABLE spends SNO1 to0. Final independent inventory: SNO0 / LIFT6, empty PSW, HUD1:20:00. No new ordinary death; three earlier deaths remain. There is no final screen or complete normal-boot proof.
YouTube | E135 report and evidence
The 71.68-second silent replay labels nine retained initial laboratory episodes, zero intermediate restores within each. Valid supported or completed-search prefixes end before excluded later inputs. Twenty-three other whole calibration episodes and all excluded suffixes are omitted but fully charged. No failed or partial-Searching snapshot is resumed and no retained death is rewound. This is development evidence, not an uninterrupted game from normal boot.
Three bounded Searching pauses precede a native closed window, an empty target and the same supported actor. The later physical Left50 exit and memory scan establish the gift: one new fragment, with all thirteen originals exact. An earlier generic-panel false positive was quarantined and never restored. Late freezing expires during return; rendered contact guards remain active.
Canonical continuation uses the valid completed Search at supported Middle146, independently validated by its downstream exit and inventory. That validation contains no additional search, reward or death. Three pods remain new: Left24/146, Left24/194 and Middle112/194. Password assembly and reproducible completion proof are still outstanding.
13 of 36 pieces, 0 of 9 password letters. The game is still incomplete.
E054 completes the Middle112/98 furniture search in SECURITY TERMINAL22. An ordinary accepted DISABLE spends SNO1 to0. The completed search earns a SNOOZE code, and physical exit through originalLeft50 plus independent inventory verifies SNO1 / LIFT6, all thirteen original shapes and native RGB colors, empty PSW and HUD1:19:21. No new fragment or death; three earlier ordinary deaths remain.
YouTube | E054 report and evidence
The 59.56-second silent replay labels five retained initial laboratory episodes, with zero intermediate restores inside each. Three retained prefixes stop before later detector failures; their excluded actions and nine whole calibration episodes remain charged to the project. An extra four-second still explicitly shows a separate independent debit-validation episode, SNO0/LIFT6; it is outside the five retained source episodes. The search completes before withdrawal. Two half-second Searching pauses, the native ZZZ reward, actual platform support and physical return are observed.
This is development evidence, not an uninterrupted normal-boot game. Four other Left/Middle pods remain in this room. More fragments, password assembly and reproducible final-screen completion proof remain outstanding.
13 of 36 pieces, 0 of 9 password letters. The game is still incomplete. E045 solves the second musical room’s six-note challenge. SNOOZES increase from 0 to 1; LIFT INITS remain 6. All thirteen earlier inventory masks are preserved. No new fragment or death; three earlier ordinary deaths remain. Independent inventory HUD 1:17:03.
YouTube | E045 report and evidence
The 69.20-second silent replay labels three initial laboratory episodes, with zero intermediate restores inside each. This exact challenge took 204 distinct order attempts: 203 incorrect orders are omitted. Winning rank 143 is (1,0), (0,0), (3,1), (2,4), (2,2), (2,1), using zero-based board coordinates. The checkerboard clears, the actor physically leaves through the original Left194 doorway, and independent inventory scanning verifies the ordinary code award. No purple-bar quit is used.
This is development evidence. Further furniture searches, password assembly and uninterrupted completion proof from normal boot remain outstanding.
13 of 36 collected fragments, 0 of 9 password letters. The game is incomplete.
E037 verifies the Pocket Computer controls needed for puzzle assembly: an ordinary workspace drop, color change, accepted pair merge and Undo. Two observed fragments form an exact 276-pixel displayed union. Undo restores the selected original, and an independent scan matches all thirteen original shapes and native colors. SNOOZES0, LIFT INITS6; no new fragment, code, letter or death. Physical Pocket OFF returns to the corridor.
YouTube | E037 report and evidence
The 69.92-second silent replay uses one initial laboratory snapshot, with zero intermediate restores. Fourteen other whole UI calibration episodes are excluded. Its final HUD is 1:14:34. The canonical E036 checkpoint stays unchanged at 1:13:30, retaining thirteen fragments and three earlier ordinary deaths. This is development evidence, not a completed normal-boot game.
Acceptance of this pair establishes the UI behavior; whether these two fragments belong to the same final four-piece puzzle is unknown. Stored inventory during the merge was not scanned. Exact original shapes and colors are checked after Undo. No complete puzzle or password letter is inferred.
13 of 36 pieces, 0 of 9 password letters. The game is still incomplete. E036 finds a new fragment in the small Left24/98 object in Room22. Independent inventory scanning retains all twelve earlier masks and verifies the thirteenth. SNOOZES decrease from 1 to 0; LIFT INITS remain 6. No new death; three earlier ordinary deaths remain. Inventory HUD 1:13:30.
YouTube | E036 report and evidence
The 45.56-second silent replay labels three initial laboratory episodes, with zero intermediate restores inside each. Ordinary DISABLE receives PASSWORD ACCEPTED and ten unchanged rendered robot samples verify freezing. The agent descends the left lift from 50 to 98, searches in two bounded Searching pauses, returns physically from 98 to 50, and leaves through the original Left50 doorway. A separate inventory scan verifies thirteen exact 32x16 masks and the ordinary SNOOZE debit.
The search episode retains its valid prefix through row131, ending at a supported Top50 checkpoint saved before a later clipped-door detector stop. That later suffix and seven other complete calibration episodes are excluded. No failed save-state is used as the successful source. Search completes before retreat; persistence during long retreats is not established.
The new thirteen-mask offline analysis finds zero non-overlapping quartets; no password letter is inferred. Further furniture searches, password assembly and uninterrupted completion proof from normal boot remain outstanding.
12 of 36 pieces, 0 of 9 password letters. The game is still incomplete. E035 solves the first musical room’s six-note challenge. SNOOZES increase from 0 to 1; LIFT INITS remain 6. All twelve earlier inventory masks are preserved. No new fragment or death; three earlier ordinary deaths remain. Independent inventory HUD 1:12:27.
YouTube | E035 report and evidence
The 72.40-second silent replay labels three initial laboratory episodes, with zero intermediate restores inside each. This exact challenge took 412 distinct order attempts: 411 incorrect orders are omitted. Winning rank 435 is (2,4), (3,5), (0,7), (2,0), (3,7), (1,7), using zero-based board coordinates. The checkerboard clears, the actor physically leaves through the original Right194 doorway, and independent inventory scanning verifies the ordinary code award. No purple-bar quit is used.
This is development evidence. Further furniture searches, password assembly and uninterrupted completion proof from normal boot remain outstanding.
12 of 36 puzzle pieces; game incomplete. E028 verifies a new fragment, all eleven earlier masks retained, SNO0/LIFT6, zero letters and no new death.
YouTube | E028 evidence | Project. 45.12s, three labelled retained laboratory episodes, zero intermediate restores per episode; three unpromoted Room22 tests omitted. Physical lift return, original Left50 exit and independent inventory verified. Full normal-boot completion proof remains outstanding.
11 of 36 puzzle pieces; game incomplete. E027 verifies LIFT5 to6 and SNO0, with all eleven exact masks preserved. No new fragment or death; zero letters.
YouTube | E027 evidence | Project. 66.08s, three labelled laboratory episodes, zero intermediate restores per episode; 74 incorrect orders omitted. Original Left194 exit and inventory verified. Full normal-boot completion proof remains outstanding.
11 of 36 puzzle pieces; game incomplete. All ten prior masks are retained. The branch includes one new ordinary electrocution and respawn: three deaths total, SNO0/LIFT4, zero letters, HUD12:58:32.
YouTube | E025 evidence | Project. 102.00s,13 labelled laboratory sources, zero intermediate restores per episode. The native reward still has an explicit4s hold. Full normal-boot proof remains outstanding.
10 of 36 puzzle pieces; game incomplete. An ordinary SNOOZE and a direct left-lift route allow the agent to search the stereo. All nine previous masks are preserved, with SNO0/LIFT4 and zero password letters.
YouTube | E023 evidence | Project. 39.96s, silent sampled RGB at1x measured emulated time. Two labelled initial laboratory snapshots, zero intermediate restores per run. Whole-game proof from normal boot remains outstanding.
9 of 36 puzzle pieces; game incomplete. Room06’s lamp yields the ninth fragment. The controller returns after SNOOZE expires by waiting for a moving robot to clear the route. All eight earlier masks remain, with SNO0/LIFT4. Three labelled initial laboratory snapshots; zero intermediate restores per run.
YouTube | E021 evidence | Project. 57.68s, silent sampled RGB at1x measured emulated time, with labelled source holds and an inventory still. Normal-boot completion remains outstanding.
8 of 36 puzzle pieces; game incomplete. Room26 yields an eighth fragment and SNOOZE resupply. Final inventory preserves all seven earlier masks, SNO1/LIFT4. Four labelled initial laboratory snapshots; zero intermediate restores per run.
YouTube | E020 evidence | Project. 50.36s, silent sampled RGB at1x measured emulated time, with labelled source holds and final inventory still. Normal-boot completion remains outstanding.
7 of 36 puzzle pieces; game incomplete. E018 validates room24 collection and return, preserving all six earlier piece masks. SNO0/LIFT3, no new death. One initial laboratory snapshot, zero intermediate restores in this run.
YouTube · E018 evidence · Project Impossible Mission. Silent sampled replay at 1x measured emulated time; controller pauses omitted. Opening and ending holds are labelled. Normal-boot completion remains outstanding.
5/5 short starting-corridor episodes. Paused decisions, pixel-only reactive rules, no training. This first C64 skill does not establish whole-game or real-time play.
A 31.08-second silent sampled replay at 0.25x measured emulated speed, with added opening/ending holds. The entire native field is shown; controller pauses are omitted. Project Impossible Mission · E001 evidence and limitations.
500,350 alive / 500,350 at tape end; 4 lives remain, zero detected deaths. 86 chronological rounds completed; ACT 7 of cycle 6 is unfinished with 2 objectives. Validated restore search; verification 0 LOAD / 0 RAM writes. 0 MODEL CALLS.
Read the milestone report and verification evidence
301,030 alive / 301,030 at tape end; four lives remain, zero detected deaths. 52 chronological rounds completed; ACT 5 of cycle 4 remains unfinished with four objectives. Validated restore search; final verification0 LOAD /0 RAM writes.0 MODEL CALLS.
Read the milestone report and verification evidence
216,480 alive / 216,480 at tape end; four lives remain, zero detected deaths. 36 chronological rounds completed; ACT 5 of the third cycle remains unfinished with one objective. Validated restore search; final native verification 0 LOAD / 0 RAM writes. A08 target300k remains NUL.
Read the milestone report and verification evidence
Complete native reset replay,zeroLOAD/RAMwrites. Validated restore search is separate from final verification. No autonomous-agent,full-game or world-record claim.
Read the milestone report and verification evidence
Complete native reset replay,zeroLOAD/RAMwrites. Validated restore search is separate from final verification. No autonomous-agent,full-game or world-record claim.
Read the milestone report and verification evidence
Complete native reset replay,zeroLOAD/RAMwrites. Validated restore search is separate from final verification. No autonomous-agent,full-game or world-record claim.
Read the milestone report and verification evidence
SEARCH — 0 MODEL CALLS
Complete native-speed 5,750-point E006 Search run, pleiads-20261005-131958-e6ceb4; zero model calls. Exact applied-input Lua replay verified. Sampled at 10 fps; no intentional cuts or speed-up; no audio. Pleiads (Tehkan, 1981). Published after human editorial and rights review.
SEARCH — 0 MODEL CALLS
Complete native-speed 5,600-point E006 Search run, pleiads-20261005-124119-f2b26d; zero model calls. Exact applied-input Lua replay verified. Sampled at 10 fps; no intentional cuts or speed-up; no audio. Pleiads (Tehkan, 1981). Published after human editorial and rights review.
SEARCH — 0 MODEL CALLS
Complete native-speed 5,540-point E006 Search run, pleiads-20261005-120227-03ddb1; zero model calls. Exact applied-input Lua replay verified. Sampled at 10 fps; no intentional cuts or speed-up; no audio. Pleiads (Tehkan, 1981). Published after human editorial and rights review.
SEARCH — 0 MODEL CALLS
Research footage awaits editorial and rights review.
A real-time Search milestone with deterministic replay verification. The video awaits editorial and rights review; the underlying game graphics and audio are not assumed to be copyright-free.
A real-time Automated Policy Search milestone with zero model calls, exact replay and pending video publication.
Read the update