ARC-AGI-3 — Единая карта проекта

Completed, evidence-qualified LS20 human Part 1 + Part 2 reconstruction and agent-implications analysis, delivered to D 2026-08-24, later standardized (2026-08-25) and primary-source-repaired (2026-08-25).

Источник: arc-prize-2026-ls20-human-combined-analysis.md

Ключевые факты

• Levels verified: 1-5 (Part 1), 6-7 (Part 2); victory timestamp 22:58.607.
• Part 1: 146 regular markers/audit rows, 15 action contact sheets, 5 dense sheets.
• Part 2: 1,430 frames (exact-PTS analysis).
• Standardized package: 16 continuous events, zero gaps/overlaps, 7 evidence paths.
• Repair coverage: 2,830 ASR segments, 6,710 words; substantive/context speech 2,335 covered / 246 uncertain / 0 missing; all speech 2,444 covered / 275 uncertain / 0 missing.
• Discrepancy resolved: 9 byte-identical pairs vs 18 rows with identical=True — pixel-delta threshold 18, not row duplication.

Скиллы — заметки

• Part 1 verifies: Levels 1-5, HELP, destructive RESET and replay, color/shape/rotation discovery, spring transport, resource planning, and phase tracking of a moving plus.
• Part 2 (exact-PTS analysis) verifies Levels 6-7 and victory at 22:58.607.
• Across both parts: scoped hypotheses, minimal discriminating tests, local correction, semantic goal tuples, hierarchical subgoals, phase-aware planning, persistent map memory under fog.
• Agent implications drawn: separate world/HUD state; semantic transition records; hard no-world-delta/repeated-effect stop guard; evidence-linked scoped hypothesis ledger; destructive UI safety policy; bounded transformation cycles; deterministic route/spring executor; moving-object phase; resource projection; persistent fog map; heavy-model calls only on novelty/violation/subgoal transitions.

Факты среды

• Mechanics referenced: color/shape/rotation discovery, spring transport, a moving plus object with trackable phase, destructive RESET (with replay), and a HELP mechanic.
• Nine Part 1 before/after PNG pairs are byte-identical — NOT treated as proof of no click (offsets come from ASR/action extraction, not exact decoded-frame PTS).

Разбор по уровням / журнал событий

• 2026-08-24: Completed and delivered to D — canonical detailed Russian report (560 lines).
• Part 1 independently re-audited: all 146 regular markers indexed; 15 action contact sheets and 5 dense sheets reviewed.
• 2026-08-25 ("Standardized package"): normalized to the FT09 event model. Validation passed: 16 continuous events, zero gaps/overlaps.
• 2026-08-25 ("Primary-source fidelity repair"): triggered because D was concerned the 9-pairs/18-rows discrepancy meant tests had failed. Repair completed against both original videos; aggregate validation passed 3/3.

Открытые вопросы

• 275 "uncertain" speech segments retained without invented wording — automatic transcription is complete, but absolute phonetic verbatimness requires manual listening and must not be claimed yet.

Durable record of D's second ARC-AGI-3 human trace (FT09): selection rationale, recording, completed evidence-qualified analysis deliverables, the cross-game standardized packaging effort, an operational correction about false completion claims, and a pointer to the third trace (VC33).

Источник: arc-prize-2026-ft09-human-trace.md

Ключевые факты

• Game slug ft09; live-verified ID 2026-08-24: ft09-0d8bbf25.
• Archetype: non-agentic visual logic — pattern matching/composition, segmentation/occlusion, relational visual state, click grounding, hypothesis discrimination; no avatar/route/HUD.
• Video: 1555s (25:55), 3024x1696, ~481 MB.
• Chronology: 41 continuous intervals (0.000–1554.970s), zero gaps/overlaps; 29/29 evidence refs valid; 3,110 dense frames; 104 overview frames; 249 delta events; ASR span 2.700–1554.010s.

Скиллы — заметки

• FT09 documents the trace-recording/analysis process rather than gameplay skill sub-rules. Meta-methodology used for extraction: "observation → hypothesis/alternatives → predicted discriminating action → observed delta → support/refute/update", observations kept separate from interpretation.

Факты среды

• FT09 is characterized as officially non-agentic visual logic involving pattern matching/composition, without an avatar, route, or HUD.
• Adds segmentation/occlusion, relational visual state, click grounding, and hypothesis discrimination; chosen partly to test whether a DSL/program-synthesis branch is useful.

Разбор по уровням / журнал событий

• Selected as second human trace after LS20 (LS20 covered navigation/affordances/resources/dynamics/fog/map-memory/phase-planning; FT09 tests non-agentic visual-logic skills).
• 2026-08-24: D submitted one continuous FT09 screen+voice video, 1555s.
• 2026-08-25: completion task finished and factually validated. Levels 1-6 and VICTORY!!! visually confirmed.
• 2026-08-25: normalization task created FT09-event-model packages for LS20 and VC33 plus a common three-game manifest. Aggregate verification passed 3/3.
• Operational correction (2026-08-25): a prior LS20 report was mistakenly presented as FT09 completion — that claim must never be reused.
• Subsequently D recorded and completed the third trace, VC33.

Открытые вопросы

• ASR (speech transcript) was not manually proofread.
• 2 fps frame sampling cannot prove sub-frame ordering.
• Level 6 vertical coupling supported by repeated observations but not proven universal.
• D's required sequence: complete individual traces first, then joint LS20+FT09+VC33 architecture analysis — still pending as of this file.

VC33 third human trace: two-part 2026-08-24 recording (Part labels reversed vs. content) analyzed into a 429-line evidence-qualified report + CSV timeline, then standardized 2026-08-25, with a decisive gate click at PTS 48.017 driving progress to Victory at PTS 794.213.

Источник: arc-prize-2026-third-human-trace.md

Ключевые факты

• Part 1 = 873s/14:33 (~317MB); Part 2 = 1000s/16:40 (~390MB); source duration 31:13 total.
• GAME OVER at PTS 800.013; decisive gate click at PTS 48.017 (lift +50.000, yellow object transfer +52.000); second confirming gate ~PTS 74.000; Victory first frame at PTS 794.213.
• Standardized package validated: 18 continuous events, 0 gaps/overlaps, 14 evidence paths, 8 decoded JPGs, 38 timeline rows.
• DOCTRINE CHECK: this file's evidence CONTRADICTS a "probably random clicks, not understood mechanics" characterization — a specific decisive causal click on an identified gate mechanic produced measured, reproducible effects, independently reconfirmed by a second gate.

Скиллы — заметки

• Transferable agent lessons: trigger a novelty interrupt and affordance sweep for every new visual class; choose tests by budget-aware information gain; separate state predicates from transition operators; represent regions, transfers and gates explicitly as a graph.
• D became stuck at one point, causing the attempt to run long across two videos — useful reasoning evidence for reconstructing what invalidated/weakened the active hypothesis.
• The orange conditional gate was noticed but initially misclassified as an indicator; repeated alignment attempts then exhausted the budget, leading to GAME OVER in submitted Part 2 at PTS 800.013.

Факты среды

• An "orange conditional gate" mechanic: clicking it is decisive — at PTS 48.017 the lift rises by 50.000 and a yellow object transfers by 52.000.
• A second gate around PTS 74.000 confirms the rule is reproducible, not one-off.
• The gate was initially misclassified as a mere "indicator" rather than an actionable/conditional trigger, causing wasted alignment attempts and budget exhaustion.

Разбор по уровням / журнал событий

• 2026-08-24: D submitted two sequential videos; Part labels semantically reversed — submitted "Part 2" contains Levels 1-4 plus GAME OVER, submitted "Part 1" continues Level 4 through to Victory.
• GAME OVER at PTS 800.013 (submitted Part 2, after repeated failed alignment attempts).
• Decisive gate click at PTS 48.017 (submitted Part 1).
• Second confirming gate ~PTS 74.000 (submitted Part 1).
• First visible Victory frame: submitted Part 1, PTS 794.213.
• 2026-08-24: 429-line evidence-qualified report and CSV timeline delivered.
• 2026-08-25: normalized to the FT09 event model; validation passed.
• Aggregate LS20+FT09+VC33 validation passed 3/3.

Открытые вопросы

— нет данных —

Completed TR87 fourth human ARC-AGI-3 trace (evidence-qualified chronology, reasoning trace, strategy summary, verification), joining the 4-package manifest with LS20/FT09/VC33.

Источник: arc-prize-2026-tr87-human-trace.md

Ключевые факты

• Video: 1968.679s (32:49), 3024x1674, ~98.1 MiB; single continuous recording to VICTORY!!! (no Part1/Part2 split).
• 53 reasoning-trace events (TR87-E001–E053), 122/122 evidence references resolved.
• Per-level attributed spans: L1 478.55s, L2 340.68s, L3 233.40s, L4 350.86s, L5 109.12s, L6 464.70s (longest level).
• Status counts: 25 supported / 4 refuted / 7 updated / 11 not_tested / 6 not_applicable.
• Pixel-delta index: 5663 frames; VICTORY spike at pts=1965.075s, mean_diff=16.64 (order of magnitude above any other value — unambiguous).

Скиллы — заметки

• Board-reading: fixed legend of paired glyphs plus one or two rows of editable glyphs with a white-bracket cursor; confirmed via on-screen HELP text.
• Navigation/routing: arrows move the cursor along the row and/or cycle the selected glyph's value — consistent across all 6 levels.
• Goal-setting/prediction: Level 4 required discovering a two-hop indirect "bridge" match (figure matches through an inverted intermediary), discovered only after two full failed scroll-cycles.
• Prediction/inference: Level 6 (final, ~460s) required recognizing a "1 blue -> 2 yellow via 2 pink intermediaries" structure plus a palindromic 1-2-3-3-2-1 positional symmetry; player revised (1,2,3) to (1,3,2) based on absence of duplicate values.
• Perception caution: level count HUD briefly misread as "/8" from a small thumbnail, corrected by a tight crop — always re-verify HUD text at full resolution.
• Perception caution: Level 5's circular grey mask on green was initially misidentified as a distinct "terrain" mechanic, then corrected to the same core game under a different skin.

Факты среды

• Core mechanic: fixed legend of paired glyphs; below it, editable glyph row(s) with a white-bracket cursor; goal is to match the legend's pairing rule.
• Levels differ in legend size (6 pairs L1/L2, 8 pairs L3/L4/L6), number of selector rows (1 vs 2), and visual skin (L5 circular grey-on-green vs rectangular grey-on-teal elsewhere — same underlying game).
• Level 4 introduces a two-hop indirect "bridge" match structure.
• Level 6 introduces a "1 blue -> 2 yellow via 2 pink intermediaries" structure with palindromic 1-2-3-3-2-1 positional symmetry.

Разбор по уровням / журнал событий

• Selected as fourth human trace (after LS20, FT09, VC33), per the 2026-08-25 week plan, as the "program/grammar/sequence" archetype-fill trace.
• Slug tr87 confirmed by browser URL and in-game HUD badge.
• D submitted one continuous video: 1968.679s (32:49), full playthrough to VICTORY!!! (no split).
• Full-video pixel-delta index built (5663 frames) to pin exact level-transition PTS and the terminal VICTORY spike.
• Local faster-whisper ASR run to completion; a separate cloud ASR pass supplied narrative content where needed, spot-checked (~15 key phrases).
• 6 independent parallel agents (one per level) each read actual frame images and grepped events.csv, producing 53 schema-compliant events merged into one trace.
• Canonical package written matching FT09/LS20/VC33 standard; manifest updated to 4/4 packages verified.
• Outcome: full playthrough won (VICTORY!!!) across all 6 levels.

Открытые вопросы

• Interior event boundaries within levels 3-6 mix exact ASR timecodes with interpolated estimates.
• The pixel-delta index locates transitions/action clusters but cannot identify which arrow key was pressed.
• The two-hop bridge rule (L4) and intermediary/palindrome structure (L6) are the player's own successful inferences, not verified against game source code.
• No separate adversarial "refute this claim" verification round was run for this trace.

TU93 human trace (5th completed trace): maze navigation + snake-predator interaction mechanics, WON 9/9 levels, ~23m23s session, confirmed genuine full clear with a stated falsifiable reason at every level transition.

Источник: arc-prize-2026-tu93-human-trace.md

Ключевые факты

• Session duration: 00:23:23.52 (1403.52s), H.264 3024x1592 @ ~6fps.
• Archetype: primarily Navigation per pre-play prediction, confirmed but with richer-than-expected mechanic depth (predator vision/facing, chase state, order-dependent elimination, goal-safety rule).
• 5th completed human trace, after LS20, FT09, VC33, TR87.
• No doctrine conflicts found; doubles down on doctrine #1 (meaningful pressing) and doctrine #4 (novelty logging).

Скиллы — заметки

• Local one-step safety fallback under high enemy count (L5, L9): when threat count exceeds what can be mentally tracked, enumerate this turn's legal moves and discard only the ones fatal NEXT turn — cheaper than full simulation.
• "New indicator that doesn't kill you must do something else" heuristic (L7): when a novel visual change coincides with NOT dying despite apparent danger, promote "enables a different mechanic" over "cosmetic" as default hypothesis, test within 1-2 moves.
• Novelty-probing method (L4): see new figure → probing moves → infer movement pattern → predict future position → path around it.
• Order-dependent elimination (L3): three red enemies eaten in specific order (back-to-front, never approaching an un-eaten enemy's eye side).
• Kiting / indirect obstacle removal (L8): chase-state persists across turns and can be exploited — D lures the chasing enemy in a loop so it commits, opening the original path behind it.
• Direction/threat encoding via inset sub-pixel (L2): enemy's "eye" side is a directional threat zone active at range 2; approaching from the non-eye side is safe. Learned by trial-death.

Факты среды

• Grid maze, LEVEL k/9 counter, fixed at 9 levels for this game.
• Player = blue square ("snake"). Goal = green square. Enemies = colored squares each carrying one smaller inset pixel encoding facing direction = threatened cell(s), not decoration.
• Magenta/pink horizontal bar fills along the bottom across the run — never narrated by D; likely a per-level/per-run move/time budget indicator, function unconfirmed.
• Goal-tile safety condition can change between levels of the same game (L5→L6): D stepped onto the goal while a threat was still adjacent and instantly LOST — landing on the goal is only a win if no enemy can eat the player on that same tile/frame, not just present.
• Orange enemies can move through/overlap red enemies (L6) — red does not block orange.
• Bordeaux (dark-red) enemy (L7): stationary until player passes near, inset pixel then flips to yellow ("фонарик"), starts chasing every subsequent turn; trigger radius vs facing-cone ambiguous.

Разбор по уровням / журнал событий

• L1 (0-150s): pure maze traversal, no threats.
• L2 (~150-300s): first enemy (red) with purple inset eye — died approaching from eye side, restarted, approached from non-eye side, safe.
• L3 (~300-480s): three red enemies eaten in specific order; first orange enemy appears at level end.
• L4 (~480-600s): orange enemy movement pattern learned via 2-3 probing moves — straight line, reverses direction.
• L5 (~600-720s): four orange enemies simultaneously; local one-step safety fallback used.
• L5→L6 boundary (~660-700s): landed on goal while threat still adjacent, LOST — new "safe on arrival" rule discovered via death.
• L6 (~720-870s): orange+red together; discovered orange passes through red (one wrong move, then corrected).
• L7 (~870-1080s): first bordeaux enemy appears, stationary until approached, then chases; inferred via elimination reasoning.
• L8 (~1080-1200s): same chase-enemy blocks only path to goal; lured via a loop.
• L9/final (~1200-1400s): combines orange patroller and bordeaux chaser in tighter space; per-turn safety fallback used. Session ends on VICTORY!!!, LEVEL 9/9.

Открытые вопросы

• Bottom magenta progress bar: never mentioned by D across 23 minutes — move counter, time limit, per-level budget, or cosmetic?
• L5→L6 goal-tile death: was the earlier level safe-on-arrival by coincidence, or is the rule genuinely different per level?
• L3 order-of-eating 3 reds: exact decision rule beyond "avoid the one that would eat me" not narrated — trial-and-error or a rule computed in advance?
• L6 orange passes through red: was there any visual cue before the mistake that could have predicted pass-through behavior, or genuinely unknowable until tested?
• L7 first bordeaux sighting: exact cell offset at which the pixel flips to yellow — fixed radius-2 trigger or facing-cone check? Ambiguous.

WA30 human trace (2026-08-27): Sokoban-style box-placement game with autonomous helper (orange) and rival (purple) NPCs, WON 9/9 levels, 6th completed human trace, analyzed via the fixed core-skill grid per D's 2026-08-27 doctrine-architecture correction (no new named micro-skills).

Источник: arc-prize-2026-wa30-human-trace.md

Ключевые факты

• Duration: 1251.33s (~20:51), H.264 3024x1592 @ 4.07fps.
• HP/move-budget bar levels observed: full at L1/L6, ~90% L7, ~70% L8, ~55% L9.
• Level-transition timestamps (t≈s): 5, 200, 355(L3), 480(L4), 600(L5), 715(L6), 840(L7), 955(L8), 1080 & 1200(L9), victory 1248.
• 6th completed human trace, after LS20, FT09, VC33, TR87, TU93.

Скиллы — заметки

• Perception/board-reading: L1 — proximity to a box changes its render (blur/"spotlight"), the affordance signal for "attachable". SPACEBAR read as context-sensitive via action-probing (not assumption).
• L3-L5 — reading each helper zone's box-count requirement is a live perception task; D revised a miscount live (open question re: exact process).
• Movement: L1 — discovers SPACEBAR attach/detach by trial.
• Prediction of others' moves: L2 — first orange NPC; wrong initial hypothesis (intercept/steal) flipped to "helper, goal-directed" after observing autonomous placement. L6 — first purple NPC; accidentally spacebar-neutralizes it, discovering the "eat a rival" mechanic empirically. L9 — purple stops rather than reroutes when blocked by player's body, blurring the A/B/C NPC classification.
• Navigation/routing: L3-L5 — player's job becomes pure logistics once walls separate player and helper zones; navigation efficiency named as the only residual risk once mechanic is understood. L8 — new topology confirms L1-L7 rules generalize.
• Goal-setting/prioritization: L7 — with two purple rivals, D sequences: neutralize both first (pure negative-value blockers), trust orange helpers to self-resolve rest. L9 — delivers to low-risk zone first, then intercepts purple; tracks HP bar to self-estimate "~10 moves left" and reorders actions.

Факты среды

• HP/move-budget bar — universal across all ARC-AGI-3 games/levels; here visibly depletes level-to-level, confirming it's consumed by moves taken, not just time or deaths (no death occurred this run, bar still drained).
• Player = green/white square; white side = facing = last movement direction (same "carried directional marker" pattern as TU93, here on the player).
• Movable pieces = blue-on-black-tile "boxes". Target = blue-gray zone sized to exactly fit the number of boxes required — implicit "how many needed" readout from L1.
• SPACEBAR is context-sensitive: near a box, attach/detach (carry); near a purple NPC approached from the right side, neutralizes ("eats") it — same input, different effect by adjacency/context.
• Orange NPC (from L2) = autonomous cooperative ally, independently places boxes.
• Purple NPC (from L6) = autonomous adversarial rival, carries placed boxes back OUT of the target zone; neutralized by approaching from the correct side + spacebar.
• Both orange and purple are "type B" figures (goal-directed) but opposite alignment — distinguishing required one observed interaction each, no visual cue (both plain colored squares).
• L8: the player cannot pass through orange (blocks movement, not decorative).
• L9: purple respects the player's body as an obstacle — stops advancing rather than rerouting when its corridor path is physically blocked.

Разбор по уровням / журнал событий

• L1 (0-150s): forms hypothesis "arrange boxes into target zone" without reading Help; regrets not checking Help sooner. Discovers SPACEBAR attach/detach.
• L2 (~240-360s): first orange NPC; wrong hypothesis corrected after observing autonomous cooperative placement.
• L3-L5 (~360-720s): multiple orange helpers in walled sub-zones; player becomes pure logistics; D revises a box-count miscount live.
• L6 (~715-840s): first purple NPC; it removes a placed box; D accidentally discovers the neutralize/"eat" mechanic while trying to reclaim it.
• L7 (~840-960s): two purple rivals; D neutralizes both first, then lets orange helpers self-resolve rest.
• L8 (~960-1080s): new map topology, same ruleset confirmed to generalize; new fact logged (can't pass through orange).
• L9/final (~1080-1248s): two zones, one orange-handled (low priority), one purple-guarded via maze corridor (high priority); delivers to low-risk zone first, then blocks/intercepts purple; tracks HP bar, self-estimates ~10 moves left.
• Victory screen confirmed on-screen at ~1248s. Result: WON, 9/9 levels, self-described "прошёл идеально".

Открытые вопросы

• ~780-820s (L6, first purple neutralize): what exactly was the trigger — facing direction, relative position, simple adjacency + spacebar regardless of side?
• L6-L9, purple's harm potential: was "can purple kill me" ever actually tested/risked, or consistently avoided out of caution?
• L6-L8, purple persistence: if never eaten, does purple keep undoing placements indefinitely, or stop after some condition?
• ~480-600s (L3-L4), multi-helper box counts: what's the actual process for reading how many boxes a given orange helper's zone needs, and why did the first read come out wrong?
• ~1080-1200s (L9), purple stopping when blocked: any visual cue hinting purple might respect an obstacle before testing, or purely trial-and-error? Also: how is the HP/move-budget bar visually tracked to estimate "moves remaining"?