Task contract. Keep the official CLRS initial array, graph, or text/pattern state unchanged in Input. Target begins with the same clean second, then visualizes every official time-indexed hint through the record's true lengths and ends on the official output.
30published algorithms
1,000official train trajectories per algorithm
5explicit Gate B renderer schemas
8Gate B samples
Current lifecycle status. Ready PR #507 is aligned to current dev at exact head c747f703. Owner PASS is officially recorded for all eight samples and 24 checks; MJXWANG is requested for independent Human Evaluator review. Binary-search samples explicitly state the lower-bound problem and label the official result as LOWER-BOUND INDEX. Open the synchronized canonical Gate B review. Full generation and Final writes have not started.
Input → target preview
Representative canonical Gate B evidence. These published exact-head videos demonstrate array-pointer decoding, weighted shortest-path state, and the two KMP phases. The complete eight-sample packet is available on the canonical review page above.
1. Insertion sort — linked predecessor order
train/insertion_sort/0 · 16 official hint states · 96 frames · 12 FPS
INPUT: unchanged initial array
TARGET: official i, j, and decoded pred_h order
2. Dijkstra — queue, settled set, distance, and parent
train/dijkstra/0 · 17 official hint states · 96 frames · 12 FPS
INPUT: unchanged weighted graph and source
TARGET: official u, in_queue, mark, d, and pi_h
3. KMP — prefix table then text matching
train/kmp_matcher/0 · 13 official hint states · 96 frames · 12 FPS
INPUT: unchanged text and pattern symbols
TARGET: official prefix, reset, alignment, and match probes
Identity and provenance
| Dataset | DeepMind CLRS Algorithmic Reasoning Benchmark / CLRS-30 |
|---|
| Official source | google-deepmind/clrs and CLRS30_v1.0.0 publisher archive |
|---|
| Paper | The CLRS Algorithmic Reasoning Benchmark, ICML 2022 |
|---|
| Published scale | 30 algorithms; the official train split exposes 1,000 trajectories per algorithm at problem size 16, approximately 30,000 train records |
|---|
| Verified local source | Complete publisher archive, 85,628,509 bytes |
|---|
| Archive SHA-256 | 7743b2f14459c34eb25f1eb6a70047924634bfd00f2ddf5872456b75f73c93b3 |
|---|
| Repository license | Apache License 2.0; publisher citation guidance is retained |
|---|
Native formats and annotation semantics
| Native item | Meaning | Task 241 use |
|---|
input_* | Initial arrays, targets, graphs, edge weights, sources, text, and patterns | Complete clean Input state |
hint_* | Padded time-indexed algorithm probes | Chronological Target animation, truncated to true lengths |
output_* | Official algorithm result such as permutation predecessor, return index, graph parent, or match | Final Target state |
lengths | True trajectory length before padding | Exact visible hint-state count |
CLRS30_v1.0.0 / <algorithm>_<split>
├── input_* → initial state
├── hint_* → T × probes
├── lengths → true T
└── output_* → official result
canonical identity = revision / split / algorithm / record_index
V2V conversion
| Output | Construction |
|---|
input_video.mp4 | Eight-second hold of the complete official initial state. |
target_video.mp4 | Frames 0–11 preserve the initial state; frames 12–95 map every official hint state chronologically and finish on the official output. |
prompt.txt | Algorithm-specific wording explicitly names visible probe types, colors, and presentation order. |
metadata.json | Metadata 1.2 binding revision, split, algorithm, record index, true length, per-tensor SHA-256 values, media facts, fingerprint, and exact producing commit. |
Gate B selection
| Algorithm | Records | Visible official probes |
|---|
| Insertion sort | 0, 7 | i, j, pred_h, final pred |
| Binary search | 0, 11 | low, mid, high, final return |
| BFS | 0, 9 | reach_h, pi_h, final pi |
| Dijkstra | 0 | u, in_queue, mark, d, pi_h, final pi |
| KMP matcher | 0 | phase, i, q, k, s, reset masks, pi, final match |
Quality gates and boundaries
- All eight samples pass the four-file contract, Metadata 1.2, unique source identity, exact archive identity, and complete media decode.
- All 16 videos are 960×540 H.264/yuv420p, 96 frames at 12 FPS; Input and Target match materially for the first second and visibly differ afterward.
- Formal resume returns
written=0, skipped=8; an independent clean generation reproduces all 32 four-file objects byte-for-byte. - Insertion-sort predecessor probes are decoded as the official linked order rather than guessed as ordinary permutation indices.
- Dense Dijkstra parent state is preserved while only current changes receive strong emphasis.
- The current implementation supports five explicit algorithm schemas. The remaining 25 CLRS-30 algorithms fail closed until probe-specific renderers and tests exist; approximately 30,000 train records is a planned upper bound, not completed Full data.
- Canonical Validation S3, Gate B evidence, and Owner PASS are current. No Final, Gate C, promotion, or Human Evaluator decision is claimed.
Pipeline → final four-file contract
official CLRS archive → verify frozen archive SHA-256
→ load exact split / algorithm / record_index
→ truncate every hint tensor to official lengths
→ schema-specific deterministic array / graph / string renderer
→ clean initial Input + chronological hint-driven Target
→ H.264/yuv420p pair + exact prompt.txt + Metadata 1.2 metadata.json
→ automated media/source/resume checks → Gate B human review
→ only after dual Gate B PASS and merge: separately authorize Full and Gate C
Completeness and source verification
| Check | Result |
|---|
| Official source and paper | Verified and recorded |
| Archive identity | 85,628,509 bytes; SHA-256 7743b2f14459…93b3 |
| Canonical Validation | 8 samples / 32 payload objects / 979,020 bytes; run 8eb4287b…0ebdc |
| Task tests | 7/7 PASS |
| Workflow control-plane | PASS on Ready PR #507 exact head c747f703 |
| Canonical review evidence | Evidence eeca3409…34b85; synchronized playback and 17/17 hosted resources verified |
| Human review | Owner PASS; MJXWANG requested; Evaluator decision pending |
| Production Final | 0 samples; Full generation is not authorized |