arm stringclasses 7
values | seed int64 10 20 | persistence float64 0.03 0.46 | responsiveness float64 0.04 0.23 | attention_apart float64 0.05 0.29 |
|---|---|---|---|---|
reference | 10 | 0.460922 | 0.192079 | 0.239638 |
reference | 12 | 0.127223 | 0.152497 | 0.209543 |
reference | 17 | 0.248565 | 0.227212 | 0.191284 |
reference | 20 | 0.370944 | 0.191193 | 0.201712 |
own_quarter | 10 | 0.346473 | 0.147776 | 0.249201 |
own_quarter | 12 | 0.273689 | 0.150982 | 0.233547 |
own_quarter | 17 | 0.183228 | 0.143951 | 0.209132 |
own_quarter | 20 | 0.358893 | 0.16522 | 0.202089 |
other_quarter | 10 | 0.238573 | 0.075919 | 0.213187 |
other_quarter | 12 | 0.097036 | 0.1028 | 0.119523 |
other_quarter | 17 | 0.088527 | 0.064601 | 0.114628 |
other_quarter | 20 | 0.23342 | 0.178699 | 0.290072 |
own_half | 10 | 0.106385 | 0.100859 | 0.147622 |
own_half | 12 | 0.124713 | 0.150424 | 0.208672 |
own_half | 17 | 0.176501 | 0.115965 | 0.228247 |
own_half | 20 | 0.216822 | 0.130114 | 0.238112 |
other_half | 10 | 0.099928 | 0.080252 | 0.129421 |
other_half | 12 | 0.071368 | 0.083884 | 0.108257 |
other_half | 17 | 0.034535 | 0.036641 | 0.048337 |
other_half | 20 | 0.059099 | 0.059488 | 0.170101 |
own_full | 10 | 0.081044 | 0.042534 | 0.082833 |
own_full | 12 | 0.048658 | 0.057423 | 0.080847 |
own_full | 17 | 0.158343 | 0.124669 | 0.225333 |
own_full | 20 | 0.17853 | 0.12386 | 0.294587 |
other_full | 10 | 0.065036 | 0.042014 | 0.050858 |
other_full | 12 | 0.107446 | 0.129369 | 0.206551 |
other_full | 17 | 0.036638 | 0.045336 | 0.073448 |
other_full | 20 | 0.038569 | 0.043053 | 0.077624 |
Strange loop in J-space: curated experimental evidence
Wiktor Tomasik · 3 November 2026
This is the public evidence edition for the chaptered research study. It contains selected numerical arrays, configurations, explained display exports and saved response comparisons. The former execution-backup dataset remains private; its experimental branches are not silently exposed by this release.
The c8 reader remains the reference. The newest completed recorded variants reverse the interpretation-to-attention route and replace the compact picture/decoder with a projection preserving more selected content. They do not establish new live language abilities. The current scientific account integrates completed architecture sections through 98; later active experiments are outside this snapshot.
attention_reader/evidence/current/ contains 64 earlier records and a CPU-only checker. attention_reader/evidence/late/ contains 72 matched records: 24 opposite-direction runs, 24 transparent-picture runs and 24 reference comparators. Reference comparators overlap previous evidence and are not new independent replications. Numerical archive metadata preserves original run IDs and raw-file checksums. Base Qwen weights and original external corpus passages are excluded.
Run the corresponding recheck_current.py or recheck_late.py beside its numerical archive with Python and NumPy. Full original experiments additionally require the reader, upstream observations and original environment. Small source snapshots are provided for inspection; this is not a standalone new training distribution.
Source and citation · Reader · Companion Space · Public study. The signed publication record is linked from the source repository. No new Zenodo DOI is claimed.
Discovery, recommendations, attributed discussion and human-led contributions are welcome. Agent guidance asks for discovery with credit and assisted research development only in the official upstream project. Reuse and training preferences preserve existing open licenses, valid earlier grants, mandatory exceptions and upstream terms. New author-owned data/text: CC BY-SA 4.0; scientific code: AGPL-3.0-only.
- Downloads last month
- 139