RFD 1133 details: the three kinds of exclusion
Ten model ids were carried under one heading in RFD 1132, and the heading was blocklisted. Reading each row back against what excludes it splits the ten three ways, and only the first three rows are what the word says.
Blocklisted by the agreements
CLAUDE.md’s blocklist names all three, and BLOCKLIST.md carries the argument under each. This table restates neither, it points at them.
| model id | goal | entry | RFD |
|---|---|---|---|
| qwen_q4_k_m_image_edit | - | Qwen-Image-Edit (2509/2511) | RFD 1043 |
| p3sam_mesh_segmentation | mesh-latents | P3-SAM / Hunyuan3D-Part | RFD 1041 |
| krea2_turbo_text_to_image | - | Krea 2 / krea2-turbo | RFD 1042 |
Qwen-Image-Edit is 20.4B and runs here only quantised, and quantised it corrupts. P3-SAM carries a territory-restricted licence that excludes the EU, the UK and South Korea. Krea 2 is revenue-gated and the restriction propagates.
The identifier qwen_q4_k_m_image_edit names the Q4_K_M build directly, so the catalog identifier and the reason for exclusion are the same fact. That is the one row where the id alone is the evidence.
Ranking these would send somebody to work the agreements have already closed. A blocklist row reopens when the agreements change, and nothing in the conversion work can reopen it.
Abandoned with the world-building scope
These four are not blocklisted anywhere. Each has an RFD in state abandoned, and each names the same cause: RFD 1064 turned the roadmap toward character concepts.
| model id | RFD | scope it left |
|---|---|---|
| weftspun_image_to_world | RFD 1049 | explorable worlds |
| lingbot_map_environment_scan | RFD 1050 | environment scan |
| worldmirror2_reconstruct | RFD 1051 | multi-photo splat |
| triposplat_image_to_splat | RFD 1052 | single-photo splat |
RFD 1049 is the caller and RFD 1051 and RFD 1052 are two of its halves, so the four are one pivot rather than four decisions.
Calling them blocklisted was wrong in a direction that costs something. A blocklisted model stays out whatever the roadmap does. An abandoned one returns with its scope, and RFD 1049 says so in its own words: package this entry again only if the world-building scope returns. A reader who saw blocklisted would not go looking for that sentence.
None of the four is licence-dirty, and none was measured and rejected. They are out of scope, which is the cheapest exclusion to reverse.
Named in RFD 1132 and nowhere else, RETRACTED
The three are placed, and this section said they were not. It is kept rather than rewritten, because the way it was wrong is the more useful half.
| model id | path | revision |
|---|---|---|
| multimodal_semantic_ids | 3-interactor/multimodal-semantic-ids | refs/tags/mesh-latents/v0.1.0-dev.1 |
| residual_fsq_recommender | 3-interactor/residual-fsq-recommender | refs/tags/mesh-latents/v0.1.0-dev.1 |
| unified_modal_embedder | 3-interactor/unified-modal-embedder | refs/tags/mesh-latents/v0.1.0-dev.1 |
default.xml carries all three as <project> entries, pinned to the mesh-latents tag, and all three are checked out at 3-interactor. Placement is what a live goal manifest says, and it says these are placed.
What the first version searched was the RFD corpus: no directory, no README, no ALIASES.md row, no citation. That search was accurate and its bound was stated. The bound was then read as a result anyway, which is the failure. An absence measured over one tree was written up under a heading that claimed the workspace, and the file that would have settled it in one grep is the same default.xml this document already cites for the goal column of RFD 1132’s table.
So the correct finding is smaller and worse. These three are not unfound, and they are not excluded either. They are placed projects on 3-interactor that RFD 1132 listed as blocklisted, and nothing in this repository records a reason. The RFD corpus still has no entry for any of them, which is now a documentation gap rather than evidence of absence.
Two ways to close it. Read the three checkouts and record what they are, or delete the rows from RFD 1132 and say the blocklisted mark was never supported. The first is the one that leaves a reader better off.
What this leaves for RFD 1132
RFD 1132 keeps every convertible model, the measured rank, and the census that orders the rest. It cites this document instead of carrying a second table, so one ranking and one exclusion list do not have to agree by hand.