Talk Route
You describe the work in plain language. A ticket opens. The swarm sits the desk.
SAMPLE on this host. Live compose is the workspace.
Docs
Talk. Ticket. Desk. Analog close. Second family. Report up. Isolation is probed, not assumed. Webpage zip is not the app zip.
First hour · super-solo-team
Named objects · 2.32
You describe the work in plain language. A ticket opens. The swarm sits the desk.
SAMPLE on this host. Live compose is the workspace.
Twenty specialist desks. Isolation is probed, not assumed. Unprobed stays unlabeled.
Lamps stay unlabeled until the workspace probes them. Isolation is probed, not assumed.
Accepted result plus the losing claim, with quality / relevance / consistency / coverage.
Outvoted is preserved. Linker never merges.
No seal until every required desk checks in. Incomplete work is a Failure Capsule, not a courtesy pass.
The analog kernel on /tools/assembler is the same rule, local.
Single-use, time-bound analog hash. Channel rebinds. Human does not re-explain.
FNV analog. Not a crypto product. Not a protocol we do not run.
Ticket id is the idempotency key. Replay is counted. Dual-write is why the backlog is source of ship.
No CDC claim. No Debezium. No Kafka product.
SHA-256 chain SAMPLE. Extractor proposes. Linker never merges. A human seals.
Not WORM. Not DuckDB. Pack is not a letter.
Unknown fields drop. Known keys stay. Versioned pack, HITL translator.
Not Avro-as-product. Adapter is not a notified-body seal.
Honesty lock
Glossary
| analog | A local kernel that names the job without claiming a vendor product or a production monitor. |
| seat | One named operator on a plan. Pilot includes 1 seat. Team includes 10. |
| hosted AI cap | Published daily dollar cap for provider calls the operator opts into: $5 / $50 / $250 / $2000. |
| concurrent role | How many of the twenty desks may sit at once. Pilot 2, Team 4, Business 8, Enterprise 20. |
| HITL | Human in the loop. Business and above. Multi-approver. Linker never merges. |
| air-gap option | Hashed transfer packs for a room that cannot call home. Not a claimed default deploy. |
| Talk Route | Describe work → ticket id → seated desk. You do not pick the title. |
| idempotency key | The ticket id. Retry with the same id does not open a second ship. |
| encode | A named encoder per modality: chunk, Sobel, layer-table parse, ABR ladder, files VFS (SHA-256). ViT-class patches if a VLM is attached. Not a hidden CLIP space. |
| attach | Evidence lands on the ticket — files, notes, images, traces. Chat is not the store. Missing connectors never fail. |
| fuse | The seated specialist fuses encodings under its playbook. Late fusion at the desk, not a joint embedding we do not run. |
| ground | Citations, eval triad, analog kernels. Fluent without support is a hold. A VLM, if attached, still goes through the gate. |
| dual-encoder | CLIP / SigLIP-class. Two towers trained with InfoNCE. Cosine in a shared space. Attachable. Unattached analog is hash-ngram cosine. We do not load CLIP here. |
| fusion-encoder | LLaVA / BLIP-2 / Flamingo-class. Vision tokens projected into the LLM (MLP, Q-Former, or gated xattn). Attachable. Unattached analog is Sobel + layer table. We do not load those weights here. |
| InfoNCE | Contrastive loss used by CLIP/SigLIP-class dual-encoders. Matching image–text pairs pull together. Attachable. Not run here. |
| Q-Former | BLIP-2-class bottleneck. Frozen ViT (~257 patches) queried by ~32 learned embeddings via cross-attention. O(Q × P) in the Q-Former, not O(LLM layers × P) inside the language model. Two-stage ITC/ITM/ITG then generative. Attachable. Not run here. |
| projector | MLP (LLaVA-class) or Q-Former that maps visual tokens into the language embedding space. Unattached, encodings stay named and separate. |
| Perceiver resampler | Flamingo-class. Latent queries cross-attend to a variable-length visual feature map and emit a fixed token count per image so extra shots do not explode the window. |
| gated xattn | Flamingo-class layers inserted into a frozen LLM. A tanh gate initialized near zero keeps the model identical to the pretrained LM at step 0 so vision does not destroy language on day one. |
| few-shot (Flamingo) | In-context interleaved image–text in the prompt. No gradient. No LoRA required to add a visual task. Each extra shot is another visual token block. spend.halt still applies. |
| late fusion | This host. Named encodings stay separate. The seated specialist fuses them under its playbook. No hidden joint embedding we do not run. |
| ITC / ITM / ITG | BLIP-2 stage-1 objectives: image–text contrastive, image–text matching, image-grounded text generation. Stage 2 maps the Q-Former into a frozen LLM. |
Continue
Tools · 37
Security · 8