FAQ

23 short answers.

Each one is under eighty words and points at the surface that owns it. We do not invent a 10M-user claim to fill this page.

Do I need to know the twenty roles?

No. You talk. A ticket opens. The swarm sits the desk. Each title is a Role Studio workspace with missions, toolkit actions, and an RBAC pass bar.

Roles

How do the twenty desks actually work?

Each desk closes a unique pain — a paper that never left the notebook, a fluent answer with no citation, an agent that said done with no contract. Shared slogans across the roster are a defect. Workday, typical ticket, playbook, Skill.md, swarm seats. A second family grades.

Twenty unique needs

How does multimodal reasoning work?

Patch / tokenize, project, fuse, ground. Vision splits into ViT-class patches or Sobel edges here. A projector (MLP or Q-Former) maps visual tokens into language if a VLM is attached. This host fuses late at the seated desk. Citations and eval ground the answer. We do not invent a joint CLIP space.

Encode · attach · fuse · ground

How does a VLM actually work?

Dual-encoder (CLIP/SigLIP) uses InfoNCE cosine for retrieve/rank. Fusion-encoder: BLIP-2 bottlenecks through a Q-Former (~257 patches → ~32 queries, O(Q×P) not O(LLM layers × P)); LLaVA projects patches with an MLP; Flamingo uses gated cross-attention with a tanh gate that starts at zero; GPT-4V/Gemini-class is a proprietary hop. This host fuses late. Families are attachable — not weights we host.

Vision-language

What is a Q-Former bottleneck?

BLIP-2-class. A frozen ViT emits ~257 patch tokens. Thirty-two learned queries self-attend, then cross-attend into those patches. The LLM sees 32 visual prefixes, not 257 concatenations. Two-stage: ITC/ITM/ITG, then a linear map into a frozen LLM. Cheaper than Flamingo-class gated xattn. Caption fluency is still not object support. Attachable — we do not run a Q-Former here.

Q-Former

How does Flamingo few-shot work?

In-context: image, text, image, text in the prompt. No gradient, no LoRA to add a visual task. A Perceiver resampler emits a fixed token count per image. Gated xattn is inserted into a frozen LLM; the tanh gate starts near zero so at step 0 the model is exactly the pretrained language model. Every extra shot is another visual token block. spend.halt still applies. Attachable — we do not host Flamingo.

Flamingo few-shot

How is fusion implemented here?

Late fusion at the seated desk. Named encoders: Sobel, hash-ngram, chunk, topology, ABR, files VFS (SHA-256). Evidence attaches to the ticket. If a family attaches: dual-encoder cosine, Q-Former bottleneck, LLaVA MLP, or Flamingo gated xattn. Ground anyway — a caption without support is a hold. Authors never grade the VLM they attached.

Fusion

Do you run CLIP or LLaVA?

No. CLIP, SigLIP, BLIP-2, LLaVA, Flamingo, and GPT-4V-class are attachable families. This host runs Sobel, hash-ngram cosine, chunk, and topology. Missing never fails the product. A caption without goldens is a hold. Trainer / Evaluator is the second family.

Vision-language

Is PromptDiff a judge?

No. Jaccard on words + line hunks + ceil(chars/4). Local. Not tiktoken and not a vulnerability score.

PromptDiff

Is the webpage zip the app?

Buyers take the operator transfer pack for air-gap. The sales-host source zip is how we restore this website — it is not a commercial SKU.

Air-gap packs

Is Enterprise a login?

No. Enterprise talks to a human on /sales. Never /login. plan=pro is rewritten to plan=team.

Talk to a human

What does Pilot include?

Pilot is $0, 1 seat included, $5 hosted AI / day, 2 concurrent roles, Agent API, Talk Route, local tools.

Pricing

How is annual Team billed?

$66 per seat / month, billed $792 per seat / year. Save is $79×12 − $66×12 = $156, published as $158 with named rounding.

Pricing math

Is isolation ever certified by default?

No. Isolation is probed, not assumed. Unprobed stays unlabeled. We never paint it as certified.

Isolation

Is a transfer pack a legal letter?

No. Art. 50 is an operational desk. A SHA-256 digest is not a notified-body seal. Pack is not a letter.

Art. 50

Do local tools leave this tab?

No, until you open a workspace and choose a hosted call. PromptDiff, VRAM, Seek, Dissent, Assembler stay in-browser.

Tools

Do you claim SOC 2 or HIPAA?

No. No SOC badge. No HIPAA certified claim. A BAA is a contract conversation. Do not send ePHI in forms.

Trust

What does it cost to hire these roles?

2026 US mid-senior cash bands live on /roles. Loaded = 1.3× midpoint. Year-1 adds ~18% recruiting. Not EngOS payroll. ARR_CLAIMED=false.

Replacement cost

What is a Seat Token?

A single-use, time-bound analog hash so the next desk can rebind the channel. Not a crypto product.

Seat Token

MCP or Skills?

MCP is tools and systems the Agent API may call. Skills are the playbook a seated role already has. You talk in tickets, not protocols.

Playbooks

What happens to a losing claim?

It is written to the Dissent Log. Outvoted is not erased. Quality, relevance, consistency, coverage stay on the record.

Dissent Log

When does a pack seal?

After every required desk checks in. Fan-in Assembler blocks the seal until then. No courtesy pass.

Assembler

Where do I report a security issue?

Email security@tuce.app or use /sales. There is no self-serve seal portal and no invented bounty payout.

Report

Still a human question? Talk to a human — never /login.

EngOS