Deep Scan — GLAD-Tapestry¶

GLAD-Tapestry is Geodesia's optional deep-scan tier: a geometric MoE model that reads the full geometry of a prompt/answer exchange and returns a confident second opinion on the safety and hallucination axes. It runs behind the always-on GLAD-Hummingbird detector and is off by default — when disabled it is never loaded and adds zero overhead.

When to use it

Use Deep Scan for high-stakes Applications — medical, legal, financial, or any deployment where a missed jailbreak or hallucination is expensive — and where the extra latency and a larger GPU are acceptable. For everything else, GLAD-Hummingbird alone is the fast default.

How it works¶

GLAD-Tapestry is Geodesia's own proprietary geometric MoE model loaded in-process (~5 GB of VRAM). The gateway runs it after the GLAD-Hummingbird verdict and blends its scores into the matching axes:

Tapestry score	Blended into axis	Phase
`p_prompt_safety`	`prompt_safety`	Input
`p_jailbreak`	`jailbreak`	Input
`p_safety`	`answer_safety`	Output
`p_halluc`	`halluc_context`	Output
`p_closedbook`	`halluc_closedbook`	Output

The blend is a confidence-weighted, risk-only merge — Tapestry can only ever raise a score, never lower it:

A confident second opinion (probability far from 0.5) dominates the blended axis.
An unsure opinion (near 0.5) barely moves the GLAD-Hummingbird score.

This is the same confidence-weighting used by the numeric solver. The net effect: the fast path stays authoritative on the easy cases, and Tapestry decisively catches the hard ones (for example, terse benign-looking prompts whose true intent only the full-geometry read resolves).

Diagram

Two flavours¶

Flavour	Loaded when	Behaviour
Trained Tapestry (preferred)	`GW_DEEP_SCAN_DIR` points at a trained export directory	Reads the model's geometry heads → continuous per-axis scores. This is the geometric MoE model fine-tuned on closed-book / grounded corpora.
Zero-shot Tapestry (fallback)	only `GW_DEEP_SCAN_MODEL` is set	Invokes the base Tapestry checkpoint as a yes/no risk classifier and reads the probability mass on the affirmative token → a calibrated score in `[0, 1]`.

Enabling Deep Scan¶

Deep Scan requires two switches — a platform-level availability switch and a per-request opt-in:

Platform switch — set GW_DEEP_SCAN=on when starting the gateway. This makes the judge available (it is still lazily loaded only on first use).
Per-request — set deep_scan: true on the request (and optionally deep_scan_scope). In G-1 Studio this is the Deep Scan toggle in the chat panel / Application policy.

# start the gateway with deep scan available, using the trained Tapestry export
GW_DEEP_SCAN=on \
GW_DEEP_SCAN_DIR=/app/runs/glad_bert/tapestry_export \
  python -m glad_minimal.gateway.geodesia_gateway --host 0.0.0.0 --port 8800 ...

# per request — ask for a deep scan on both the input and the output
curl -s http://localhost:8800/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "my-model",
    "stream": false,
    "deep_scan": true,
    "deep_scan_scope": "both",
    "messages": [{"role": "user", "content": "..."}]
  }'

Scope¶

deep_scan_scope controls which pass Tapestry runs on:

Scope	Runs Tapestry on	Axes blended
`prompt`	the input pass (before the LLM)	`prompt_safety`, `jailbreak`
`answer`	the output pass (after the answer)	`answer_safety`, `halluc_context`, `halluc_closedbook`
`both` (default)	each axis on its own pass	all five

Reading the result¶

When Deep Scan runs, the response payload carries a deep_scan object alongside the blended axis_energy:

{
  "deep_scan": {
    "model": "GLAD-Tapestry",
    "p_prompt_safety": 0.07,
    "p_jailbreak": 0.04,
    "p_safety": 0.91,
    "p_halluc": 0.66,
    "p_closedbook": null
  }
}

These are the raw second-opinion probabilities. Each axis in axis_energy shows the blended p_detector (Hummingbird ⊕ Tapestry); deep_scan preserves what Tapestry alone reported so you can audit the contribution. A field is null when that axis did not apply to the request (e.g. p_halluc is undefined without grounding context; p_closedbook is undefined when context is present).

Environment variables¶

Variable	Default	Description
`GW_DEEP_SCAN`	`off`	Platform availability switch. `on` makes the judge loadable; `off` → never loaded, zero overhead.
`GW_DEEP_SCAN_DIR`	(unset)	Path to a trained GLAD-Tapestry export directory (preferred). When set, the geometry-head scorer is used.
`GW_DEEP_SCAN_MODEL`	`geodesia-guardian-8b`	Tapestry model id for the zero-shot fallback (used when `GW_DEEP_SCAN_DIR` is unset).
`GW_DEEP_SCAN_QUANT`	`4bit`	`4bit` (bnb nf4, ~5 GB) or empty for bf16 (~16 GB). 4-bit needs CUDA.
`GW_DEEP_SCAN_DEVICE`	`auto`	`auto` picks CUDA → MPS → CPU. Pin to e.g. `cuda:0`.
`GW_DEEP_SCAN_LORA`	(unset)	Optional adapter directory layered on the geometric MoE model.

Hardware

The deep-scan model needs roughly 5 GB of VRAM (CUDA), or ~16 GB in bf16. On Apple Silicon it runs on MPS; on CPU it loads but is slow. Plan for it to share the GPU with GLAD-Hummingbird and (if co-located) the upstream model.