Provenance
Where did this come from? Source count, age, chain depth, and source health decide whether the answer has a floor.
working lab artifact
trust architecture
OpenAI logprobs
batch-invariant proof engine
Trust architecture, reduced to a working instrument.
Most AI demos show the answer. This one shows the release decision around the answer: provenance, confidence, consistency, attribution, and the controls that decide whether a human sees it first.
It is not a project list. It is the kind of artifact I build when a leadership room asks whether their model output is safe enough to ship.
GRAIN
GRIP
KNOT
VEIN
Receipt
Confidence is the weak layer. The answer may be right, but the release gate holds it until a human checks the uncertain claim.
/LAB
Live instrument · provider-selectable
Inspect one answer before it gets trusted.
Choose a prompt, pick OpenAI or HuggingFace, adjust the gate, and run the inspection. The answer comes from a real LLM call through the Raising Agents serverless endpoint. OpenAI is the instant logprob path; HuggingFace is the full-access trace path that can expose internal signals when the RunPod worker is configured.
Decision
Review required
0.71
weighted trust score
Answer under inspection
Signals surfaced
Inspection trace
Expert Agent review
Architecture review pending
Run an inspection to ask the LLM-based expert to explain the result and the Trust Architecture behind it.
GRAIN checks provenance. GRIP checks confidence. KNOT checks consistency. VEIN checks attribution. The gate decides release, review, or block.
Mechanistic trace
/R
Repeatability audit · live capped run
One answer is not a reliability signal.
The original demo had a full deterministic AI surface. This page turns that into a live capped run: same prompt, several calls, hashes grouped by output. The point is operational, not theatrical.
Run audit
Repeated live calls
preview mode
click Run live for real calls
/REF
Architecture reference
Four questions before an answer becomes an action.
Trust Architecture is a release system for model output. It does not claim the model is truthful. It asks whether enough evidence exists to let the answer pass without intervention.
Input
Prompt
Evidence
Policy
→
Model
LLM call
Logprobs
Trace
→
Signals
21 readings
6 families
→
Layers
GRAIN
GRIP
KNOT
VEIN
→
Decision
Release
Review
Block
Confidence
How certain should we be? Entropy, logprob, calibration, hidden-state drift, and late-layer override catch confident wrongness.
Consistency
Does it stay stable? Repeated calls and branch agreement reveal whether the answer survives rerun pressure.
Attribution
Is each claim traceable? Claim support, coverage, citation density, and unsupported claims turn prose into an audit object.
Provider capability matters.
Some providers expose logprobs or top-k data. Some do not. Local Hugging Face backends expose deeper traces. A serious trust surface says which signals are actually measured.
| Provider | logprobs | hidden | attention | repeatable |
|---|---|---|---|---|
| openai | limited | no | no | weak |
| anthropic | no | no | no | weak |
| huggingface | full | yes | yes | strong |
Information theory
Entropy, NLL, token margin, confidence calibration, and variance read the model's probability surface when the provider exposes it.
Behavioral checks
Self-consistency and semantic equivalence ask whether repeated generations preserve the same operational answer.
Mechanistic trace
HuggingFace full access adds attention entropy, representational drift, cross-layer consistency, and logit-lens convergence.
Geometry
Reasoning smoothness and curvature anomalies look for jumps in the answer's semantic trajectory.
Ensemble gate
The final decision combines available signals, layer weights, and release thresholds into release, review, or block.
Course connection
This is the course thesis in artifact form.
AI Transformation Leads do not need more positioning language. They need instruments: inventory scanners, trust gates, repeatability harnesses, and the judgment to explain the result to leadership.