LoomSignal
Docs← Back to site
Architecture

Trustworthy by construction

LoomSignal is a single service that turns raw tool output into governed context and serves it over MCP. Its whole job is to make sure an AI never acts on a half-truth — so every design choice is about where trust is earned and where it is enforced.

The path of every fact
WRITE GATE
The single point where anything is written. New facts are appended, never overwritten — a conflicting update supersedes the old version instead of erasing it.
appends to a tamper-evident chain
GOVERNED STORE + PROVENANCE
Versioned context with a hash-linked provenance chain. Every fact is traceable to its source and moment; the chain can be independently verified.
read through the gate
READ-GATE
The only way context leaves the system. It serves canonical, cited answers — and withholds or clearly flags anything contested. It never blends conflicting claims into one confident-sounding answer.

The medallion: three tiers of trust

Context is refined in stages, and each stage has a different level of trust. Nothing is deleted along the way — promotion is additive.

BRONZERaw observations, exactly as your tools reported them. Append-only and immutable — the audit trail.
SILVERThe many names for one real thing (a service, a ticket, a component) reconciled into a single stable entity.
GOLDThe governed answer an agent is allowed to read — promoted only when it clears the gate.

One gate, verifiable provenance

Every write goes through a single gate. It never silently overwrites a prior fact: a conflicting update supersedes the old version and both remain in the record. Each step is appended to a hash-linked provenance chain, so any fact can be traced back to the exact source and moment it came from — and the chain can be verified independently.

When two sources disagree, LoomSignal does not pick a winner and hide the loser. The conflict is kept as an explicit, first-class signal — and surfaced, not smoothed over.

The read-gate: nothing contested slips through

Context only ever leaves through the read-gate. It serves canonical, cited answers and withholds or clearly flagsanything contested or stale. It is designed to fail closed — if a non-canonical fact can’t be flagged, it is not served at all, rather than shipped as if it were settled.

The MCP surface

Agents don’t call a proprietary API — they speak MCP, the open protocol. The surface is small and stable:

query_context
Ask for the governed context around a topic or entity, with citations.
search
Find relevant documents across the substrate.
get_document
Fetch a specific document — full, structured, or summarized.
push_observationwrite
Write a new raw signal into bronze (the only writable tier).
list_contradictions
See where sources currently disagree.
get_provenance
Walk the provenance chain behind any fact.

On your infrastructure, offline-tolerant

The whole substrate runs inside your environment — sources, credentials, and knowledge never leave. Context is stored locally and also projected to a portable, human-readable mirror that survives independently of the database. Search keeps working even with no model and no network: full-text is always the floor, and semantic search layers on top when available.

Next: Installation