01provenance per edge
Every edge cites the extractor and the location. consumes-endpoint from openapi at acme/payments/openapi.yaml#L412. observed-rpc from otel span tagged http.target=POST /v2/charge. You can trace any claim back to a line.
trust · the floor everything stands on
Why an agent on this model
can be allowed to act.
An agent that is confidently wrong is worse than no agent. The reason blastmap can be the substrate for autonomous action is the discipline this page describes.
§01 · the trust contract
Six rules that make the labor safe to sell.
These six rules are non-negotiable. They are the boundary conditions that make a confident answer worth more than a plausible one.
02confidence in [0, 1]
Every edge carries one number. Merged across extractors with noisy-OR so multiple weak signals can corroborate. Distance pays a cost across the walk. The number ranks every answer the model gives.
03observed vs declared
Declared (the schema said so). Observed (runtime saw it). Inferred (the LLM extractor read the source). The model holds the three side by side and tells you when they disagree.
04unknown is never safe
The absence of an edge is not the absence of a relationship. The model renders unknown as unknown, with a reason. An agent reading the model never confuses no evidence with no risk.
05no ai decides truth
LLMs extract candidate edges and the model scores them. The model decides what is true. The discipline that picks a number is code, not a generated narrative.
06refusal is first-class
When the evidence is not there, the model and the agents on it say undetermined. Not maybe. Not probably. Refusal is recorded and audited the same way an action would be.
§02 · four states
Every edge is in one of four states.
The state determines what an agent built on the model is allowed to do with it.
verified>= 0.85
Multiple independent extractors confirm the edge, including runtime. The agent can act inside the model's policy. The human reads the receipt.
confirmed0.6 to 0.85
Declared sources agree and at least one runtime signal does too. The agent proposes, the human approves.
inferred0.4 to 0.6
An LLM extractor read the source and produced the candidate edge. Other extractors are quiet. The agent surfaces it; nothing acts on it.
unknown< 0.4 or absent
The model does not know. Rendered as unknown, not as no edge. The agent says undetermined and stops.
§03 · security floor
The promises underneath all of it.
The discipline above is for correctness. These are the promises about where the data lives and who decides what to do with it.
local-first
Run blastmap on your laptop or inside your VPC. The model never has to leave your perimeter for you to use it.
BYO-LLM
Plug in Anthropic, OpenAI, Google, or local via Ollama. Your code never reaches an LLM you did not choose.
no telemetry
We do not phone home. No analytics on what the model says about your system. Opt-in if you ever want to share anonymized shape data.
OSS-first
The core engine is open source. You can audit the extractors, the noisy-OR merge, the refusal logic, the whole substrate.
read-only by default
The model and the agent surface are read-only by default. Writing is a separate, narrower contract that you opt in to per action class.
auditable refusals
Every undetermined verdict is recorded with the evidence checked. You can always answer 'why did it not act'.
§04 · why it matters
An agent without this is a story generator.
The contrast is not theoretical. Agents in the wild today produce plausible explanations because nothing they read tells them what is verified, what is inferred, and what is unknown.
without the model
- · every edge has the same epistemic weight
- · no source on any claim
- · no notion of stale vs current
- · silence reads as no risk
- · refusal looks like failure
with the model
- · every edge ranked by confidence
- · every edge cites its extractor and source
- · every fact knows its last-verified time
- · unknown is unknown, not zero
- · refusal is a first-class output