Appearance
Blackridge Trust Architecture
We Don't Ask You To Trust Our Conclusions.
We ask you to inspect our evidence.
Every claim Blackridge produces can be traced back to immutable records, versioned algorithms, and explicit evidence.
If we cannot prove a claim, we say so.
If we infer a claim, we label it.
If we simulate a scenario, we distinguish it from observed reality.
Trust is not a feature.
It is an architectural property.
What Blackridge Knows
Blackridge records observable facts.
Examples include:
- Model invocations
- Tool calls
- Costs
- Policy decisions
- Budget and routing verdicts
- Cache outcomes
- Routing choices
- Timing
- Retries
- Imported human participation events, when supplied
- Workflow lineage
These become immutable canonical records.
They are never rewritten.
What Blackridge Does Not Know
Blackridge does not pretend to know information it cannot observe.
Unless a framework explicitly provides it, we do not claim to know:
- Internal reasoning
- Hidden chain-of-thought
- Planner deliberation
- Memory scoring
- Retrieval ranking rationale
- Private model state
Unknown remains unknown.
Unknown is never converted into certainty.
Every Claim Has Evidence
Every forensic finding carries:
- Evidence grade
- Confidence
- Supporting records
- Algorithm version
- Reproducibility hash
Pricing versions, policy versions, and custody hashes are recorded when they are evidence for that finding or artifact. Unknown remains explicit.
Nothing appears without provenance.
Five Evidence Grades
Observed
Captured directly by Blackridge.
Imported
Provided by trusted external systems.
Derived
Deterministically computed from canonical records.
Inferred
Estimated from incomplete evidence.
Always labeled.
Simulated
Counterfactual.
Represents what could have happened, not what did happen.
Unknown Is Not Zero
Missing information is never silently converted into a default value.
Missing pricing does not become zero cost.
Missing attribution does not become an assumed owner.
Missing usage does not become zero tokens.
Unknown remains unknown.
Immutable Evidence
Canonical records are append-only.
Corrections create new records.
History is never rewritten.
This allows every investigation to be reconstructed from the original evidence.
Replay Is Reconstruction
Blackridge does not replay from logs.
It reconstructs execution from immutable evidence.
Replay quality depends on evidence quality.
Every replay includes an integrity score describing how complete the reconstruction is.
Chain of Custody
Every generated artifact records:
- Input evidence
- Generator version
- Algorithm version
- Evidence hash
- Output hash
Policy and pricing versions are recorded when they were in scope for the derivation.
Every report is traceable.
Framework Neutral
Blackridge does not require a specific agent framework.
Whether you use:
- LangGraph
- CrewAI
- OpenAI Agents
- Claude Code
- Cursor
- Custom runtimes
Blackridge observes the execution boundary.
It explains economic and runtime evidence. Broad governance and action authorization belong outside the current Blackridge release.
We Prefer Honest Uncertainty Over False Precision
Many systems answer every question.
Blackridge answers only the questions it can support with evidence.
That is slower.
It is also more trustworthy.