The model is not the only black box.
blkbx is an evidence infrastructure layer for autonomous systems. Our core abstraction is the evidence receipt — but underneath every receipt is a local-to-global reconstruction engine.
A real AI incident is not caused by a single output. It emerges from a chain of local evidence fragments that must agree: the prompt, the retrieved documents, the policy constraint, the model output, the agent plan, the tool call, the human review, the system mutation, and the downstream consequence.
The hard problem is not merely recording each fragment. It is determining whether those fragments agree. When they glue into a coherent global record, the action is reconstructable. When they do not, the system surfaces an obstruction.
Evidence cover
prompt context
retrieved documents
policy references
model output
agent plan
tool call
human review
downstream action
Why sheaves?
Insurance depends on reconstructability. AI systems create distributed evidence. A sheaf is the mathematical object for exactly this situation: local data, overlap rules, gluing conditions, and obstruction when gluing fails.
This is the difference between observability and insurance evidence. Observability asks what happened? blkbx asks can what happened be reconstructed, verified, and defended?
Grounded in sheaf-theoretic BLT work — local sections, learned restriction maps, cohomological diagnostics, sheaf Laplacian spectra, stalk-level inconsistency, and obstruction classes.
Local section
Evidence fragment from one part of an AI action
Hook window
Capture point in a model, agent, workflow, or system
Restriction map
Rule describing how two fragments should agree
Coboundary
Measured disagreement between fragments
Global section
A coherent, reconstructable event story
H⁰
Space of globally consistent explanations
H¹
Irreducible obstruction in the record
Sheaf Laplacian
Consistency and robustness operator
Stalk defect
Local contradiction at a specific decision point
Obstruction energy
Fraction of the event not explained by coherent evidence
Capture
Capture the consequential action at the model, agent, API, workflow, or system boundary — claim denial, loan approval, refund, coverage recommendation, account mutation, exception approval.
Issue receipt
Create a signed evidence receipt with model, policy, input, output, tool, human-review, and action metadata.
Build evidence cover
Decompose the action into local evidence regions U₁…U₇.
Test consistency
Check whether overlapping regions agree — does the output match the policy? does the tool call match the approved action? does review cover the actual consequence?
Surface obstruction
If evidence does not glue, identify the obstruction: missing review, policy mismatch, version conflict, undocumented tool call, retrieval gap, unauthorized action.
Export insurance packet
Produce artifacts for underwriting, audit, incident response, claims review, broker submissions, carrier review, and reinsurer analysis.
Not decorative math — insurance questions.
global-section existence
Can the event be reconstructed?
H⁰ consistency dimension
How much of the record is coherent?
H¹ obstruction count
Where does the evidence disagree?
spectral gap
Is the workflow insurable?
stalk defect
Which control failed?
obstruction energy
Is the record complete enough for claims review?
From technical evidence to insurance-facing artifacts.
- Underwriting control exhibits
- Broker submission packets
- AI incident reports
- Claims reconstruction files
- Audit trails
- Regulatory review packets
- Reinsurer exposure summaries
- Renewal evidence