Relevant can look like correct.
92% similarDifferent customer
89% similarRight transaction
The generator receives both and must guess which relationship matters.
The control layer between retrieval and response
Control the evidence. Then let AI speak.
PAIL turns possible evidence into one machine-readable decision: release a verified packet, expose a conflict, or refuse when proof is missing.
Works afterSQLVector searchFilesLogsGraph
Database rowsRetrieved chunksEvents and logs
PAIL AUTHORITY BOUNDARYEvidence
runtime
runtime
- Identity
- Relation
- Time
- State
- Source
- Conflict
VERIFIEDRelease packetCONFLICTStop generationMISSINGRefuse honestly
Retrieval finds possibilities. PAIL sets the authority.
Keep the search stack you already use. PAIL evaluates its candidate set before any model, workflow, or operator treats those candidates as fact.
PAIL
Every released fact has a boundary.
DROPWrong customer
KEEPExact transaction
The application receives a stable outcome and a source-linked packet.
Capability families
One runtime. Four ways to put evidence under control.
Start with one failure mode. Add only the capability your workflow needs.
01 · RAG EVIDENCE GUARD
Stop the wrong record before it becomes a confident answer.
Verify exact anchors, permitted joins, source boundaries, time, and contradictions after retrieval.
- Bring your own retriever
- Fail-closed outcomes
- Evidence IDs preserved
02 · DECISION RUNTIME
Return a bounded decision when an LLM is unnecessary.
Use explicit facts and constraints to produce JSON or a deterministic response.
- Model optional
- Same input, same decision
- Unknowns stay visible
03 · RCA EVIDENCE VIEW
Build a traceable incident view without claiming causal proof.
Isolate related events, preserve sequence, and surface disagreement for human investigation.
- Trace and event isolation
- Point-in-time scope
- Correlation labeled honestly
04 · LOCAL RUNTIME
Check bounded evidence where cloud generation is unavailable.
Run deterministic evidence decisions locally and choose JSON, templates, or an optional local model.
- Local-first path
- Generation not required
- Private runtime stays private
The operating path
From candidate data to an answer your application can trust.
Each stage has one responsibility. Retrieval may rank; only the evidence runtime may authorize.
Retrieve candidates
Use exact search, SQL, vectors, graph, metadata, or an existing RAG framework.
Preserve structure
Keep identifiers, source, time, state, and record boundaries available for checks.
Apply authority rules
Evaluate identity, relation, time, source, contradiction, and required evidence.
Issue the contract
Release verified evidence, show a conflict, or return no verified evidence.
Choose the narrator
Return JSON, use a template, or let an optional model explain only the packet.
THE OUTPUT IS NOT PROSE
A decision contract your software can enforce.
status: "CONFLICT_WITHHELD"
evidence_ids: ["F12", "F19"]
generation_allowed: false
reason: "customer_id disagrees"Choose the evaluation boundary
Explore publicly. Validate privately. Operate deliberately.
No public demo is presented as a customer production deployment.
Public sample lab
Use labeled synthetic scenarios to see verified, conflict, and missing-evidence outcomes. No private files or private core.
Open without sign-in →Signed file trial
When email, workspace storage, and the protected runtime are connected, a tenant-bound trial supports sanitized files and ten guarded queries.
Check trial availability →Private runtime
Place the runtime behind your application boundary and connect your own retriever, identity, storage, model, and monitoring.
Read the public contract →Disconnected workflow
Use deterministic packets locally without a hosted model. Hardware, recovery, and workload limits must be tested in the target environment.
Read the boundaries →Measured on a labeled fixture
A result you can inspect, not a universal promise.
The published evidence-control campaign used 4,980 generated messy records across 3,000 cases. It measured the release decision itself, without an external LLM.
Inspect methods and limitsThe honest fit
Strong where evidence has structure. Explicit where it does not.
PAIL protects an evidence boundary. It is not a replacement for every search, reasoning, or operations system.
Records, events, policies, and bounded decisions.
- Explicit IDs, entities, versions, or timestamps
- Facts distributed across known sources
- Workflows where refusal is safer than guessing
- Applications that need provenance and machine outcomes
Universal understanding or production proof.
- Not a replacement for broad semantic retrieval
- Does not prove correlation is causation
- Does not resolve source truth without policy or review
- No production SLA or independent certification yet
Bring one difficult workflow
Start with the answer your current RAG should refuse.
Use a small sanitized dataset with correct evidence, one contradiction, and one question the sources cannot answer.