94% “payment failed after gateway timeout”
92% “TXN-9912 failed after database lock”
89% “TXN-0025 completed locally”
Evidence control for RAG and AI systems
Make every AI answer earn its evidence.
PAIL sits between retrieval and generation. It checks which records belong together, exposes contradictions, and releases a smaller traceable packet only when the evidence holds.
Works with
FilesSQLVector searchLogsAny LLM
CRMTXN-0025
customer CUST-125PAYMENTSTXN-0025
customer CUST-999EVENT LOGTRACE-9025
gateway timeout
PAIL
Evidence
runtime identity · relation · time
runtime identity · relation · time
VERIFIEDRelease packet
CONFLICTWithhold answer
MISSINGRefuse honestly
Candidate evidenceAuthority boundaryMachine-readable outcome
Scroll into the evidence system
RETRIEVALFinds possible evidence
→
PAILDecides what may support the answer
→
LLM · OPTIONALExplains verified facts
The missing layer
Search can find a similar record. It cannot prove it is the right record.
A vector database is excellent at broad semantic recall. But a customer ID, transaction ID, policy version, timestamp, or contradictory source may matter more than similarity. PAIL gives those facts authority before generation begins.
VS
KEEP TXN-0025 · gateway timeout
DROP TXN-9912 · wrong transaction
STOP customer IDs disagree
One runtime, two operating modes
Wrap your current RAG, or make a bounded decision without an LLM.
Ingest
Files, records, logs, database rows, APIs, or retrieved chunks.
Locate
Exact lookup, SQL, metadata, keyword, graph, or vector retrieval.
Verify
Bind entity, role, time, state, source, and allowed relations.
Decide
Release, show conflict, ask for clarification, or withhold.
Deliver
Return JSON, a deterministic answer, or an LLM-ready packet.
Audit
Preserve the evidence IDs, rules fired, unknowns, and token count.
Keep your vector database and model.
PAIL checks the candidate set after retrieval and gates what reaches generation.
Explore middleware mode →Let the model speak, not decide.
For bounded policies and record workflows, rules create the decision object. Templates or an optional model explain it.
Explore deterministic mode →A stable decision contract
Three honest outcomes. No hidden guess.
Your application does not need to parse confident prose to understand what happened. Every query returns an explicit machine-readable state.
Evidence agrees.
Release a bounded packet with provenance and optional natural-language explanation.
generation_allowed: trueSources disagree.
Expose the conflicting fields and stop before a model silently chooses a side.
generation_allowed: falseProof is missing.
Return what is unknown and refuse to answer from model memory.
generation_allowed: false
What the prototype actually does
An evidence investigator, not another chatbot.
PAIL separates retrieval, evidence authority, deterministic inference, and language generation so each part can be tested independently.
Prioritize transaction, order, customer, trace, case, and device anchors over vague similarity.
Check that records can legally and structurally combine before forming a claim.
Order events, respect point-in-time cutoffs, and distinguish provisional from final states.
Keep unresolved conflicts visible instead of averaging them into a confident answer.
Send only the selected facts, provenance, unknowns, and answer contract downstream.
Use deterministic templates, a local model, or a hosted LLM without giving it evidence authority.
Isolate one trace and construct a bounded timeline without claiming correlation proves causality.
Run a single local node with its own store, audit trail, and signed bounded exchange artifacts.
Measured, with the boundary attached
Prototype evidence—not borrowed credibility.
These results come from one generated, deliberately messy record-based campaign. They demonstrate deterministic behavior on that fixture, not universal accuracy or a customer production deployment.
Read methods and limitationsWhere it fits first
Record-heavy workflows where a wrong join matters.
PAIL is strongest when the answer depends on identifiable entities, states, versions, relationships, or event sequences—not broad literary interpretation.
Payments
Transaction evidence
Bind orders, payments, customers, retries, reversals, and settlement events.
Explore → OperationsIncident timelines
Separate one trace from surrounding noise and expose missing causal proof.
Explore → PolicyEligibility decisions
Combine current policy, user state, quantities, alternatives, and approval requirements.
Explore → OfflineLocal evidence nodes
Continue bounded verification where cloud connectivity is weak or prohibited.
Explore →
Plain answers
What is PAIL doing that normal RAG does not?
This is the explanation to send to a nontechnical colleague before asking them to try the product.
Read every question →Is PAIL only filtering records?+
No. Filtering removes obvious noise. PAIL also binds entities, validates relations, orders events, checks state and source, exposes contradictions, and records why evidence was released or withheld.
Does PAIL replace a vector database?+
Usually no. Vector search finds semantically similar candidates. PAIL decides which candidates may support a conclusion. In exact record workflows, PAIL can also operate without vector search.
Is an LLM compulsory?+
No. PAIL can return JSON or a deterministic template answer. An LLM is useful for understanding flexible language and producing friendly prose, but it should not override the verified decision object.
Do teams manually create chunks every time?+
No. They define policies once: sections for PDFs, functions for code, transactions for payment logs, incidents for traces, and entities for database rows. PAIL adds entity-aware boundaries where fixed token windows would mix unrelated records.
Can one pipeline work for every sector?+
The evidence core can be shared, but adapters and policies must vary. Banking events, HR policy, stories, sensor streams, and SQL rows should not be chunked or interpreted in the same way.
What role does the Panini-inspired grammar play?+
It is a deterministic rule architecture: normalize different surface forms into actors, actions, objects, quantities, roles, states, and constraints; then apply explicit transformations and conflict rules. It is not a claim that Sanskrit grammar alone understands arbitrary modern language.
Bring one difficult workflow
We will start by proving what the system should refuse.
Use synthetic or sanitized records. Plant correct answers, missing answers, and contradictions. Then compare ordinary retrieval with the guarded result.