01 / REPLAY
Replay the state.
Run the same question against what the database knew at a specific moment — even after later corrections, backfills, or policy changes.
Explore bitemporal queries →
The decision database
Reconstruct any decision exactly as it happened.
InvariantDB preserves what was true, what was known, and the evidence behind every decision — so months or years later, you can explain it without guessing.
Bitemporal by design
Cypher-native
Tamper-evident history
Underwriting twin · loan-1842
verified
What the twin knew
March 11, 2026
✓ chain intact
Underwriting verdict
HIGH RISK
70% confidence
Two adverse signals were known at decision time. The later bureau correction had not yet been recorded.
recorded timeknowledge accumulates →
Risk evidence recorded
Income verification added
Bureau correction received
Twin state
Relationships evolve too.
At this point, the decision is supported by the original adverse bureau signal.
revision 1
EntityApplication
BeliefAdverse bureau signal
EvidenceIncome verification
DecisionLoan-1842 verdict
RevisionBureau correction
Mar 04Apr 18Jun 01
Drag the timeline. The history stays put; the answer changes.
Two clocks, one record
Most databases tell you what is true now. InvariantDB can tell you what was true then — and what you knew about it at the time.
That distinction matters whenever facts arrive late, policies change, evidence is corrected, or an automated system must later explain why it acted.
01
Valid time
When the fact was true in the world.
02
Recorded time
When your system learned, stored, or revised it.
From explanation to evidence
Ask the historical question directly.
No bespoke event-sourcing layer. No log archaeology. Time is part of the query, provenance is part of the data, and the result can carry its own receipt.
Question
// What did the system know when it approved the loan? MATCH (b:Belief {id: "belief-99"}) AT RECORDED '2026-03-04' AT VALID '2026-03-11' RETURN b.claim, b.confidence, b.source, b.superseded_at;
Later, after the correction
MATCH (b:Belief {id: "belief-99"}) AT RECORDED '2026-06-01' RETURN b.claim, b.confidence;
Receipt
HIGH_RISK · 0.70r7a1…9cc2
valid2026-03-11
recorded2026-03-04
revisionbelief-99 · rev 1
sources2 evidence items
audit chain verified
LOW_RISK · 0.95bd41…0f73
valid2026-03-11
recorded2026-06-01
revisionbelief-99 · rev 2
reasonbureau correction
both answers remain provable
Replay. Trace. Prove.
InvariantDB turns the historical record into a queryable surface, so teams can reconstruct decisions instead of rebuilding them from scattered logs and snapshots.
02 / TRACE
Trace the evidence.
Walk from a decision to the beliefs, inputs, sources, actors, and revisions that produced it. Provenance lives with the record.
See decision lineage →03 / PROVE
Prove the history.
Hash-chained audit records make historical reconstruction tamper-evident, so an incident or audit starts with evidence instead of inference.
See regulated workflows →When the past matters.
AI is the high-urgency wedge. The underlying problem is broader: any consequential decision that may need to be reconstructed later.
Agentic twins
Agents execute. Twins persist.
Give autonomous systems a durable, evolving model of people, processes, assets, and organizations — including what the twin knew, believed, and relied on at each point in time.
“What did this twin believe when it made that decision — and which evidence and relationships supported it?”
Regulated decisions
Answer the auditor with the record itself.
Reproduce historical state and lineage for underwriting, claims, healthcare, financial controls, privacy, and other governed workflows.
“Show me the exact policy and evidence in force when this decision was made.”
Critical systems
Keep revisions without erasing history.
Model corrections as revisions rather than destructive overwrites, so old and new knowledge can coexist and remain explainable.
“What changed, when did we learn it, and which decisions relied on the old fact?”
Two clocksValid time + recorded time as first-class query semantics.
CypherTemporal graph queries without inventing a separate query model.
ProvenanceWho, when, source, reason, confidence, and lineage attached to values.
ReceiptsTamper-evident history you can verify and replay.
Start with one decision
Build systems that can explain themselves later.
Connect an agent, maintain an agentic twin, or replay a historical state. The underlying record stays inspectable from day one.