Screen 01[BUILT]

Connect Evidence

The AI supplier submits an immutable candidate reference; evidence sources establish accountability for the decisions already being made. Notary AI preserves and verifies submitted evidence — it does not create or deploy the change.

Priya Ramanathan

Enterprise AI decision accountability

Company XYZ decision accountability · LoanCore Underwriting 3.2.0Lending · Applicant A-1027— Representative lending boundary case

ACCOUNT1–3→
VERIFY4–7→
DEFEND8–9

ACCOUNTScreen 01

Question: What evidence does Company XYZ have about the decisions its AI systems make?
Answer: Evidence is arriving from AI systems, observability platforms, and systems of record, but coverage and authority vary by decision family.
Output: Provenance-preserving evidence inventory.

Where Notary AI fits

Evidence comes from the stack. Assurance travels back into it.

Notary AI does not replace observability, governance, GRC, or business systems. It binds their evidence into a decision-level verification record and returns scoped assurance results to the workflows where accountable teams already operate.

[BUILT]operational foundation[PLANNED]production integrations

Systems that hold the facts

Decision-time capture

Notary AI SDK · customer evidence bundles

Logs, traces & evaluations

LangSmith · Arize · Datadog

Governance & GRC

Credo AI · OneTrust · ServiceNow

Enterprise systems

Business workflows · Snowflake · ticketing

Independent verification layer

NOTARY AI

ACCOUNT

Bind the record

VERIFY

Test the change

DEFEND

Issue the evidence

Decision + outcomePolicy + authorityEvidence + provenanceVerification + scope

PLANNED · asymmetric signing and offline verification

Systems that use the answer

Evidence readiness

Coverage · provenance · accountability gaps

Release workflow

Candidate result · accountable release decision

Audit & dispute record

Scoped evidence package · explicit limitations

Governance write-back

GRC · model risk · ticketing · CI/CD

Logs explain a run. Governance records a control. Notary AI verifies the decision across them.

Evidence sources — Company XYZ tenant

45,657 decisions · 7d

Every AI decision family Company XYZ operates has at least one connected source. Coverage varies — some sources emit rich, replayable evidence; others still have gaps.

LoanCore Underwriting · Lendingconnected · 86%
CarePath Prior-Auth · Healthcare Prior-Authconnected · 72%
NorthStar Support Bot · Customer Supportconnected · 64%
ContactHub Escalations · Contact Centerconnected · 58%
TalentScreen ATS · Hiringnot connected · 31%

LoanCore candidate provenance · Applicant A-1027, personal loan $18k

Rule failure

This is the evidence Notary AI has preserved for Applicant A-1027 — the candidate reference this walkthrough carries through every subsequent screen.

12:04:17 · LoanCore SDK
forensic-snapshot A-1027 sealed · root sha256:9f2c…a17b
12:04:18 · Snowflake · LoanCore DW
observation: bureau_pull → 2 tradelines, 0 delinquencies (provenance: LOANCORE.BUREAU_PULLS)
12:04:18 · Datadog · OTLP
observation: rule_gate hit → if score<700 deny

How evidence flows in

45,657 decisions · 7d

Producing systems

Instrumented agents

LoanCore Underwriting · CarePath Prior-Auth

AI observability

LangSmith · Arize AX · Datadog OTLP

Systems of record

ServiceNow · Snowflake · Greenhouse

Ingest surface

Decision Evidence Protocol

Canonical resource types: forensic-snapshot, observation, context, assessment, evidence-bundle. Provenance preserved.

SDK captureDEP providerSource connectorOTLP bridgeManual submission

Sealing: HMAC-SHA256 Merkle → root hash · duplicate → quarantine

Notary AI evidence store

Immutable log

Append-only. Content-addressed. Scoped to org · company-xyz / env · demo. Status is derivable from transitions; never overwritten.

confirmed

inferred

gap

Configured connectors — Company XYZ tenant

9 sources

SDK captureIn-process Forensic Logger. Sealed at decision time.

LoanCore Underwriting agent
LoanCore (AI vendor)
connected
Notary AI Forensic Logger wrapped around the underwriting agent. Every application produces a sealed Forensic Snapshot at decision time.
emits: forensic-snapshot · decision-evidence-graph · context-binding
Lendingsync · 2m ago · 9,310 bundles
CarePath Prior-Auth agent
CarePath Health AI
connected
In-process SDK capture of prior-auth determinations, including the criteria bundle version and any high-risk clinician notes.
emits: forensic-snapshot · context-binding · risk-metadata
Healthcare Prior-Authsync · 6m ago · 4,128 bundles

DEP providerNative Decision Evidence Protocol emitter.

LangSmith
LangChain, Inc.
connected
Emits DEP resources from the NorthStar support bot's traces — LLM completions, retrieval events, and tool-call payloads.
emits: observation · context · assessment
Customer Supportsync · 34s ago · 21,440 bundles
Arize AX
Arize AI
connected
Model-invocation telemetry from the support LLM v1.4. Provides token-level context Notary AI correlates to Forensic Snapshots.
emits: observation · risk-metadata
Customer Supportsync · 1m ago · 3,200 bundles

Source connectorConfigured integration with a system of record.

ContactHub Escalations (ServiceNow)
ServiceNow
connected
Pulls escalation cases and disposition history, preserving ServiceNow record IDs as provenance on every ingested case.
emits: observation · review-decision
Contact Centersync · 12m ago · 6,905 bundles
LoanCore data warehouse
Snowflake
connected
Cursored pull of bureau-pull results and adverse-action records. Provenance carries the Snowflake account, database, and table path.
emits: observation · evidence-bundle
Lendingsync · 18m ago
TalentScreen ATS
Greenhouse
pending
Reads screening scores, feature vectors, and reject/advance events. Currently partial — only 31% of decisions include feature attributions.
emits: observation · risk-metadata
Hiringsync · 3h ago · partial · 3,874 bundles

OTLP bridgeBridged from existing OpenTelemetry traces.

Agent traces bridge
Datadog · OTLP
connected
Existing OpenTelemetry traces are bridged into DEP observations without re-instrumenting agents.
emits: observation
Cross-domainsync · 9s ago

Manual submissionFile / bundle import. Provenance preserved.

Cassette bundle import
Local — Company XYZ evidence pack
connected
Manually submitted evidence bundles from historical incidents. Provenance is preserved as 'manual-submission'.
emits: evidence-bundle
Cross-domainsync · on demand

[BUILT]Import sealed snapshot

Loads a content-addressed cassette into the . Append-only, duplicate → quarantine.

What each connector produces

forensic-snapshot: Full decision-chain trace, sealed and Merkle-rooted at capture.
observation: A discrete telemetry event: model invocation, retrieval, tool call.
context: Decision-time config & policy versions — .
assessment: Human or reviewer disposition attached to a decision.
evidence-bundle: A packaged set of the above, portable across systems.

Seeded source corpora (legacy view)

LoanCore Underwriting86%
CarePath Prior-Auth72%
NorthStar Support Bot64%
ContactHub Escalations58%
TalentScreen ATS31%

Sealing & protocol detail

Decision Evidence Protocol (DEP)

Every source terminates in the . Bytes on the wire differ per source, but each arriving bundle becomes a canonical set of DEP resources: forensic-snapshot, observation, context, assessment, evidence-bundle. A is HMAC-SHA256 sealed as a Merkle tree — its root hash is the identity of the record.

The sealing boundary — not a trust promise

Notary AI seals what a source emits; it does not vouch that the source was honest. Every downstream certificate carries an explicit . This is what stays inspectable and attributable to its source under review.

Screen anatomy

What this screen establishes

Notary AI establishes the provenance, integrity, and inventory of the AI-decision evidence submitted from each connected source — instrumented agents (SDK), vendor observability platforms that speak DEP natively, systems of record via connectors, and existing OpenTelemetry pipelines. Ingestion does not establish that the underlying source data is itself true; it establishes exactly what was submitted, by whom, and that it has not been altered since arrival.

How this works

Every arriving bundle becomes a canonical DEP resource set, sealed and Merkle-rooted before it reaches the . See "Sealing & protocol detail" above for the full mechanism.

The sealing boundary

Sealing is not a trust promise about the source. See .

In the requirements

REQ-FP-DES-002, REQ-FP-MSSC, REQ-FAL-005, REQ-FP-005 — see full quotes in "Sealing & protocol detail" above.

Evidence has been preserved with its provenance. Next, Company XYZ needs to understand what that evidence can actually establish.

Decision Landscape →