EvidenceWorks
From inspection event to customer-safe proof
Every pilot event becomes sealed evidence, review context, and a customer-safe go/no-go packet.
Step 01
Pilot brief
Define what the pilot must prove before promotion.
Inputs
- Inspection point
- Part family
- Defect classes
- Pilot scope
Proof generated
- Pilot scope record
- Defect-class list
- Promotion criteria
Where it appears in the app
- Commissioning
- Stations
- Capabilities
AI Sense support
- Flags incomplete pilot scope
- Flags missing evidence requirements
Safety boundary
- Defining scope never promotes a station.
Step 02
Capture event
Capture what happened at the station.
Inputs
- Camera frame
- Part outline
- Timestamp
- Station ID
Proof generated
- Frame reference
- Defect marker
- Capture timestamp
Where it appears in the app
- Inspect
- Continuous
- Evidence
AI Sense support
- Detects missing frames
- Explains low confidence
- Flags blur, glare, and incomplete capture
Safety boundary
- Capture records evidence; it never forces a verdict.
Step 03
Evidence capsule
Package proof by reference, not raw images.
Inputs
- Frame reference
- Overlay reference
- Coverage map
- Measurement
- Model version
- Rule version
Proof generated
- Sealed evidence capsule
- Content hash
- Model + rule versions
Where it appears in the app
- Review
- Evidence
- Support bundle
AI Sense support
- Explains what evidence exists
- Explains what evidence is missing
Safety boundary
- Sealing references, not raw images — no raw frame leaves the station.
Step 04
Coverage proof
Prove the required views were seen. No coverage, no PASS.
Inputs
- Required faces
- Captured views
- Coverage map
Proof generated
- Covered / missing regions
- Coverage receipt
Where it appears in the app
- Vision Twin
- Review
- Work packages
AI Sense support
- Finds repeated coverage gaps
- Recommends human checks
Safety boundary
- Missing required views block a PASS — coverage is proven, never assumed.
Step 05
Rule decision
Apply the active criteria pack. Rules decide, AI detects, evidence proves.
Inputs
- Active criteria pack
- Measurement
- Threshold
Proof generated
- Rule-fired record
- Decision receipt
Where it appears in the app
- Governance
- Review
- QA decision ledger
AI Sense support
- Explains why a rule fired
- Explains whether evidence supports it
Safety boundary
- A rule decision is a deterministic result, never an AI approval.
Step 06
Operator review
Ambiguous cases go to a person, not auto-clear.
Inputs
- Uncertain case
- Reviewer note
- Evidence capsule
Proof generated
- Human review card
- Review decision receipt
Where it appears in the app
- Review
- Review analytics
AI Sense support
- Finds review bottlenecks
- Flags missing QA context
Safety boundary
- Uncertain cases are never auto-cleared; a person decides.
Step 07
False-reject candidate
Near-threshold rejects are tracked, not hidden.
Inputs
- Near-threshold reject
- Confidence band
- Human review marker
Proof generated
- False-reject candidate record
- Candidate receipt
Where it appears in the app
- Review analytics
- Work packages
AI Sense support
- Ranks false-reject patterns
- Suggests human checks
Safety boundary
- Candidates are tracked for review — never silently discarded.
Step 08
Escape candidate
Possible missed defects stay visible for human review.
Inputs
- Possible missed defect
- Support flag
- Review trail
Proof generated
- Escape candidate record
- Risk marker
Where it appears in the app
- Review analytics
- Support handoff
- SignalOps
AI Sense support
- Flags repeat escape risks
- Flags evidence gaps
Safety boundary
- Possible escapes are surfaced, never hidden.
Step 09
Drift watch
Watch the station before trust degrades.
Inputs
- Model confidence trend
- Lighting drift
- Pose drift
- Station health
Proof generated
- Drift signal report
- Station health line
Where it appears in the app
- Vision Twin fleet
- Vision Twin knowledge
- Ops metrics
AI Sense support
- Predicts attention priority
- Prepares calibration work packages
Safety boundary
- Drift is observed and advised; it never changes a verdict.
Step 10
PLC route agreement
HoldField observes and compares the route. PLC authority remains highest.
Inputs
- Governed result
- Observed route
- Safety-chain state
Proof generated
- Route-agreement record
- Agreement check
Where it appears in the app
- Governance
- Commissioning
- Ops metrics
AI Sense support
- Explains routing mismatch
- Flags missing PLC evidence
Safety boundary
- Route agreement is observed and compared — never a PLC write.
Step 11
Go / no-go packet
Assemble the customer-safe recommendation. The packet recommends; it does not approve itself.
Inputs
- Evidence completeness
- Coverage completeness
- Review status
- Known limitations
Proof generated
- Go/no-go recommendation packet
- Completeness summary
Where it appears in the app
- Commissioning
- SignalOps
- Readiness packet
AI Sense support
- Summarizes blockers, risks, and missing proof
Safety boundary
- The packet is a recommendation for a human — it never approves production.
Step 12
Known limitations
Proof includes what the system cannot claim.
Inputs
- Limitations ledger
- Model validation status
- Sampling note
Proof generated
- Known-limitations ledger
- Human-acceptance-required flag
Where it appears in the app
- Trust
- Governance
- Customer acceptance
AI Sense support
- Prevents unsupported claims
- Prevents confidence inflation
Safety boundary
- Human acceptance is required; the ledger records what cannot be claimed.
AI Sense
AI Sense reads across the whole flow
Observes evidence, finds missing proof, explains uncertainty, ranks human checks, and prepares handoffs — it never commands hardware.
Reads
- Evidence bundles
- Review events
- QA decisions
- Vision Twin drift
- Commissioning blockers
- Governance decisions
- Station registry
- Ops metrics
Produces
- Findings
- Evidence-gap warnings
- Work-package hints
- Commissioning questions
- Support summaries
Never
- No PLC writes
- No force PASS
- No recovery clear
- No robot commands
- No camera/light commands
- No production approval
- No evidence mutation
- No QA decision mutation
AI Sense observes evidence and guides humans — it records nothing and changes nothing. It does not command a station, write a PLC, clear recovery, reset safety, force a pass, approve production, sign off, or mutate any review, QA decision, commissioning, governance, evidence, or runtime state. Every recommendation is a suggestion for a human to carry out; the PLC and safety circuit remain authoritative.
Customer-safe export
Proof by reference, never raw data
The redacted packet carries references, hashes, receipts, and an explicit omissions list — never raw images, file paths, identities, tokens, or PLC registers.
- evidence_refs
- capsule + frame/overlay references
- hashes
- content + decision receipt hashes
- coverage
- covered / missing regions (counts)
- review
- human review status + note references
- recommendation
- go / no-go (recommendation only)
- known_limitations
- what the system does not claim
- redaction_status
- redacted — no raw images or identities
- omissions
- explicit list of what was withheld
Signed-in teams run this operationally in the HoldField app, under EvidenceWorks and Capabilities. Open the workspace →