Run the disproportionality and time-to-onset analyses on your own safety data, assemble the evidence behind a candidate signal and track it — your validation lead and safety review board decide what it is.
Take the safety-database extract, the literature findings and the class information agreed for each product in scope.
02
Read the exposure denominator from the source your company uses, with its date and its basis recorded beside it.
Reason
03
Run disproportionality on the strata you agreed, and fit time to onset where the reported dates will carry it.
04
Test what else moves a count — a reporting surge, a masking product in the same class, or the indication treated.
05
Assemble the case series and its quality, the current label position, the literature and the exposure behind it.
Decide
06
List candidates with the evidence behind each, and mark what only a named person may settle.
07
Carry forward what no statistic raised — a designated medical event, a single documented case, a class finding.
Out
08
Record the candidate in your tracking system with the run, the parameters and what was ruled out.
09
Keep the extract, the parameter sets, the package as it went and the changes a reviewer made.
→Product statement
The agent runs, tests, assembles and tracks. Validating a signal, prioritising it, refuting it, closing it and deciding what an authority is told stay with your named validation lead, the safety review board and your QPPV.
Example workflow
One candidate, end to end
AgentHuman
1Data assembledThe safety-database extract, literature findings and class information, with the exposure it will be read against
2Analyses runDisproportionality on the strata your procedure names, and time to onset where the reported dates will carry it
3Context testedWhether a reporting surge, a masking product in the class or the indication treated explains the movement instead
4Package assembledThe case series and its quality, the current label position, the literature and the exposure behind the candidate
No human action required
Stages 1 to 4 run without a person in the loop — the analyses, the context tests and the package are finished before a reviewer opens anything. Nothing is validated in that stretch.
5DecisionSplits on evidence strength and the context tests
Evidenced, context tested
Reaches the board's list ready to read.
Thin, or an artefact suspected
Held with the failing test named, not dropped.
Validation lead and review board
The validation lead settles whether the candidate is a signal; the board and your QPPV settle what is prioritised, what is refuted and what an authority is told.
Board validates · Correct the package · Ask for more evidence
Validated — handed back▼
6Candidate trackedWritten to the tracking system only where write access and policy allow; the validation and closure fields stay empty
7Outcome evaluatedWhat the board validated, what it refuted, what reached it without a statistic, and how long each candidate waited
Send-backs
Every candidate the board sends back is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required8 items
Deciding a signal is real, or refuting one.
Closing a signal or setting one aside.
Determining causality at population level.
Prioritising the safety review board's agenda.
Automation boundaryAgent acts unaided
✓Run disproportionality and time to onset on the strata you agreed.
✓Test for a reporting surge, a masking product or the indication.
✓Assemble the case series, label and exposure.
✓Carry forward what no statistic raised.
Write actions run only inside the approval boundaries agreed during implementation. Validation status is not among them.
Changing a label or a risk-management plan.
Deciding a signal need not be reported.
Notifying an authority of a validated signal.
Changing thresholds, strata or tracking-system status.
Example output
One candidate, annotated
Everything the agent puts forward stays attached to the run and the extract it came from.
Signal-detection output · single candidateIllustrative example
Pair
Sources
Exposure
Status
Confidence
Causality
One product, one event
Safety database, literature
Patient-years on file
Candidate, not validated
89%
Not assessed by the agent
As receivedThe extract, the sources it came from and the exposure it is read against — nothing here is computed.
Evidence usedStratified disproportionalityTime to onset, dates usableNot in the current label
Why it is only a candidateDisproportionate reporting is an association in a database, not a rate and not a cause.
ActionBoard validatesCorrect the packageAsk for more evidence
What the score decidesConfidence orders the reading queue. It does not validate the candidate or rank it.
Value
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value landsValue 01 – 04
Every product and event pairFrom the safety database, literature and the class
03Signal management
Read a statistic in its own context
Test the exposure denominator, the reporting surge, the masking product in the class and the indication treated before a candidate is put to anyone.
01Approved path
Get past assembling the package
The runs, the case series, the label position and the literature are pulled together before a reviewer opens anything, so the hour goes on judgement.
02Human review
Send the board what it can rule on
A candidate arrives with the exposure it stands on, what was ruled out and what could not be tested, rather than a table of ratios with no context.
04Build an evidence trail
Retain the extract, every parameter set, what was ruled out, the package as it went, the board's decision and the tracking status — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between a statistic and the agenda
Each control wraps the one inside it. A candidate clears every layer before a reviewer reads it, and validation sits outside all six.
L6 · Outermost — last line of defenceInward → L1 · closest to the model
L6Rollback / safe modeReturn signal detection to your team if evaluations or production signals degrade.Roll back
L5TraceabilityRecord the extract, the parameters, what was ruled out, the package and every reviewer change, under your change control.Record
L4Non-statistical routesDesignated medical events, literature and class findings are routed without waiting on a statistic; a serious one ranked below a statistic is measured here.Route
L3Validation gateValidation, prioritisation, refutation and closure stay with your validation lead, the board and your QPPV.Gate
L2Context testsReporting surges, masking products and the indication treated are tested for, and the tests that could not run are listed.Test
L1Denominator checkThe exposure a rate stands on is named and dated; where it will not fit, the candidate is marked rather than rated.Check
Model coreCandidate raised — the statistic, the stratum, the case series, the exposure and confidence
L1 – L2Decide whether a movement may stand
L3Decides who validates and prioritises
L4 – L5Keep the quiet routes open and traced
L6Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the method — not only the candidates it put up.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the candidate package the board opens
Depth of coverage ▼
E1Package faithfulnessDoes the package state what the extract, the label and the literature show?
E2ReproducibilityDo the same parameters on the same extract return the same candidates?
E3Tool evaluationDid it read the right product, stratum, date range and exposure file?
E4Reference-set recallWould it have put up the candidates your board validated before?
E5Artefact discriminationAre surges, masking and indication told apart from a real movement?
E6Slice evaluationHow does candidate quality change across products, events and data ages?
Floor — the signal a named person validates and answers for
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle. Missing a real signal and burying the board in artefacts both start here.
Agent lifecycleDirection of processing →
01 · Data assembly1 mode
SG-01
Exposure that never fitted
A rate is computed against a denominator for another market.
Stage gathersThe extract, the sources and the exposure denominator
02 · Quantitative run2 modes
SG-02
Smaller signal stays masked
A large reported event in the class holds the ratio down.
SG-03
Reporting surge read as risk
Press coverage or a litigation campaign moved the count.
Stage computesDisproportionality and time to onset on agreed strata
03 · Evidence package1 mode
SG-04
Indication read as effect
The event belongs to the disease the product treats.
Stage assemblesThe case series, the label position and the literature
04 · Queue / output2 modes
SG-05
Ranked by alert, not by harm
A serious event with no statistic sits below one with.
SG-06
Window spent in the tracker
A validated signal waits unassigned while its clock runs.
Stage presentsWhat the board reads, with the evidence behind each
05 · Change / Version1 mode
SG-07
Strata moved without a record
A subgroup boundary changes and the candidate list changes.
Stage tracksModel, prompt, threshold and stratification changes
Sev-1 · a real signal is not put to anyoneSev-2 · the board decides on a distorted pictureSev-3 · review time goes on a reporting artefact
Thin data and crowded classes cost the most rework
The count here is candidate signals a reviewer sent back — re-run on another stratum, dropped as an artefact, or returned for evidence the package lacked. None of it says a cohort deserves less scrutiny than the baseline.
Slice performance — reported separately, not only in aggregateIllustrative example
Slice
Failure rate
Lift
Lift vs. threshold
Status
Newly authorised products
4.9%
3.2×
Review
Rare events, few reports
3.8%
2.5×
Review
Products in a crowded class
2.7%
1.8×
Watch
Long-established products, stable use
1.4%
0.9×
Normal
Bar: sent-back-candidate lift vs. established-product baseline · scale 0–4.0× · tick at the 2.0× threshold2 of 4 slices over threshold
Evidence-linked improvement
A refuted candidate is evidence about the method
What the board threw out says something about the stratum, the threshold or the context test behind it. That is what moves, before the next run.
Improvement cycle · five stagesSwitchback — the path turns at Improve and returns at Learn
01Detect
Send-backs or refuted candidates gather on one kind of product.
02Diagnose
If another stratum makes the candidate vanish, the stratum was the finding.
03Improve
The stratum, threshold or context test moves under change control, approver named.
04Verify
The reference set is run again, the candidates your board validated among it.
05Learn
The refuted candidate is kept as a case the next run has to account for.
Learn → DetectThe return edge. A threshold or a stratum is part of a described, inspectable method — it moves with a named approver and a record, not quietly between runs.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, data and denominators, methods and context tests, evaluation, tracking integration, then production validation and handover.
WorkstreamWeek 1Week 2Week 3Week 4Week 5Week 6
01Signal-management discovery and boundary.
02Exposure denominators and their sources.
03Safety-database extract and refresh rules.
04Strata, thresholds and method selection.
05Disproportionality runs on your own data.
06Time-to-onset analysis where dates allow.
07Masking, surge and indication context tests.
08Designated-medical-event and literature routes.
09Reference-set recall and artefact discrimination.
10Evidence packages in your board's format.
11Tracking-system write-back and audit trail.
12Observability, deployment and Agent Care handover.
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallelFinal scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
Capability✓ in scope · — not at this tierPilotOne product, one extractProductionProduction tracking-system linkAdvancedMulti-product / multi-region
Introduced at Pilot
Disproportionality on your own data✓✓✓
Time-to-onset analysis where dates allow✓✓✓
Exposure denominator with its source named✓✓✓
Masking, surge and indication tests✓✓✓
Routes that do not wait on a statistic✓✓✓
Evidence package per candidate✓✓✓
Validation and prioritisation by named people✓✓✓
Traceability and baseline evaluation✓✓✓
Introduced at Production
Tracking-system write-back and observability—✓✓
Additional products and data sources—✓✓
Introduced at Advanced
Multi-region reporting and enterprise controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build priceConfirmed after discovery based on products and data sources in scope, safety-database and tracking-system integrations, exposure data, the methods and strata your procedure names, review controls and deployment requirements.
Separate from buildBuild pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bringWe build with it
01The products and the safety data your signal management runs on→Signal-management discovery and boundaryWeek 1
02Where your exposure figures come from, and what they count→Exposure denominators and their sourcesWeek 1
03Access to the safety database and the extract you work from→Safety-database extract and refresh rulesWeek 2
04The strata, thresholds and methods your procedure names→Strata, thresholds and method selectionWeek 2
05The format your review board and your QPPV read→Evidence packages in your board's formatWeek 4
06Candidates your board validated, and the ones it refuted→Reference-set recall and artefact evaluationWeek 4
07A named validation lead and tracking-system access→Tracking-system write-back, then supervised runsWeeks 5–6
Nothing else is requiredDeployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Phases are drawn over the weeks they actually occupy. Week 5 carries both the evaluation slices and the first packages a board reads.
PhaseW1W2W3W4W5W6
DiscoveryW1
BuildW2 – W3
EvaluateW4 – W5
Pilot & LaunchW5 – W6
Week focusW1The boundary, the denominator sources and who validatesW2Safety-database extracts, strata and the refresh scheduleW3Disproportionality, time to onset and the context testsW4Reference-set recall, artefact tests and the board's packageW5Packages read by your board, tracking write-back, correctionsW6Production validation, threshold review and Agent Care handover
Reading the bandThe exposure denominator is settled in week 1 rather than fitted later, because a rate nobody can source is a rate the board cannot use. That dependency is what the bars show, not a smooth ramp.
At the end of W6Your board has read a package it did not have to assemble. Thresholds, strata and denominator sources sit with Agent Care from week 7 onwards.
DurationSix-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Biotechnology AI agent
Build signal detection your board and your QPPV can follow.
Show us one product, the data your signal management runs on and how a candidate reaches your safety review board today. The first conversation is not about detection rate — it is about what your method can reproduce, what your exposure data can carry, and which routes into that board do not run on a statistic at all.
