Geological & Assay-Data AI Agent (Sample to Database)
Validate drill-hole and sample data on import, reconcile assay certificates against the database and score the QA/QC controls — handing a geologist the exceptions, while a named Competent Person carries every reported figure.
Take drill-hole, collar, survey, logging and sample records as they arrive from the field or the geological database.
02
Read assay certificates as the laboratory issued them, carrying the method, the detection limits and the batch.
Reason
03
Reconcile each certificate value against the database entry it became, and mark where the two disagree.
04
Score the batch's standards, blanks, field and pulp duplicates and umpire results against the programme's own limits.
05
Check collars, downhole surveys, sample intervals and logging codes for the contradictions that move a sample.
Decide
06
Flag batches that fail their own control limits, samples with no certificate behind them, and method changes.
07
Route what it cannot resolve to the project geologist, the database administrator or the laboratory.
Out
08
Present the exception pack — each item beside the certificate, the batch and the control that failed.
09
Retain every certificate read, the check applied, the rationale and every geologist correction.
→Product statement
The agent checks, reconciles and flags. Accepting a batch, releasing a dataset for estimation and every reported figure stay with the named Competent Person.
Example workflow
One assay batch, end to end
AgentHuman
1Batch receivedA laboratory certificate, a logging upload or a survey file arriving against a drill programme
2Records resolvedHole, collar, survey, sample interval, batch, method and detection limit as the certificate states them
3Checks runCertificate reconciled to the database, QA/QC controls scored, collar and survey geometry checked
4Exceptions markedControl failures, samples with no certificate, interval gaps and overlaps, and changes of method
No human action required
Stages 1 to 4 run without a person in the loop — reading, reconciling and scoring the controls finish before anyone is asked to look at anything. Nothing is accepted in that stretch.
5DecisionSplits on check confidence and whether a control limit was breached
Controls inside their limits
Reaches the geologist as a checked batch.
Control failure or an unmatched value
Goes to the project geologist before anything uses it.
Project geologist
Reads the exceptions, the certificates behind them and the controls that failed, then decides what happens to the batch.
Accept findings · Correct · Request re-assay
Findings accepted — handed back▼
6Pack routed to the geologistHanded over as exceptions; no batch is accepted or rejected and no dataset is released by the agent
7Outcome evaluatedWhat the geologist corrected, what the re-assay returned, and what the database audit found by cohort
Corrections
Values a geologist corrects at review are counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required8 items
Estimating or classifying a Mineral Resource or Ore Reserve.
Reporting any figure as an Exploration Result.
Signing off a dataset as fit for estimation.
Accepting or rejecting an assay batch.
Automation boundaryAgent acts unaided
✓Reconcile each certificate value against the database entry it became.
✓Score the batch's standards, blanks and duplicates.
✓Check collar, survey and interval data for errors.
✓Assemble the exceptions with the failed control beside each one.
Write actions run only inside the approval boundaries agreed during implementation. A reported figure is never one.
Overriding a QA/QC control limit or failure rule.
Changing a logging code, standard or sampling protocol.
Editing a certificate value held in the database.
Deciding a sample swap is only a labelling error.
Example output
One assay batch, annotated
Everything the agent flags is attached to the certificate and the sample it came from.
Exception pack · one certificate batchIllustrative example
Batch as received
Method as stated
Detection limit
Flagged
Confidence
Batch
Certificate, 240 samples
Fire assay, 30 g charge
0.01 ppm
Certificate mismatch
91%
Not accepted by the agent
As receivedThe batch, the method and the detection limit as the certificate states them — nothing on this side is re-keyed.
Evidence usedLaboratory certificateBatch QA/QC insertsSample dispatch record
Why it is heldThe certificate and the database entry for that sample disagree.
ActionAccept findingsCorrectRequest re-assay
What the score decidesConfidence decides how hard the geologist reads the pack, not whether it is accepted.
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 batch entering the databaseCertificate, logging upload or survey file
03Checking & reconciling
Check what the laboratory issued
Use the programme's sampling protocol, insertion rates and control limits, and the method the laboratory states for that batch.
01Approved path
Find the error nobody thought to query
A swapped sample or a re-keyed value looks like ordinary data, so the check runs across every batch rather than the ones someone already doubted.
02Human review
Send the failed controls up first
Batches that breached their own limits, and values with no certificate behind them, reach the geologist before the data reaches a model.
04Build an evidence trail
Retain the certificates read, the controls scored, the exceptions raised, the confidence and every geologist correction — 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 the model and the database
Each control wraps the one inside it. A batch clears every layer before a geologist reads it, and the reported figure sits outside all six.
L6 · Outermost — last line of defenceInward → L1 · closest to the model
L6Rollback / safe modeReturn checking to the geology team if evaluations or production signals degrade.Roll back
L5Change controlRecord the certificate, the protocol version, the check applied, the named approver and every correction.Record
L4Competent-Person gateAccepting a batch, releasing a dataset and reporting any figure stay with the named person.Gate
L3Open-exception holdA batch with an open exception stays marked unresolved until a geologist closes it.Hold
L2Control limitsStandards, blanks, duplicates and umpire results are scored against the programme's own limits.Flag
L1Certificate as sourceValues are scored against the certificate the laboratory issued, not the database entry; one with no certificate behind it is raised.Cite
Model coreCheck result produced — reconciliation, QA/QC scores, data-integrity flags and confidence
L1 – L2Decide whether a value may stand
L3Decides what a geologist must close
L4 – L5Keep reporting with a person and the record intact
L6Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the whole check — not only the pack a geologist opens.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the exception pack the geologist opens
Depth of coverage ▼
E1Final-output evaluationDid the pack carry the exceptions that actually mattered?
E2Step-level evaluationDid it read the right certificate, batch, method and limit?
E3Tool evaluationDid it resolve the correct hole, interval and sample number?
E4Control recallHow many control failures were missed against the audited batch?
E5Slice evaluationHow does check quality change across specific batch cohorts?
E6Business outcomeHow much was corrected, and what did the re-assays return?
Floor — the database an estimate would be built from
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle.
Agent lifecycleDirection of processing →
01 · Retrieval2 modes
GA-01
Certificate not reconciled
A database entry differs from the certificate and passes.
GA-02
Superseded certificate used
The laboratory reissues a certificate and the first one stands.
Stage gathersCertificates, sample records, survey files and the protocol
02 · Reconciling2 modes
GA-03
Method change read as geology
A change of assay method is scored as a change in the ground.
GA-04
Over-limit result carried
A value at the upper detection limit is used as if measured.
Stage proposesMatched values, QA/QC scores and data-integrity flags
03 · Verification1 mode
GA-05
Sample swap not caught
Two exchanged samples each look ordinary on their own.
Which batches does this miss? Not grade control, which runs the same method every week — the ones where the method changed, the data is old, or the tool sat by steel. Nestack reports performance by slice, not only in total.
Slice performance — reported separately, not only in aggregateIllustrative example
Slice
Failure rate
Lift
Lift vs. threshold
Status
Batches after a method change
4.8%
3.1×
Review
Historic holes, legacy database
3.5%
2.3×
Review
Magnetic downhole surveys
2.9%
1.9×
Watch
Routine grade-control batches
1.2%
0.8×
Normal
Bar: corrected-or-re-assayed item rate, lift vs. routine grade-control baseline · scale 0–4.0× · tick marks the 2.0× review threshold2 of 4 slices over threshold
Evidence-linked improvement
One corrected value does not close the batch
Correcting the value leaves the reason it got there untouched. The check, the control limit or the import rule behind it changes under change control.
Improvement cycle · five stagesSwitchback — the path turns at Improve and returns at Learn
01Detect
Geologist corrections or re-assay results move in a cohort.
02Diagnose
Nothing is corrected until the check that let it through is named.
03Improve
The check or the limit goes through change control, with an approver.
04Verify
Run again over archived batches whose audit result is already known.
05Learn
The missed exception is added to the protocol later batches are scored against.
Learn → DetectThe return edge. Nothing shipped here changes what a Competent Person must still decide.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, certificates and database access, reconciliation and control scoring, evaluation, integration, then production validation and handover.
WorkstreamWeek 1Week 2Week 3Week 4Week 5Week 6
01Workflow discovery and automation-boundary definition.
02Sampling protocol and QA/QC limit mapping.
03Laboratory certificate and LIMS access.
04Drill-hole database and survey data access.
05Certificate-to-database reconciliation.
06Standard, blank and duplicate scoring.
07Collar, survey and interval validation.
08Method and detection-limit change detection.
09Evaluation suite, control recall and regression cases.
10Geologist review and correction capture.
11Exception pack and database write-back.
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 programme, one laboratoryProductionProduction database integrationAdvancedMulti-site / multi-commodity
Introduced at Pilot
Certificate-to-database reconciliation✓✓✓
Standard, blank and duplicate scoring✓✓✓
Umpire and check-assay comparison✓✓✓
Collar, survey and interval validation✓✓✓
Method and detection-limit change flags✓✓✓
Reporting stays with the Competent Person✓✓✓
Baseline evaluation✓✓✓
Introduced at Production
Geologist review workflow and write-back—✓✓
Observability and evaluation—✓✓
Introduced at Advanced
Multi-site and multi-commodity programmes——✓
Multiple laboratories and database systems——✓
Build priceFrom $5,000From $8,000Custom quote
Final build priceConfirmed after discovery based on certificate and database integrations, the number of programmes, laboratories and commodities, the QA/QC scheme in use, 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
01Your sampling and QA/QC protocol, with its insertion rates→Sampling protocol and QA/QC limit mappingWeek 1
02The reporting boundary and who your Competent Person is→Workflow discovery and automation-boundary definitionWeek 1
03Access to the laboratory's certificates or its LIMS feed→Laboratory certificate and LIMS accessWeek 2
04Access to the drill-hole database, collars and survey files→Drill-hole database and survey data accessWeek 2
05The standards, blanks and duplicates you insert, and their limits→Standard, blank and duplicate scoringWeek 3
06Batches that failed a control, and what you did about them→Evaluation suite, control recall and regression casesWeek 4
07Named geologists to read the exception packs→Geologist review workflow, then pilot batches and production validationWeeks 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 work and the first packs a geologist reads.
PhaseW1W2W3W4W5W6
DiscoveryW1
BuildW2 – W3
EvaluateW4 – W5
Pilot & LaunchW5 – W6
Week focusW1Sampling protocol, QA/QC limits and the automation boundaryW2Laboratory certificates, the drill-hole database and survey files wired inW3Reconciliation and control scoring on the first drill programmeW4Evaluation suite, control recall and method-change detectionW5Database write-back, supervised batches and targeted correctionsW6Corrections made by the geologists who own the data, then handover
Reading the bandWeek 4 scores the checks against batches you have already audited, so the recall figure comes out of your data, not ours.
At the end of W6Geologists who own the data have worked live packs and read corrections back, then Agent Care takes over monitoring.
DurationSix-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Mining AI agent
Build a geological data agent around your sampling protocol.
Show us your sampling and QA/QC protocol, one closed drill programme and the database it landed in. That programme is reconciled against the certificates it was built on, and the disagreements come back named, with the certificate behind each one.
