Collapse the citation chain to its origin before a figure is counted, and hold each finding with the source it rests on, the date that source was published and the method behind it.
A question is set, and the sources able to answer it are listed before any figure is quoted.
02
A source is offered, and the chain behind it is walked back to whoever published the number first.
Reason
03
A figure recurs across three articles, and one unsourced release beneath them is still one source.
04
A vendor report is cited, and its author is labelled a seller in the category the report sizes.
05
A citation is drafted by a model, and it is retrieved and checked to exist before it is reported.
Decide
06
A panel wave closes, and who was asked, how many, when and in which markets rides with the result.
07
A respondent answers, and the response sits under your own retention and consent settings.
Out
08
A source cannot be reached, and the figure is reported untraceable rather than repeated.
09
Execute write actions only inside the approval boundaries agreed during implementation.
→Product statement
The agent gathers, traces, labels and drafts. It vouches for no source, only for what that source says and where it came from. A named researcher accepts the finding into the record.
2Question context assembledThe question, the markets it covers, the sources able to answer it and when each was published
3Draft finding assembledThe finding, its sources, the chain behind each and completeness
4Controls appliedChain-collapse checks, retrieval checks, sample and recency checks and completeness confidence
No human action required
Stages 1 to 4 run unaided, and nothing is accepted at any of them — the agent is sourcing, and the review lane opens at the completeness gate.
5DecisionSplits at the completeness gate
Evidence sufficient
Goes to the named researcher to accept.
Anything thin
Adds an insights lead read first.
Insights review
The finding is held with its sources, their dates and the chain collapsed behind each.
Accept · Append source · Send to insights review
Accepted — by a named researcher▼
6Finding and source records updatedOnly where write access and records policy allow it
7Outcome evaluatedSource traceability, chain depth, reviewer corrections and what the read found
Corrections
Each insights correction is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required8 items
Accepting a finding into the research record.
Deciding a vendor number may stand unlabelled.
Signing off a market size for a strategy paper.
Setting the panel consent and retention policy.
Automation boundaryAgent acts unaided
✓Collapse each citation chain to the first publication behind it.
✓Record the date a source was published and the method it used.
✓Retrieve every cited source and check that it actually exists.
✓Label a category report published by a seller in that category.
Nothing is accepted or published except by a named person, inside the agreed boundaries.
Judging whether a source is sound.
Telling the business what the market is worth.
Setting the sampling standard a study is held to.
Changes to panel data, consent or retention.
Example output
One finding, annotated
No market size is vouched for anywhere on this page; the record below is simply what one research question turned out to rest on.
Record entry · single questionIllustrative example
Question
Recorded as
Source
Evidence of record
Confidence
Held for
Category size, desk research
Traced to origin, vendor-published
Vendor report
Source dated 3 August 2026
Held unaccepted
The accepting researcher, by name
As receivedTaken from the source document and the chain collapsed behind it, and it reaches no further than they do.
What the record holdsSource documentPublication dateMethod statement
Why no acceptance hereWhether a source is sound enough to rely on is a researcher call.
ActionAcceptAppend sourceSend to insights review
What the score decidesBelow the configured threshold a finding picks up an insights read before acceptance.
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 findingFrom the source that carries it
03Evidence
Where the evidence is used
The advertising pages we ship are agency-side and serve a client; this one answers to the brand's own insights and legal functions, which defend the number in a strategy paper years later.
01Approved path
A finding is not a fact
Three articles citing each other back to one unsourced press release are not three sources, which is why every chain is collapsed to its origin before anything is counted.
02Human review
What was checked, and not found
No original publisher was located for the category figure the trade press repeats, the panel provider publishes no method statement for the wave in question, and no independent replication of the vendor sizing was found.
04Build an evidence trail
The finding, the source it rests on and the researcher who accepted it stay on file.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Published and desk sourcesTrade press · statistical offices Source documents and dates
Panel and survey platformsPanel providers · survey tools Respondent-level responses
First-party evidenceCRM · sales and pricing data Internal comparisons
Agent
Market research sourcing
Reads the questions Traces the sources Holds for the researcher
Records and case systemsResearch repository · ticketing Finding and acceptance records
Integration availability depends on the client's existing systems and API access.
Agent controls
Six sieves between the model and the researcher
Six probes pushed in order, the last the deepest. Whatever still stands is set out in the map below.
L6 · Outermost — last line of defenceInward → L1 · closest to the model
L6Rollback / safe modeNarrow the agent to source listing when evaluation or production signals degrade.Roll back
L5Version monitoringTrack model, prompt and sourcing rules, and re-run the fabricated-citation suite whenever any of the three moves.Track
L4TraceabilityRecord each finding, the sources under it, the chain collapsed behind each and every read of the file.Record
L3Researcher acceptanceHold the finding for a named researcher; the hold governs acceptance, not whether the source underneath is any good.Gate
L2Sourcing guardrailsTest each finding against the sourcing standard you configure: chain collapsed, author labelled, sample and recency stated.Restrict
L1Confidence thresholdsRoute a thinly sourced finding to an insights read; a plausible citation that resolves to nothing is the worst output here.Require review
Model coreEvidence assembled — the question, its sources, the chains and completeness
L1 – L2Test whether a finding may stand
L3Puts the acceptance in a person's hands
L4 – L5Keep the finding and the source behind it
L6Returns to source listing when signals degrade
How Nestack evaluates it
Evaluate the whole assembly — not only the finding that comes out.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the finding a strategy paper quotes
Depth of coverage ▼
E1Final-output evaluationDid the entry record the source a finding actually rests on?
E2Step-level evaluationDid the agent read the right question, the right market and the current edition?
E3Tool evaluationDid it retrieve the cited source and file it against the right question?
E4Confidence calibrationDo low-confidence findings actually attract more insights corrections?
E5Slice evaluationHow does performance change across specific question types?
E6Business outcomeHow many findings needed a correction before a researcher accepted?
Floor — the finding the business defends
Failure modes
Where each failure originates in the agent
Seven failure modes, each pinned at the stage where it first shows.
Agent lifecycleDirection of processing →
01 · Retrieval1 mode
LN-03
Stale source read
The edition read has been superseded by a later one.
Stage gathersThe questions, the sources, the dates and methods
02 · Reasoning2 modes
LN-04
Fabricated citation
A cited source is plausible and resolves to nothing.
LN-06
Vendor claim run neutral
A seller in the category is reported unlabelled.
Stage proposesThe findings, their chains and completeness
03 · Tool / write2 modes
LN-02
Thin finding passed forward
A finding moves on without the insights read.
LN-05
One chain counted repeatedly
Articles on one release are counted separately.
Stage writesOnly where write access and approval policy allow it
04 · Output1 mode
LN-01
Accepted, source unrecorded
The record shows acceptance but not what it rested on.
Stage returnsThe finding a researcher accepts and a paper quotes
05 · Change / Version1 mode
LN-07
Silent sourcing regression
A prompt change loosens the chain rule, not the record.
Stage tracksModel, prompt, sourcing rules and finding fields
Sev-1 · a fabricated source is citedSev-2 · vendor claim reported as neutralSev-3 · source degrades, finding held back
A question-level source-quality figure can read clean while market-sizing questions carry most of the rework. Nestack reports the correction rate by research question, not only in total.
Slice performance — reported separately, not only in aggregateIllustrative example
Slice
Failure rate
Lift
Lift vs. threshold
Status
Market-sizing questions
8.2%
3.6×
Review
Competitor-claim questions
5.9%
2.6×
Review
Pricing and willingness questions
3.6%
1.6×
Watch
Usage and behaviour questions
1.4%
0.6×
Normal
Bar: correction-rate lift vs. usage-question baseline · scale 0–4.0× · tick marks the 2.0× review threshold2 of 4 slices over threshold
Evidence-linked improvement
What an uncited claim costs
A loop ends when the uncited claim has become a case the next release must pass. That suite is what the next study issued is measured against.
Improvement cycle · five stagesSwitchback — the path turns at Improve and returns at Learn
01Detect
Correction rate rises on market-sizing questions.
02Diagnose
The market figure everyone repeats and nobody can trace to a source is worked back until one cause remains.
03Improve
Number the change; the findings that drove it are filed beneath it.
04Verify
Each touched finding case runs once more, and one red holds it back.
05Learn
It stays on as a standing test, and the sourcing rules travel with it.
Learn → DetectThe return edge. The next study is measured against a suite one case longer.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, finding assembly, evaluation, integration, then production validation and handover.
WorkstreamWeek 1Week 2Week 3Week 4Week 5Week 6
01Citation-chain discovery and automation-boundary work.
02Desk, panel and vendor-report sources.
03Question-to-source and vendor-labelling rule mapping.
04Source and respondent ingestion.
05Finding, source and date binding.
06Traceability scoring and review routing.
07Researcher acceptance workflow.
08Research-repository integration.
09Sourcing and sampling cases.
10Guardrails and acceptance controls.
11Finding-trail instrumentation.
12Deployment, documentation 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 question type, one quarterProductionProduction acceptance workflowAdvancedMultiple markets / categories
Introduced at Pilot
Source tracing to your questions✓✓✓
Named researcher acceptance✓✓✓
Source-inventory baseline✓✓✓
Introduced at Production
Reporting by research question—✓✓
Acceptance workflow in your systems—✓✓
Approved write-back—✓✓
Panel-and-desk integration—✓✓
Introduced at Advanced
Multi-category programmes——✓
Cross-market evidence packs——✓
Large source libraries——✓
Multi-market sampling controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build priceConfirmed after discovery based on integrations, workflow complexity, study volume, approval 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 research questions and the markets they cover→Question inventory mapping and source captureWeek 1
02Representative desk sources, panel waves and vendor reports→Source binding, chain logic and the traceability baselineWeek 2
03Your sourcing standard and the labels it requires→Question mapping, source binding and the automation boundaryWeek 1
04Access to relevant APIs, feeds or exports→Desk, panel and vendor-source assessment, then integration setupWeek 2
05Studies you would not want sourced→Sampling cases and failure-mode testingWeek 4
06What no research finding may settle→Traceability scoring, review routing, guardrails and acceptance controlsWeek 3
07A named researcher to accept the finding→Acceptance workflow, then pilot and production validationWeeks 5–6
Nothing else is requiredDeployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Width here is time genuinely worked and not a drawing choice, so the fifth band has to hold two.
PhaseW1W2W3W4W5W6
DiscoveryW1
BuildW2 – W3
EvaluateW4 – W5
Pilot & LaunchW5 – W6
Week focusW1Research workflow discovery, question mapping and the automation boundaryW2Source integration and the traceability baselineW3Finding assembly, chain logic and acceptance controlsW4Evaluation suite, sourcing cases and failure-mode testingW5Repository integration, pilot questions and targeted correctionsW6One research year run under the insights lead, then Agent Care handover
Reading the bandEach bar covers only the weeks its own work is named for. The fifth spans a pair because the work does.
At the end of W6When the source record validates, Agent Care picks the agent up.
DurationSix-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Marketing AI agent
Build a research agent around the figure your last strategy paper could not trace to a publisher.
Show us one research question and the number your last strategy paper leaned on. If a figure has to survive a challenge years after the study closed, then its source, its date and its method have to travel with it. We vouch for no source, only for what it says.
