Clarifying Questions

Questions to align scope and requirements before proceeding with architecture

Response Needed: Dec 20, 2025

Total Questions

From Discussion

High Priority

Needs Clarification

Open

Answered

ID	Source	Question	Category	Priority	Status
Q1	Steerco	Which data sources feed into QC Excel files?	Technical	High	Needs Clarification
Our Understanding: Excel exports from Unify platform containing market share, volume, and distribution data. Need to confirm exact file formats and frequency.
Q2	Steerco	What variance thresholds trigger anomalies?	Scope	High	Needs Clarification
Our Understanding: Per Steerco: ±2% for market share is mentioned. Need to confirm if this applies to all metrics or varies by category/client.
Q3	Steerco	How do we access historical Unify snapshots to compare data hierarchies?	Integration	High	Open
Our Understanding: Unify is Circana's internal data platform. We need point-in-time snapshots to detect store dropouts, UPC changes vs. actual market shifts. Access method TBD.
Q4	LOA	Which clients are in scope for POC?	Scope	High	Answered
Our Understanding: Per LOA and Steerco: PepsiCo is the primary pilot client. USA market only. Possibly Nestle as secondary.
Q5	Steerco	What external context sources are approved for anomaly explanation?	Technical	Medium	Needs Clarification
Our Understanding: Steerco lists: competitor promos, local events, brand sentiment, market trends, supply chain. Need approved source list (news APIs, social, etc.).
Q6	Steerco	How should tickets be drafted and where do they go?	Integration	Medium	Open
Our Understanding: Steerco mentions 'tickets' for CIT team. LOA mentions 'draft tickets'. Need clarification: ServiceNow? Jira? Email? What format?
Q7	LOA	Who are the pilot users for testing?	Timeline	Medium	Open
Our Understanding: CST members from PepsiCo team. Need specific names/contacts for user acceptance testing in Weeks 8-9.
Q8	LOA	Is "Visual QC" (output data visual checks) in scope or deferred?	Scope	High	Needs Clarification
Our Understanding: LOA includes 'Output data visual QC' but this wasn't in original Steerco. Need clarification if this is MVP or Phase 2.
Q9	Steerco	What is the definition of 'plain English QC' mentioned in Steerco?	Scope	Medium	Needs Clarification
Our Understanding: Steerco mentions generating 'plain English' explanations. This is NOT explicitly in LOA use cases. Confirm if in scope.
Q10	Steerco	Are there existing QC rules/heuristics we should incorporate?	Technical	Medium	Open
Our Understanding: CST likely has tribal knowledge about what to look for. Should we encode existing rules or let AI learn from examples?
Q11	LOA	What is the expected volume of QC files per week?	Scope	Low	Open
Our Understanding: Need to size infrastructure. How many Excel files? How many cells/rows per file? Weekly or daily cadence?
Q12	LOA	What security/compliance requirements apply to data handling?	Technical	High	Open
Our Understanding: Client data (PepsiCo) is sensitive. Need to confirm: data residency, encryption requirements, access controls, audit logging.
D1	Discussion	What exact scenarios require Computer Vision vs. structured data parsing?	Technical	High	Open
Our Understanding: Ash mentioned computer vision multiple times for handling documents where data 'moves around' and isn't always in the same cell. Need to clarify: is this for layout detection or actual visual anomaly detection?
D2	Discussion	Should MVP support PDF and PowerPoint, or Excel only?	Scope	High	Needs Clarification
Our Understanding: Discussion mentioned 'regardless of format - PDF, PowerPoint, Excel, web screen.' However, PDF/PowerPoint adds significant complexity. Recommend Excel-only for MVP, add formats in Phase 2.
D3	Discussion	What is the target scale/throughput for MVP?	Technical	High	Open
Our Understanding: Ash mentioned 2.2 billion analyses/year in US alone (~6M/day, ~70/second). Even 'curtailed' is hundreds of millions. Need to define MVP scale target vs. production scale.
D4	Discussion	Is Presentation/Qualitative QC in scope (blank pages, wrapped headlines, formatting)?	Scope	Medium	Open
Our Understanding: Discussion mentioned checking for blank pages, wrapped headlines, 'widow words', error messages in generated PDFs. This is qualitative QC vs. quantitative data QC. Clarify if MVP or Phase 2.
D5	Discussion	How pluggable does the architecture need to be?	Integration	High	Open
Our Understanding: Ash emphasized: 'Whatever is architected has to have an easy way to hook into existing apps.' Need to understand integration points: schedulers, FTP folders, existing pipelines. Build standalone or middleware?
D6	Discussion	Should we support FTP folder watching or manual upload only?	Integration	Medium	Open
Our Understanding: Discussion mentioned 'client team has an FTP folder where all this stuff has landed.' For MVP, manual upload is simpler. FTP integration adds complexity but enables automation.
D7	Discussion	Is a conversational interface part of MVP?	Scope	Medium	Open
Our Understanding: Chris suggested 'throw a conversational interface on top so analyst can say I want to understand X.' This is a powerful feature but adds scope. Clarify if MVP or future enhancement.
D8	Discussion	How should the system learn from user corrections?	Technical	Medium	Open
Our Understanding: Discussion mentioned systems that 'learn and improve' from user feedback. Options: full model retraining, prompt tuning, rule updates, feedback loop. Need to define learning mechanism.
D9	Discussion	How do we capture client-specific context that machines can't know?	Technical	Medium	Open
Our Understanding: Ash gave example: client intentionally moved products between categories, causing expected variance. CST teams have this context. Need mechanism to capture 'expected changes' to suppress false positives.

Source Documents

Steerco QC Bot Steerco Presentation (September 2024)
LOA Letter of Agreement - Bounteous x Client (Dec 2025 - Feb 2026)
Discussion Client Discovery Call (Nov 14, 2025) - Computer Vision & Scale Requirements