MN warehouse access topology — 2026-08-12

Primary testimony + routing record. The full analysis lives in the Kairos repo at CONTEXT/2026-08-12-warehouse-access-topology.md — read that for the mechanisms, the open branch and the three named tests. This note holds what the vault must not lose: Lee's own account, and who actually holds which answer.

Lee's testimony (primary source, 2026-08-12)

Asked whether MN people using the chat interface could have found the revenue-table duplication documented in the Kairos blindspot map:

using the chat interface is the default way for most (might be all) of the MN team. some people might have the SDK connection (surely, Simonas does), but having in mind it took us roughly one full day to wrestle it out from Simonas from ourselves, even with explicit "go" approval from Ro... I doubt most people would ever go for it.

Three facts in that: the chat interface is MN's default data surface, not a supplement; direct SDK access exists only for a few; and obtaining a direct connection cost Kairos a full working day with the sponsor's explicit prior approval.

Why it matters

The duplication was structurally undiscoverable from inside. The chat interface returns aggregates and the defect is at row grain; the contamination has been stable at 24–31% of gross since 2023, so every trend, ratio and growth rate reads correct and only the absolute level is wrong; and with one default surface the whole company draws agreeing answers from one configuration. The only person positioned to catch an error in the curated layer is the person who built it — there is no second reader. Access friction functions as an epistemic policy.

Standing consequence for all Kairos work: every figure an MN person quotes us arrived through one of two pipes and the number alone doesn't say which. Any said-vs-did cross-reference must record where the speaker got their number, not only what they said.

Routing — who holds which answer (recorded 2026-08-12 after getting it wrong)

Šaras — Kairos's own engagement lead (PROJECT.md: operated by Lee + Šaras, sponsored by Robertas). Lee's partner, on our side of the table. He is the conduit into MN and the right owner for permission and navigation asks. He is not an MN data authority, and questions to him need no diplomatic packaging.
Simonas — MN's analyst; builds the curated claude.* layer. He holds the answer to what the chat interface reads and to the dedup question. Canon writes this chain as "Šaras → Simonas/analyst": ask through Šaras, but the question is for Simonas.
Robertas — sponsor, MN founder; the approval path, not the data path.

The general rule this cost is in memory: dont-extrapolate-org-facts — whose credentials an integration runs under and whose name sends the messages are plumbing, not authority.

Open

Which dataset the chat interface actually reads is unresolved. If claude.*, the team gets the correct figure while decks and Looker carry the inflated one — the inversion where leadership reads the worst data. Tests and owners: Kairos CONTEXT/2026-08-12-warehouse-access-topology.md §E. Both warehouse-side tests were blocked 2026-08-12 by expired bq auth; Lee's call that day was that Šaras connects, not him.