Register in only one of two newly-crossed nexus states for FY24
Why — recorded
The second state's threshold was crossed only marginally, and the client's CFO approved the narrower reading in writing.
Institutional memory · compiled
The decision is in the file.
The reason usually isn’t.
Firms don’t lose their work. They lose the reasoning behind it — in a thread nobody searches, a memo nobody reopens, a partner’s head. Cortex reads what your firm already wrote and compiles it into what was decided, who owns it, and where anyone recorded why.
- CitedEvery claim resolves to a source
- PermissionedEnforced in the query, not the answer
- YoursYour infrastructure, your model
01 · Ingest
It reads what you already have.
Documents, spreadsheets and slides from Drive. Threads from Slack. Mail. Meeting notes and work papers. No new system for anyone to keep up to date — the record is whatever your firm already produced.
32 documents in the demo tenant — a fictional tax firm, not a customer. Drive and Gmail run live today; Slack and Microsoft Graph are built but not yet pointed at a real tenant.
02 · Compile
It compiles a record, not a search index.
Every engagement is reduced to decisions, procedures and ownership — each one carrying its owner, its status, and the sources it came from. Where the firm wrote down why, that reasoning is captured. Where nobody did, Cortex says so.
Globex crossed the economic-nexus receipts threshold in two additional states in FY24. Rather than register in both, we applied the existing apportionment approach and registered only where the threshold was clearly exceeded. The client's CFO approved the narrower reading in writing. The second state's threshold was crossed only marginally and is flagged for re-testing next year. Owner: Meera Sundaram.
The reasoning was written down, so it is carried into the record and attributed to the sentence it came from.
Clara: Globex are claiming the R&D credit on the same activities we excluded last year as routine engineering. Either we changed our mind or someone didn't read the prior file — and there's no note explaining which.
No sentence in this source states a reason.
Conflicting R&D credit treatment on Globex's engineering activities
Why — not recorded
No source explains the reasoning. Cortex will not invent one.
Nothing in the source explains why. The record says so and stops — a plausible reason here would be a fabrication the firm could act on.
5 of 8 artifacts below carry a recorded rationale. The rest are shown with the gap visible — that is the output, not a failure of it.
Register in only one of two newly-crossed nexus states for FY24
Globex crossed the economic-nexus receipts threshold in two additional states in FY24; rather than registering in both, the firm applied the existing apportionment approach and registered only in the state where the threshold was clearly exceeded.
Why — recorded
The second state's threshold was crossed only marginally, and the client's CFO approved the narrower reading in writing.
Conflicting R&D credit treatment on Globex's engineering activities
Globex is claiming the R&D credit in the current year on the same activities that were excluded last year as routine engineering; Clara Boone flagged that either the position changed or the prior-year file was overlooked, with no note in the file explaining which.
Why — not recorded
No source explains the reasoning. Cortex will not invent one.
R&D credit: routine engineering treatment differs between Globex and Harlow based on documentation
Routine engineering was excluded from the R&D credit for Globex (FY24) but is being claimed for Harlow this year, even though both are described as process improvement. The distinguishing factor is that Harlow has contemporaneous project records supporting the claim (documented in the Harlow study), while Globex did not. The Globex file itself does not record this reasoning, and whether the Globex position should change is still an open discussion item as of the September 8 compliance review.
Why — recorded
"Harlow has contemporaneous project records and Globex did not, which is the distinction" (Nisha Patel); the Globex file itself lacks that explanation, per Clara Boone.
- Clara Boone to pull the FY24 Globex file before responding to the client on the R&D question
Don't extend Globex's nexus treatment to other clients without re-testing
The FY24 handover said Initech and Umbra follow the Globex approach of treating a second state as not triggering nexus. That treatment was actually a Globex-specific fact from a call with the client's CFO, not a firm-wide position, so it should be re-tested before being applied to other clients such as Umbra.
Why — recorded
The no-nexus conclusion came out of a call with Globex's CFO — a client-specific fact rather than the firm's house position (Ravi Iyer, Clara Boone).
Marketplace facilitator nexus/registration analysis
When determining where a multi-channel seller must register for sales tax, exclude marketplace-facilitated sales from the threshold analysis and evaluate only direct sales volume by state, since marketplaces already handle collection and remittance in nexus states.
Why — recorded
Marketplaces collect and remit in every state where Meridian has nexus, making it unnecessary to include marketplace volume in the registration threshold analysis.
- Confirm which marketplaces collect and remit tax in every state where the company has nexus
- Exclude marketplace-facilitated sales from the state-by-state threshold analysis
- Analyze direct sales volume against each state's economic nexus threshold
- Register in states where direct sales volume exceeds the threshold
- Monitor states where direct sales are approaching the threshold
Engagement letter must be countersigned before work begins
The company's onboarding procedure requires a countersigned engagement letter from the client before any fieldwork or preparation work starts.
Why — not recorded
No source explains the reasoning. Cortex will not invent one.
- Send engagement letter to client for countersignature
- Confirm countersigned letter is received
- Only then begin fieldwork or preparation work
September extension due-date list — three returns
Anup Rege owns the three extended returns due 15 September: Globex Ltd, Initech and Umbra Systems.
Why — not recorded
No source explains the reasoning. Cortex will not invent one.
Treat Harlow Devices' process-improvement engineering as qualified research (R&D credit)
For Harlow Devices, process-improvement engineering activities are being treated as qualified research for this year's R&D tax credit study.
Why — recorded
The activities are documented against the four-part test and the client has contemporaneous project records, which is what made the position supportable.
03 · Ask
Then a partner can ask it a real question.
Question
Why did we treat routine engineering differently for Globex and Harlow on the R&D credit?
The distinction was documented for Harlow but never written down for Globex, and by September it's become a live, unresolved question.
The stated rationale (June 2026): When Clara Boone flagged apparent inconsistency — "we excluded routine engineering for Globex last year and we are claiming it for Harlow this year. Both are process improvement" — Nisha Patel gave the reasoning directly: "Harlow has contemporaneous project records and Globex did not, which is the distinction. That is in the Harlow study." This is consistent with Harlow's R&D credit study, which documents the process-improvement engineering against the four-part test and notes the client's contemporaneous project records as "what made the position supportable."
But that reasoning was never recorded for Globex. In the same June 19 exchange, Clara pushed back: "then the Globex file should say that is why it was excluded. It does not say anything." So while Nisha's explanation exists in chat, the sources show no evidence it was ever written into the Globex engagement file itself.
This gap resurfaced as a live problem. By September 8, Clara reported that Globex is now claiming the credit on the same activities that were excluded last year as routine engineering, and framed it starkly: "Either we changed our mind or someone didn't read the prior file — and there's no note explaining which." The firm's weekly compliance review that same day logged this as unresolved — "Needs further discussion: whether the Globex R&D position changes from last year, and who signs off if it does" — and assigned Clara to pull the FY24 Globex file before responding to the client.
Bottom line: the sources record a reason (lack of contemporaneous project records at Globex vs. Harlow having them), but that reason lives only in a June chat message, not in the Globex file, and as of the most recent record (September 8) the firm had not yet confirmed whether that's actually why Globex was excluded or resolved whether the position should now change.
Sources
- 1 #tax-prep — Globex R&D credit question
- 2 R&D credit study — Harlow Devices
- 5 #tax-prep — are we consistent on R&D?
- 6 Meeting notes — weekly compliance review, 8 September
Hover or tap a citation to trace it to its source
Captured verbatim from the running system. Retrieval takes at most two chunks from any one document, so an answer spanning a thread, a study and a set of meeting notes is a property of the design rather than a lucky draw. Every claim above resolves to a source a partner can open.
The evidence chain
One question, four sources, three months
The evidence chain behind the R&D answer — reconstructed from what the firm already wrote.
-
R&D credit study — Harlow Devices
“documented against the four-part test; contemporaneous project records are what made the position supportable”
The reasoning, recorded — for Harlow.
-
#tax-prep — are we consistent on R&D?
“Harlow has contemporaneous project records and Globex did not, which is the distinction.”
The distinction is stated aloud — in chat, not in the file.
-
#tax-prep — same thread
“then the Globex file should say that is why it was excluded. It does not say anything.”
The gap is identified at the time — and not closed.
-
#tax-prep — Globex R&D credit question
“Either we changed our mind or someone didn't read the prior file — and there's no note explaining which.”
Eleven weeks later the same gap resurfaces as a live client question.
-
Weekly compliance review
“Needs further discussion: whether the Globex R&D position changes from last year, and who signs off if it does.”
Logged as unresolved, with an owner assigned to pull the FY24 file.
No one wrote this chain down. It spans Drive and Slack across three months, and every step is a document a partner can open.
04 · Permissions
Nobody sees what they could not already open.
Permissions are enforced in the retrieval query — restricted material is never selected, so it never reaches the model and cannot leak through a paraphrase. Redacting an answer after the fact is not the mechanism, because that approach fails quietly.
Document set — shaded blocks are withheld from the selected viewer.
View as
13 of 13 documents readable
Nothing withheld at this level.
11 of 13 documents readable
- Deal-desk discount matrix group:deal-desk
- Leadership compensation plan group:leadership
A different demo tenant — Northbridge's corpus is public-ACL'd, so it cannot show gating.
The lens reports counts and the principal that gated each document — never the withheld content itself, so the explanation cannot become the leak it describes. A source that reports no permissions at all fails closed: it is indexed, withheld from every answer, and counted as unresolved until the source supplies them.
05 · Scope
What it is, and what it is not.
Worth being blunt about, because the category is crowded with things this is deliberately not.
It is
- Internal expert augmentation
- It works for your people, behind your login. Clients never touch it.
- Institutional memory
- Prior-year positions, elections, and the reasoning behind them — held across staff turnover.
- Citation-grounded
- Every claim resolves to a document someone in the firm can open and check.
- A record of what is missing
- Where nobody wrote down why, it reports the gap instead of filling it.
It is not
- Not a chatbot
- There is no assistant persona and nothing conversational to put in front of a client.
- Not client-facing
- It is never exposed to the people you serve. It is a tool for the desk.
- Not a replacement for judgment
- It surfaces what the firm already concluded. A person still decides.
- Not a filing or advice engine
- It does not compute positions, prepare returns, or give an opinion.
06 · Deployment
It runs inside your boundary.
Your client files are the reason this is hard to buy. So the corpus, the index and the model all sit inside your VPC or on your own hardware, and nothing is sent to a third-party AI API. There is no vendor tenant holding your documents, because there is no vendor tenant.
- UK / EU
- UK GDPR and the ICO's IDTA where data would otherwise leave the jurisdiction. Self-hosting removes the transfer question rather than papering over it.
- US financial & tax
- GLBA and the FTC Safeguards Rule (16 CFR 314), with IRS Pub 4557 and the Pub 5708 WISP for firms that need a written plan.
- India
- DPDP Act consent and purpose limits, with processing kept in-country when a client requires it.
Listed as regimes the deployment posture satisfies, not as legal advice. The posture is the same in each case: if the data never leaves, most of the questionnaire answers itself.
07 · Where this stands
Honest state, then the ask.
Built and running
- Compiler, retrieval, permissions and directory — running, with a test suite
- Google Drive and Gmail connectors, live against a real mailbox and drive
- Permission enforcement in the query, with a lens that reports what was withheld
- Deployment inside your VPC or on your own hardware, self-hosted model
Not yet
- No customers yet — the first pilots are being scoped now
- Slack and Microsoft Graph connectors are built but not yet run against a live tenant
- Outcome numbers: the pilot establishes the baseline rather than quoting someone else's
The pilot
- 6 weeksstart to findings
- One service linenot the whole firm
- ~20 engagementsenough corpus to be real
- Fixed feeno per-seat commitment
You give one service line’s existing files and threads. You get a Documentation Gap Report for those engagements — what was decided, who owns it, and specifically where the reasoning was never written down. That report is useful whether or not you ever deploy the system.
Start a conversationhello@sayuk.cloud · a reply within 24 hours · no mailing list.