Question

---

Sanitization by Construction: The "Edge Compiler"

Regex-based PII stripping on arbitrary exports is fragile: one new column, one merged cell, one localization change—and you leak. We chose structural exclusion: the network never sees a free-form ledger string because we never build one for the model.

---

Naming truth: there is no EdgeCompiler package

Edge Compiler is our term of art for the deterministic reduction in buildPortfolioContext (app/lib/ai/contextBuilder.ts). There is no separate package or binary with that name—just this function’s contract.

---

Two different pipelines (do not conflate them)

Stage	What happens	PII risk
CSV → trades	Importer / broker adapters (packages/importer, etc.) parse files into normalized Trade objects in the browser	Parsing must respect broker columns; that is import logic, not Ask AI
Trades → LLM context	buildPortfolioContext turns Trade[] + positions into a fixed template string	No row dump — only totals + top 10 tickers by value

Part 3 is about the second stage: the AI boundary. The function does not parse CSV text or “drop Description / Account Number columns” line-by-line — by the time it runs, data is already Trade. Anything that is not part of the allowed output lines simply cannot appear, because the function only pushes the template fields (ticker, shares, currency, value, allocation %, P/L %).

That is structural exclusion, not redaction.

---

Why PII “never hits the network tab” (default path)

For the default Ask AI flow:

1. The client builds context from buildPortfolioContext only.
2. The HTTP body contains context, not the raw CSV file and not a concatenation of every trade row.
3. Open DevTools → Network on a typical “ask about my allocation” question: you should see a short context string — not your export.

Paid attachment is explicit: if the user sends file content, that is a deliberate second boundary — not the default “sanitized snapshot” path.

---

The core logic (same as Part 2, different emphasis)

TOP_HOLDINGS_COUNT = 10 caps ticker-level disclosure. Totals (calculatePortfolioTotals) give portfolio-level signal without listing every fill. Comment in source:

// No raw ledger rows, no PII, no account identifiers—sanitization by construction.

For CSV import, optional column mapping may send headers and a few sample rows only; the full file stays in the browser. Same privacy instinct, different pipeline than Ask AI context.

---

Performance

Reduction is O(n) over trades for position derivation + O(m log m) for sorting positions where m is position count — typically milliseconds on-device. Sub-200ms is plausible on modern laptops; measure if you publish a number.

---

Summary

• Edge Compiler = buildPortfolioContext — fixed schema, aggregates + top-N.
• Not a regex scrubber on arbitrary strings.
• CSV column semantics belong to import; semantic summary belongs to context builder.

---

Part 3 of Sovereign Engineering.

Read the full Sovereign Intelligence book or try the app.