100eyes.watch — ARGUS log

Reading mail without touching it

Thu, 25 Jun 2026 00:00:00 GMT

The email scanner must never alter mailbox state. A scanner that marks mail as read — or worse, moves it — is indistinguishable from the attacks it is supposed to catch. So every fetch is read-only: BODY.PEEK throughout, on a read-only SELECT.

The fetch itself is two-phase. Phase one pulls the headers and the BODYSTRUCTURE — the full MIME tree — with zero body bytes transferred. Phase two fetches only the text parts the feature extractor actually needs.

Hard caps bound the worst case: any single MIME part over 5MB is skipped and recorded rather than fetched, and a running total stops the fetch entirely past 25MB.

The privacy boundary holds here too: bodies are parsed for structure and indicators. Raw content never leaves the machine.

XGBoost sets the bar at 97.78 — on validation

Tue, 16 Jun 2026 00:00:00 GMT

Three approaches are being compared for ARGUS’s Gate-2 classifier: symbolic rules against named features, XGBoost on the 2381 EMBER 2018 v2 features, and a 1D CNN reading raw file bytes.

First result: XGBoost, trained on 509,932 samples and evaluated on 89,988 samples it had never seen, scores 97.78% accuracy and 0.98 F1 — on the validation set. Both classes hit 0.98 precision and recall; no significant bias toward benign or malware.

These are validation numbers, not final results. The held-out test set was isolated before any model was trained and stays locked until all three approaches are complete — it gets touched once, at the end. Anything else would let the test set leak into model choices, and the final numbers would be quietly dishonest.

CNN and fusion are next.

The daemon stands up

Tue, 09 Jun 2026 00:00:00 GMT

ArgusDaemon is the always-on core: one processor thread receives events from the file watcher and the email scanner and routes them through the gate-keeper pipeline.

The universal contract, enforced from day one: the incident record is written to the hash-chained, append-only SQLite log first, synchronously, before anything else acts. If the log write fails, nothing downstream happens.

The second invariant is the staging zone. New downloads land under a deny-execute ACL and nothing in that zone executes without a verdict — quarantine-first, guilty until proven innocent.

Five smoke tests pass: the daemon starts, watches, routes, logs, and shuts down cleanly.