AI routing and the linguistic air-lock
fix(ai): add dated Correction callouts to 7 of 10 ai/ articles — decode-time constraints, SLM stack deps, sovereign-ai-routing PII claim (compliance-relevant), draft-generate endpoint underclaim, plus 3 minor factual slips; 3 articles verified clean
@@ -16,6 +16,8 @@ paired_with: sovereign-ai-routing.es.md --- **Correction (2026-08-02) — compliance-relevant, escalated:** the sanitize-outbound/rehydrate-inbound mechanism described in this article does not exist. The only real sanitization code, `service-slm/crates/slm-doorman/src/redact.rs`, is a plain regex-based **secret/credential redactor** (PEM keys, AWS/GitHub/Slack tokens) — its own doc comment states it is "the only redaction surface in the apprenticeship pipeline," called exclusively when writing training-corpus tuples (`apprenticeship.rs`, `verdict.rs`), never on the Tier C external-routing path. A corpus-wide search for "PII," "pseudonym," "location identif," or any rehydration-table code returns zero hits anywhere in `service-slm`. Separately, the real Tier C client (`tier/external.rs`) documents "no live API calls in v0.1.x" and gates every call behind a compile-time allowlist — contradicting this article's unhedged claim of live routing to external providers today. This is the same class of finding as the `security/data-sovereignty-telemetry.md` correction (an active gap between a public compliance-adjacent claim and real code, not just a stale technical detail) — escalated to Command by mailbox given this article explicitly targets regulated-industry readers (real estate, financial advisory, clinics, law firms) evaluating the platform's actual privacy protections. **Flagged, not resolved.** > AI routing processes language model requests through a local sanitization step before any data reaches external models, keeping internal structured data off third-party servers. **AI routing** in the [[pointsav-overview|PointSav]] platform is the mechanism by which [[service-slm|`service-slm`]] — the [[doorman-protocol|Doorman]] — mediates every request that involves a language model, whether the model runs locally on the customer's hardware, on a burst compute provider, or on an external API. The routing design treats the boundary between customer-controlled infrastructure and external compute as a one-directional filter: before any text leaves the customer's private network, the payload passes through a local Small Language Model that sanitizes sensitive information, strips location identifiers, and masks Personally Identifiable Information (PII). Only the sanitized prompt — containing the mathematical structure of the request but not the customer's private data — routes to external compute. When the external model returns a result, the router re-hydrates the response with the correct internal context before delivering it to the operator. The external model never holds the actual structured records from the customer's ledger.