SilentShift

An order arrives as 20 CASE. Its approved contract says 240 EA. The transport succeeds and the JSON is valid, but there is no approved conversion explaining the change. SilentShift holds the order before delivery and gives the integration operator the original message, reference records and reason in one investigation.

The same incoming value can be legitimate. An effective approved amendment can authorize 20 CASE; SilentShift checks that authority at event time rather than guessing a conversion. The payment example makes the same distinction for interface contracts: 12550 minor EUR units and 125.50 major units may be numerically equivalent, but that does not authorize a representation change.

Investigate inside the native production

SilentShift runs its decision inside an InterSystems IRIS interoperability production:

Partner Business Service → Semantic Business Process
  → reference, precedent and native AI Hub Operations
  → Embedded Python policy → destination or quarantine Operation

The native AI Hub agent reads four case-scoped, read-only tools: the current message, approved contracts, related records and similar precedents. It proposes a recommendation with evidence IDs and uncertainty. Deterministic Embedded Python policy authorizes the final route. A previous delivery, a helpful runbook or an agent's explanation cannot override a critical contradiction in current approved evidence.

Give the operator a clear next action

The React evidence workbench keeps the original payload, immutable reference snapshot, actual tool reads, policy checks and delivery receipt together. Similar precedents expose real source IDs, outcomes and cosine scores from native IRIS Vector Search. They supply context without borrowing authority from an earlier decision.

Held cases become native IRIS Workflow tasks. Authorized reviewers can approve eligible cases or reject them with a reason, current version and idempotency key. Critical contradictions cannot be approved. If native completion fails after a decision is saved, Resume recorded review carries the original request forward. Each verified concurrent-approval and recovery journey produced one simulator delivery.

Partners can submit raw message JSON through the authenticated API or a custom native ready-file adapter. The adapter retains exact input bytes, records the native case and trace, and writes an acknowledgement from the persisted outcome. Replays reuse the original case and receipt. Each actual local provider HTTP exchange also appears as a typed request/response pair in its case's native Visual Trace, including a verified failure path.

Run and inspect the demonstration

The verified runtime is official ARM64 Community AI Hub 2026.3.0AI.154.0 (154U), with local Ollama qwen3-silentshift:4b-instruct and local all-minilm:22m embeddings in 384-dimensional IRIS vectors. The backend also passed a clean local IPM installation without a source mount. After explicit setup, that installed production returned 5/5 expected final outcomes and 4/5 expected agent proposals, with all four native tools read.

All supplied orders, payments and references are synthetic. Delivery means a receipt in the included IRIS simulator. The five frozen scenarios are a coverage demonstration: the deterministic baseline also solves all five, so they establish no AI accuracy uplift or general production-readiness claim. A public hosted application and public IPM registry installation are not provided.

Start with the operator walkthrough: hold the unsupported unit change, allow the effective amendment, then inspect a review task and native Visual Trace. The README provides exact runtime, model, installation and verification commands. The licensed EAP runtime and model weights are obtained separately.

Source: github.com/himanshu748/silentshift

Demo video: https://youtu.be/a4dOcqu89EA