Inspiration

A customer requests a new delivery address and time. A dispatcher acknowledges the message, but the driver is still following the previous instruction. “Received” is not the same as “I accept this new version.” FreightSwitch explores this precise coordination problem: how can a reviewer see the last complete agreement, the proposed change, and the evidence still missing without a language model silently deciding that the shipment is authorized?

The current demonstration uses one synthetic shipment and three fictional parties. It is a workflow prototype, not evidence of customer adoption or a field-tested logistics product.

What it does

FreightSwitch has two deliberately separate workspaces:

A synthetic instruction-history workbench. Follow a changed address/time, receipt-only replies, stale-version confirmation, withdrawal, and later explicit reconfirmation. Inspect the source messages behind each state. The workbench preserves past confirmation evidence even after a later rejection and keeps current validity separate from history. A newer agreement in the authored scenario requires all three parties' latest confirmations on the same instruction.
A source-bound interpretation review. Inspect an unverified suggestion beside the complete original message, literal version, revision and SHA-256. Record a separate local human interpretation with a reason. Acceptance interpretations require explicit review attestations; direct approval of a model acceptance is blocked. Changing the source or sender label makes its old interpretation stale. The audit preserves both the original suggestion and the human decision; the suggestion remains UNVERIFIED even after that interpretation.

Important limitation: these human interpretations do not update the shipment agreement. They are not authenticated customer, dispatcher or driver confirmations. The app sends no messages, connects to no carrier and dispatches nothing. The three-party scenario and human-review panel are not an end-to-end operational confirmation system.

How it is built

The provider adapter implements calls to NVIDIA Nemotron 3 Nano through Nebius Token Factory's HTTPS chat-completions endpoint; the historical evaluations used actual provider calls. Requests ask for a semantic label—propose, accept, received, reject, withdraw or unclear—and a source-specific question. Application code, not the model, binds the full source text, its hash and span, and the sole literal version token. That separation keeps evidence integrity distinct from the fallible interpretation of meaning. The local reference UI is offline; its authored practice and legacy historical replay do not make new inference calls.

The local app uses Python's standard library and a JavaScript interface. Its review store is local and append-only. Exports carry build and source identifiers so a reviewer can identify the context inspected. Hashes establish consistency with source bytes, not real-world identity or truth.

A separate browser-local preview presents captured authored scenario outputs and a separately tested human-review adapter. It does not execute Python/SQLite or live inference. Its audit is stored in this browser's localStorage, not authenticated, tamper-proof or cross-device. It is a bounded demonstration of the review interaction, not an arbitrary Python evaluation service.

What testing actually showed

In a separate frozen October 29 run, 12 real Nebius calls produced valid model JSON. The independent evaluation found 7/12 complete cases and 30/36 source labels correct at the raw semantic level—but the strict integrated pipeline rejected all 12 cases, returning no usable results. Its narrow language gate was a bottleneck; the model also made six label errors. This is a failed end-to-end result, not a successful autonomous workflow.

The published v1.6 source is a post-evaluation repair: preserve mechanically bound suggestions as visibly UNVERIFIED, quarantine invalid rows, and let a person record a separate interpretation without granting operational authority. Replaying the known responses through the preceding repair retained 36 unverified suggestions and automatically approved none. That is diagnostic visibility, not 36 correct predictions, a revised score, or a new blind test. No new inference has evaluated this human-review release. Its public source includes reported aggregate results, not the underlying private fresh-set inputs or audit.

The public release's 201 passing development tests and six authored scenarios exercise its logic and workflows; they do not establish model accuracy, production safety or real-world impact. The separate preview passed 86 adapter checks, including 43 differential cases, and 23 real browser checks on the official internal preview. Browser tests exercised recording a local interpretation, preserving UNVERIFIED labels, reload persistence, export, stale-source blocking and the authored scenes. The 320/390/1280-pixel layout checks were CSS iframe tests, not mobile-device emulation or production-hosted access tests.

Challenges and lessons

The hardest lesson was that exact provenance and semantic correctness are different problems. Binding the right bytes does not prove that “accepted” is the right interpretation. A conservative gate can also become so narrow that it discards every useful response. We kept the original failed result and changed the product boundary instead of relabeling the outcome: show uncertainty, preserve the original suggestion, and make human judgment inspectable without calling it verified agreement.

Accomplishments

The working local prototype connects version history, source inspection, withdrawal/reconfirmation scenarios and an auditable interpretation workflow. It preserves the original unfavorable model scores rather than replacing them with development-test counts. Its distinctive contribution is the separation of model suggestion, local human interpretation, historical evidence and operational authority—not a claim that AI has solved shipment dispatch.

What's next

The next evidence needed is actual dispatcher research, authenticated counterparty confirmation, a reviewed integration between interpretation and instruction state, and a newly frozen independent evaluation on unseen cases. Any new model test must report semantic quality, retained suggestions, quarantine and delivered outcomes separately. Until those steps are completed, FreightSwitch remains a synthetic review prototype rather than a production dispatch system.