Cost grounding

fOS never makes up a number. Every figure it states about money — what a chain costs, what a lane pays, what to bid — comes from an authoritative source: the cost engine for cost, APEX for market, and the bid recommendation for the price to put forward. The language model orchestrates; it does not calculate. This page explains why that rule exists and how it is enforced.

The single source of cost truth

The cost engine is the single source of cost truth. It computes real, per-country, per-leg costs from a carrier's own rate tables and the vehicle's specs. When fOS needs a chain's economics, it reads them from the engine via a decision-grade tool — it does not approximate them itself.

This is not a soft preference. A tool that would independently estimate a cost the engine already computes is redundant at best and fabricated at worst, so such tools are cut from fOS entirely. The cautionary example is a former "toll estimate" tool that guessed distances and mislabeled countries; it was removed because the honest answer to "what's the toll?" is to read the engine's real per-leg breakdown, not to guess.

Reference tables are different from cost-of-record

There is a real distinction fOS preserves:

• Cost-of-record — the authoritative per-leg P&L for this chain and this carrier. Only the cost engine produces it, and fOS only ever reads it.
• Reference tables — real benchmark rows a dispatcher consults, such as diesel prices or driver rates by country. Displaying these real rows is legitimate; they inform judgment but are not presented as the cost of a specific chain.

A tool that displays real benchmark data is a valid reference surface. A tool that derives a chain's cost itself is not. fOS keeps the former and refuses the latter.

Market numbers come from APEX

The same rule applies to market prices. When fOS quotes what a lane pays, that comes from APEX's percentiles, not from the model's intuition. And when fOS suggests a price to bid, the authoritative figure is the cost-floor-aware bid recommendation — not a quick market-only number. fOS may show market context, but it acts on the floor-aware recommendation.

Why this matters for trust

Grounding is what makes fOS safe to act on. A dispatcher can accept an fOS recommendation because every number behind it is traceable to the cost engine or APEX and would be identical if a human pulled it up. The model's job is to run the mission loop — decide what to do next and invoke the right tool — not to be the calculator. When a required input is missing, fOS surfaces lower confidence rather than inventing a value.