Methodology

What a replay is

A replay takes a workload of usage events, already normalized into one canonical stream, and simulates it against a plan's documented mechanics: rolling and calendar windows, token and request limits, credit pools, model rules, promotions, overage behaviour and hard stops.

It is a simulation of rules, not a bill and not a prediction of what a provider would charge you. Providers change rules, apply unpublished limits and make mistakes. A replay tells you what the documented mechanics would have done with your workload, and how confident it is in that answer.

Attempted versus accepted demand

Every event counts as attempted demand, including events a limit rejects. Only events a plan actually serves advance accepted consumption. This is what makes a hard stop visible: a workload can attempt far more than a plan accepts, and a replay reports both numbers rather than silently discarding the rejected work.

Accounting is declared, never guessed

Each source declares how its token numbers relate to each other: whether cache reads are a subset of input tokens or additive, whether reasoning tokens are part of output, and which categories it cannot know. The replay engine keeps token buckets disjoint and treats an undeclared or contradictory relationship as unknown rather than assuming one.

When a source does not record a category, the replay says unknown. It does not estimate a value and present it as measured.

Windows and reset behaviour

Rolling windows are anchored exactly as the plan documents them, and calendar windows use the calendar period the plan states. Latching limits stop consumption until the window resets; rejecting limits drop the event; overage limits consume beyond the allowance at the documented rate; record-only limits are reported without enforcement.

Coverage and confidence

Coverage is reported on three separate dimensions: requests, usage and models. They are never blended into one number, because a replay can know perfectly well how many requests happened while knowing very little about the tokens inside them.

Confidence is a level with named factors, not a single score. A partial result says what it could not determine.

Money

Money is carried as decimal strings and computed with exact decimal arithmetic. A replay never presents a floating-point dollar figure, and every derived amount records the basis it was computed from (fixed plan price, plan price plus overage, or API list price equivalent).

Direct API targets

A replay can also target a provider's published API list prices instead of a subscription plan. Nothing is admitted, rejected or deferred: every recorded event is served, and its token categories are priced from the list-price records in force at the rules date, with each event's own timestamp selecting any conditional tier or schedule inside the record it was priced from.

Availability still comes from the catalog: a model is priced only when the catalog records the selected provider as offering it. A model with no list-price record, a record that is not in force at the rules date, or a token category the record does not cover is reported as a gap rather than filled in. No cost is shown unless the whole workload could be priced, because a partial sum would be read as what the workload would have cost. Discounts, provisioned capacity, taxes, minimums and negotiated rates are not modelled, and a real invoice can differ.

Versions and reproducibility

Every result records the engine version, the result schema version, the catalog version, the methodology version, the date the target rules were taken as of, and the plan version used. A replay is reproducible from those, and a share link carries them so a shared result can be audited.

What a replay does not do

StackReplay reads the workload files you select and processes them locally in a browser worker. Raw workload files are not uploaded. Saving the normalized workload in this browser is optional; temporary imports can be replayed without saving.

It does not claim to know unpublished provider behaviour, and it does not turn an unknown into a number.

It does not compare plans by blending unrelated dimensions into a score. If two plans differ in ways a single number cannot express, the replay reports both.