Counterfactual Routing (I25)

The upper beam is what the sessions actually did: cache bleed and context pressure, drawn in amber and rose. The lower beam is the fast-tier route the operator-pinned price table would have simulated, drawn in emerald and cyan. The shaded band between them is pressure that the simulation says was recoverable.

Drag the context-pressure threshold. Sessions at or above that cache-bleed inflection stay in the gate, and the recoverable context and cost totals recompute from the loaded rows. If the tier table is unpinned, the lower beam is withheld and the cost stays unknown. That refusal is not a zero. The picture is not an effect and not a routing change.

observed cache-read ratio = Σ cache_read / Σ (input_uncached + cache_write + cache_read)
cache_bleed_inflection = QUANTILE_CONT(cache-read fraction, 0.9) after the ASOF alignment
simulated_cost = (input_uncached + cache_write + cache_read) × input_rate / 1e6 + output_tokens × output_rate / 1e6
simulated_context_pressure = peak input side / pinned context limit
Refuse when the profile pin, cohort pin, or rates are missing, or the usage harness does not match the pin.
Unpinned tiers keep observed I02 and do not invent a cost.
The simulation is not an effect and not a routing change.

Scrubber, local to these rows:
  in gate when cache_bleed_inflection ≥ threshold
  observed pressure = observed_context_pressure, else context_pressure, else cache_bleed_inflection
  context delta = max(0, observed pressure − simulated_context_pressure) when both are numbers
  recoverable context = Σ context delta inside the gate
  recoverable cost = Σ simulated_cost inside the gate
  null simulated fields stay unknown and are not treated as zero

Waiting for rows.