# Cost Model
This document describes the analytical cost and memory model used by the crate reports.
“The experiment is intended to demonstrate behavioral differences rather than establish optimal performance.”
## Scope
The cost model in this crate is architectural. It is not a measured GPU benchmark. Its job is to make the integration burden explicit enough for diligence:
- which buffers exist
- which stages are local
- what the approximate read/write pressure looks like
- what changes between minimum, host-realistic, and research/debug modes
## Modes
### Minimal
Smallest decision-facing path:
- residual-like local discrepancy
- trust / intervention proxy
- alpha modulation
This corresponds to the lowest-burden attachment path and is used in the timing report as the minimum cost reference.
### Host-Realistic
Current minimum serious path:
- residual
- depth disagreement
- normal disagreement
- neighborhood inconsistency
- thin proxy
- history instability
- grammar/state contribution
- trust, intervention, alpha
Important current decision:
- motion disagreement is not part of the minimum path anymore
- it remains available as an optional motion-augmented extension
That change is deliberate. The current suite does not justify treating motion disagreement as mandatory in the minimum path.
### Full Research / Debug
Comparison-only path:
- synthetic visibility hint
- optional motion disagreement
- thin proxy exports
- history instability exports
- structural-state exports
This mode exists for ablation, trust diagnostics, and report generation. It is not a deployment claim.
## Current Trust and Cost Interaction
The current trust behavior is near-binary / gate-like in this crate. That matters for cost because it makes two reduction ideas more credible:
- half-resolution trust or intervention
- per-tile trust aggregation followed by alpha upsampling
The crate does not claim these are already tuned on hardware. It only shows that the dataflow is compatible with them.
## Core Statements
“The DSFB supervisory layer can be implemented with local operations and limited temporal memory, with expected cost scaling linearly with pixel count and amenable to reduced-resolution evaluation.”
“The framework is compatible with tiled and asynchronous GPU execution.”
These are architecture statements, not measured deployment claims.
## What The Cost Model Helps Decide
- whether the supervision is local enough to be a realistic GPU pass candidate
- whether the minimum path is materially smaller than the debug path
- what the memory scaling looks like at larger resolutions
- which buffers are plausibly droppable outside analysis mode
## What The Cost Model Does Not Prove
- real GPU milliseconds
- cache behavior on NVIDIA, AMD, or Intel hardware
- production pass scheduling quality
- shipping-engine memory-system efficiency