---
name: undominated-activation-recomputation-tax-audit
description: "Audits selective vs full gradient checkpointing memory savings vs compute FLOP taxes for long-context prefill. Recomputed fields must match the other inputs. A pass is not a compliance, safety, or price verdict."
license: MIT
metadata:
  author: Undominated.ai
  version: "1.1.0"
---

# Activation Recomputation Tax Audit

## Description
Audits selective vs full gradient checkpointing memory savings vs compute FLOP taxes for long-context prefill.

## Category
`architecture`

## Tags
`checkpointing`, `recomputation`, `vram`, `flops`, `memory`

## When to Use
Audit memory-constrained prefill configurations balancing VRAM headroom against generation latency.

## Operational Limitations
Calculates theoretical FLOP addition based on attention and MLP re-forwarding.

## Verification Script
Run the deterministic verification checker:
```bash
python3 scripts/check.py examples/synthetic.json
```
Exit codes:
- `0`: Check passed and thresholds satisfied.
- `1`: Threshold breach or operational violation detected.
- `2`: Schema error or invalid usage.

## What the checker recomputes

- `flop_tax` = 100 * recompute_flops_tflops / forward_flops_tflops, then the original `>` check against `max_allowed_flop_tax_percent`. `recompute_flops_tflops` cannot exceed `forward_flops_tflops`.

- Bound: `activation_memory_saved_gb` <= `total_vram_gb`.

Every other declared field is type-checked. A negative value is rejected except for deltas, drifts, and gains. A type check does not, by itself, prove the measurement.

`population` and `observedAt` are required. The bundled examples are synthetic fixtures, not market measurements.

```bash
python3 scripts/check.py examples/synthetic.json
python3 scripts/check.py examples/contradiction.json
python3 scripts/check.py examples/broken.json
```

Exit 0 is a pass, exit 1 is a review, and exit 2 is invalid input. A pass does not certify a provider, a regulation, a release, or a price.