# Agentic-waste — reproduce the headline figures (public, no repo access)

These are the precomputed result JSONs behind the claim
`66.85% blast radius / 51.76% VRAM locked / ~99.5% output<input waste`.
They are derived from already-public datasets (Azure LLM 2024 traces, Alibaba
GPU 2026 traces). Full registry: https://lib.trustfortress.ai/claims/agentic-waste

## 1. Download + verify integrity
```bash
curl -sL -o tp4.json    'https://lib.trustfortress.ai/objects/cubie-tf%2Fagentic-waste-evidence%2F2026-07-04%2Fintc-v1.0-tp4_nvlink_results.json'
curl -sL -o hunter.json 'https://lib.trustfortress.ai/objects/cubie-tf%2Fagentic-waste-evidence%2F2026-07-04%2Fintc-v1.0-hunter_results.json'
curl -sL -o az2024.json 'https://lib.trustfortress.ai/objects/cubie-tf%2Fagentic-waste-evidence%2F2026-07-04%2Fintc-v1.0-azure_2024_unconstrained.json'
curl -sL -o alibaba.json 'https://lib.trustfortress.ai/objects/cubie-tf%2Fagentic-waste-evidence%2F2026-07-04%2Fintc-v1.0-alibaba_v2026_squatter_metrics.json'
sha256sum tp4.json hunter.json az2024.json alibaba.json
# EXPECT:
#   7d1a90bfc6036e7d6eb8483ba1ddb4bf69cc69003a306016368d221286119bc9  tp4.json
#   6fb720db2600441e6a7155d4a2ac6ee8cdc023de65fceef016edc7742d979685  hunter.json
#   4a2a451e4e104bc543cce67aae42f8446513ea7160c7c32a3a6a880c2bfada06  az2024.json
#   2e4adf2dcd79233d603e1c68498434a469413e292b8843bca7bc5131844d43ad  alibaba.json
```

## 2. Re-derive the headline numbers from the JSONs
```bash
# 66.85% fleet blast radius + 51.76% VRAM locked:
python3 -c "import json;d=json.load(open('tp4.json'));b=d['full_tp4_blast_radius'];print('blast %%:',b['capacity_pct'],'intervals:',b['intervals'],'/',d['total_llm_intervals']);print('VRAM locked %%:',d['vram_locked_by_tp4']['pct_total_vram_locked'])"
# -> blast %: 66.8479  intervals: 105226 / 157411 ; VRAM locked %: 51.76

# ~99.5% request waste (scoped to the Azure-2024 trace):
python3 -c "import json;d=json.load(open('az2024.json'));c=d['code'];v=d['conv'];num=c['total_rows']*c['output_lt_input']['pct']/100+v['total_rows']*v['output_lt_input']['pct']/100;den=c['total_rows']+v['total_rows'];print('output<input %%:',round(100*num/den,2),'of',den,'requests')"
# -> output<input %: 99.47 of 44107694 requests
```

## 3. Regenerate the JSONs from raw telemetry (optional, full pipeline)
The raw datasets + analysis scripts (`tp4_nvlink_breakdown.py`, `hunter.py`)
live in the public datacenter-crash-logs pack:
https://lib.trustfortress.ai/objects/cubie-tf%2Fdatacenter-crash-logs%2F2026-07-02%2FMANIFEST.md

## Scope
`~99.5%` output<input waste is measured on the **Azure-2024 trace only**
(44.1M of the ~45.3M analyzed rows). Only the `66.85%` figure appears in a
whitepaper. See the registry for full scoping and pinned provenance.