# WWT Cubie A100 Demo Workload — 1,500 Request Package (MANIFEST)

- Date: 2026-07-07
- Owner: Nick Venezia / Centillion.AI
- R2 bucket: `trustfortress-resource-library`
- Prefix: `cubie-tf/wwt-demo-workload/2026-07-07/`
- Public object route: `https://lib.trustfortress.ai/objects/<url-encoded-key>`
- Source zip: `wwt_cubie_demo_workload_1500.zip` (2,131,270 bytes, 33 files, ~38 MB extracted)

## Safety / secret scan

Secret-scanned before publication with
`grep -rEi "(api[_-]?key|secret|password|bearer |glpat-|sk-ant|AKIA|-----BEGIN)"`
across all 33 extracted files. **No real credentials found.** All matches are the
intended synthetic content: prompt-injection / PII demo patterns labeled
`approved_synthetic_pattern`, redacted placeholders (`sk-demo-redacted`,
`sk-following`), and one environment-variable reference (`Bearer $LITELLM_PROXY_KEY`)
inside `submit_litellm_batch_example.sh`. No `-----BEGIN` private-key blocks, no
`glpat-`, no `AKIA…`, no `sk-ant-…`, and no real `sk-…` API keys are present. The
package README affirms all prompts are synthetic and sanitized with no real
credentials, PII, crash logs, or customer payloads.

## Category distribution (1,500 total)

| Category | Count |
|---|---:|
| clean_enterprise | 600 |
| token_bloat_long_context | 300 |
| retry_loop_duplicate | 225 |
| unauthorized_tool_scope | 150 |
| prompt_injection_pii_exfiltration | 150 |
| malformed_doomed | 75 |

## Published objects (uploaded to R2)

| Object | R2 key | Bytes | SHA-256 |
|---|---|---:|---|
| `wwt_cubie_demo_workload_1500.zip` | `cubie-tf/wwt-demo-workload/2026-07-07/wwt_cubie_demo_workload_1500.zip` | 2131270 | `14c85d865b9446271d5a938d92fec44ea9ab67342af549093f81c2c83be6e8f3` |
| `README.md` | `cubie-tf/wwt-demo-workload/2026-07-07/README.md` | 1787 | `f3ee9eef0b85944ff23d90effa19fd03974b981953be59edb171356efdae8240` |
| `run_manifest.yaml` | `cubie-tf/wwt-demo-workload/2026-07-07/run_manifest.yaml` | 2298 | `2f60981f12c5f2e6c01b9c1333ea460679a06f1dd74d9ca53532450af4bc9875` |
| `source_register.csv` | `cubie-tf/wwt-demo-workload/2026-07-07/source_register.csv` | 10103 | `2932f0f520542f6883a4c5187a88ec9222fc64677adbca22a408aa19c5fa0393` |
| `nist_mapping.csv` | `cubie-tf/wwt-demo-workload/2026-07-07/nist_mapping.csv` | 467740 | `35fa4c2441a3bac5a8f2882fdb65a1661d197e851d1f26d42f923094737c372b` |
| `bulk_request_validation_report.json` | `cubie-tf/wwt-demo-workload/2026-07-07/bulk_request_validation_report.json` | 2144 | `d34fa6f0e3dfd1ca3e893bc51802ef9792af2f3891eb5dfe0305b7e4a22a443d` |
| `workload_variants.md` | `cubie-tf/wwt-demo-workload/2026-07-07/workload_variants.md` | 464 | `f9e85476138e3755f65f7e857ea8390ec7e258049b72dff634bd554958726bfa` |
| `exclusions.md` | `cubie-tf/wwt-demo-workload/2026-07-07/exclusions.md` | 2039 | `9f673f97f9793155bc96184522a1b74a1ed5381082cc1ef91c2611c8d84c742c` |
| `MANIFEST.md` | `cubie-tf/wwt-demo-workload/2026-07-07/MANIFEST.md` | (this file) | (recorded in locator doc) |

The two large audit/runtime JSONL prompt sets, the 1,200-request variant files, the
llama-70b follow-on files, and the 15-file split batch directory are contained inside
`wwt_cubie_demo_workload_1500.zip` rather than published as 30+ individual objects.

## SHA-256 of every file inside the zip (33 files)

```
f3ee9eef0b85944ff23d90effa19fd03974b981953be59edb171356efdae8240  README.md
4328782e49118475c4a61b7523b38128fa25dd3d974f54d2982985216316f40b  blended_prompt_set_1200.jsonl
713c8f9a0dca483b68fa1e19aa766182464813bea383175d8e20307a5c15064d  blended_prompt_set_1500.jsonl
729d10a67afc409a5129f8c436e9d843c38f753751c70ca60c888bd7d0b91887  bulk_processing_notes.md
d34fa6f0e3dfd1ca3e893bc51802ef9792af2f3891eb5dfe0305b7e4a22a443d  bulk_request_validation_report.json
0294b40c16a5d9f17e885334773664017b342d3181a93dbd506cb568f0cd4680  candidate_workloads.jsonl
9f673f97f9793155bc96184522a1b74a1ed5381082cc1ef91c2611c8d84c742c  exclusions.md
96e0da0b649b2a36094938a60b1bbd0167c54f264c6dcee7414c2f0de4c22430  litellm_batch_requests_1200_llama70b_followon.jsonl
b2a24b5907060278c784705cd742af680e4b77aa1d522f310f2ce421024a0942  litellm_batch_requests_1200_llama8b.jsonl
a0dcdceefc653b598b93ede09c6fe0c1206ab475f3a39394b541c88add9c3b83  litellm_batch_requests_1500_llama70b_followon.jsonl
d1f1afb2e3f98bc0ce913ce9533f4da423e0644d50ccc9779181604026ac3475  litellm_batch_requests_1500_llama8b.jsonl
7682629491a337a6816483aca422e6ee4918287b088ac627ee382da3504b53cf  litellm_batches_llama8b/batch_01_llama8b.jsonl
133d15a0cde2821192f70034398d8fd9ad43d1cb0ea7a1a61b9fb24b0de072e0  litellm_batches_llama8b/batch_02_llama8b.jsonl
56ef8b7bdf1006494da37f0ac1f7649948971cfd8fbcf913eec3075d5f4a5629  litellm_batches_llama8b/batch_03_llama8b.jsonl
94e07913a58dedf00a8b46e64083d48154f6e10552c43178ee689985499efd47  litellm_batches_llama8b/batch_04_llama8b.jsonl
65b390226d262dbbfa76d97c6993899517bb93e8e3a90fad45a9a90843ee2dce  litellm_batches_llama8b/batch_05_llama8b.jsonl
b9eec26f459b3bceae1c9cfc7e996750a0b5e7e5354f61e3274802c09992bd10  litellm_batches_llama8b/batch_06_llama8b.jsonl
d3168401b743d6b5453a417f843b039b47551f48e8b3eae7e4d810adddb1d6b5  litellm_batches_llama8b/batch_07_llama8b.jsonl
e6a7ea58f030215404003f94f1f6331c51f89c2e0249c40480f2092d7bfee1d6  litellm_batches_llama8b/batch_08_llama8b.jsonl
821147d310dedf2512b7b51a47e9f839ede37b67a4a6686c5aa6230dd3254949  litellm_batches_llama8b/batch_09_llama8b.jsonl
ba85fd61dc369a403e8a707405ac85b213578f4f6d1435a0b97444459e694dce  litellm_batches_llama8b/batch_10_llama8b.jsonl
91a3ead5812e503404ad3ac519809d0ca4ff33c73ea2d066beaaf2c8fab6947f  litellm_batches_llama8b/batch_11_llama8b.jsonl
1f28091dd47e47a86721abc8b539bb95e0a459d2956e2b245ddc46e1147cbe84  litellm_batches_llama8b/batch_12_llama8b.jsonl
7ec86a6d33edde1d9e81f51b5b2cd7b85ca80a314ab20644a277cc8dd1fe85fa  litellm_batches_llama8b/batch_13_llama8b.jsonl
b3bb1076630edf49907cfd573cea67403bcdc28e2d7acf7fa67180881f278db0  litellm_batches_llama8b/batch_14_llama8b.jsonl
cabdf5c75dda84de6383ee2411a29736bbfeff502eed74e78ae45270511f25d8  litellm_batches_llama8b/batch_15_llama8b.jsonl
3c7f8fcd4ca636510fe886e977ccff2ac44be1711a9cb6cca86134407f2f1879  mcp_openrouter_crash_patterns.jsonl
35fa4c2441a3bac5a8f2882fdb65a1661d197e851d1f26d42f923094737c372b  nist_mapping.csv
2f60981f12c5f2e6c01b9c1333ea460679a06f1dd74d9ca53532450af4bc9875  run_manifest.yaml
b0b1921f14eecb58e56d6c3fed775e858d3d175087778a457fddc719342af301  runtime_batch_manifest.csv
2932f0f520542f6883a4c5187a88ec9222fc64677adbca22a408aa19c5fa0393  source_register.csv
3b9dc3315dcc936eed1cf0daa27e75efcf3e96ddd064d760f30a50e86e1e6d99  submit_litellm_batch_example.sh
f9e85476138e3755f65f7e857ea8390ec7e258049b72dff634bd554958726bfa  workload_variants.md
```

## Reproduction

The runtime file `litellm_batch_requests_1500_llama8b.jsonl` is upload-ready JSONL
(one `POST /v1/chat/completions` request per line). Submit via
`submit_litellm_batch_example.sh` after exporting `LITELLM_BASE_URL` and
`LITELLM_PROXY_KEY`. All requests are temperature 0, max_tokens 512, model
`llama-8b`; the llama-70b follow-on must be submitted as a separate batch job to
avoid model-mismatch errors.