Research Note

AI System Identity Stack

One model name can hide several systems.

Aug 4, 20261 min readBy Dalton Anderson
In this article

AI System Identity Stack

One model name can hide several systems.

LayerWhat can change
WeightsRevision, adapter, merge, precision, quantization, tokenizer, and license
RuntimeLibrary, kernel, template, batching, hardware, and decoding
EndpointProvider, served revision, safety controls, caching, and updates
RouterModel selection, fallback, load balancing, and experiments
WrapperSystem prompt, retrieval, tools, memory, postprocessing, and retries
ProductUser state, policy, interface, rate limits, and connected data

An output from a private endpoint cannot be assumed to come from the public checkpoint with the same label. An output from a wrapper cannot be assigned entirely to the underlying weights.

The minimum comparison record needs the request path, artifact revision or provider, time, prompts, settings, and response evidence. Stylistic similarity can raise a question but cannot prove hidden identity.

Sources

Follow the evidence.

  1. arxiv.org: 1810arxiv.org
  2. crfm.stanford.edu: indexcrfm.stanford.edu
  3. HELM MMLU recordcrfm.stanford.edu
  4. daltonanderson.ghost.io: the ai that wasnt unmasking reflection 70bs hypedaltonanderson.ghost.io
  5. huggingface.co: Reflection Llama 3.1 70Bhuggingface.co
  6. huggingface.co: 458962ed801fac4eadd01a91a2029a3a82f4cd84huggingface.co
  7. huggingface.co: mainhuggingface.co
  8. huggingface.co: a376762159d10b8077c6a162ebd2f72267fe8a2fhuggingface.co
  9. huggingface.co: discussionshuggingface.co
  10. huggingface.co: 58huggingface.co
  11. huggingface.co: 59huggingface.co
  12. open.spotify.com: 4xX50HChI6FBLaYetiVZQHopen.spotify.com
  13. venturebeat.com: meet the new most powerful open source ai model in the world hyperwrites reflection 70bventurebeat.com
  14. daltonanderson.net: the ai that wasnt unmasking reflection 70bs hypedaltonanderson.net
  15. NIST AI Risk Management Frameworknist.gov
  16. nist.gov: towards best practices automated benchmark evaluationsnist.gov
  17. youtu.be: hmXfbvJOBY8youtu.be

From this episode

Two useful next steps.

Guide · 1 min

How to Reproduce a Language Model Benchmark

A reproducible LLM benchmark method covering model identity, datasets, prompts, runtime settings, scoring, raw outputs, deviations, and uncertainty.

Episode Story · 1 min

What Venture Step Got Wrong About Reflection 70B

A correction to E038 that separates Reflection 70B's versioned model-card claims, public and private evaluations, unresolved system identity, and unsupported conclusions.

Return to the episode