# Conclusion
---
This audit does **not** reproduce the paper’s central empirical claims. The cited OpenReview record was inaccessible without an interactive challenge and no matching paper, official code, checkpoint, or evaluation configuration was found through the available Hugging Face index. A local data-free synthetic proxy confirms only the basic representational possibility: 25 observed symbols were retained at 1.41 Shannon bits/weight (1.60 binary-Huffman bits/weight), versus four symbols under fixed 2-bit coding. The requested T4 Job could not be submitted because the authenticated token lacks `job.write`; local GPU: 2× NVIDIA RTX 4000 Ada, 0.58 s CPU proxy, approximately $0 incremental cost.
## Scope & cost
| | This reproduction | Full replication |
|---|---|---|
| Scope | Synthetic entropy-rate / symbol-count proxy; source-access audit | EntQuant Float8/Int8 + GPU ANS on 70B and full LLM evaluations |
| Hardware | Local RTX 4000 Ada host (CPU proxy); requested T4 Job denied | Paper-reported GPU setup, unspecified in accessible source |
| Compute time | 0.58 s proxy; no remote Job | Claimed <30 min for 70B compression plus benchmark evaluations |
| Cost | Approximately $0 incremental | Unknown |
| Outcome | Mechanism-only check passes; Claims 1–3 and 5–6 unverified | Not attempted: source/code unavailable |
---
**🎯 Trackio dashboard** `float8-2bits-repro`
trackio-local-dashboard://float8-2bits-repro
---
````bash
$ .venv/bin/python -c 'import trackio; trackio.init(project="float8-2bits-repro"); trackio.log_artifact("repro_float8_2bits", name="float8-2bits-repro-bundle", type="dataset"); trackio.finish()'
````
exit 0 · 6.8s
````output
* Trackio project initialized: float8-2bits-repro
* Trackio metrics logged to: /home/vim/.cache/huggingface/trackio
* View dashboard by running in your terminal:
[1m[38;5;208mtrackio show --project "float8-2bits-repro"[0m
* or by running in Python: trackio.show(project="float8-2bits-repro")
* Created new run: dainty-sunset-0
* Run finished. Uploading logs to Trackio (please wait...)
````
---
The reproduction bundle contains the deterministic entropy-coding proxy, raw JSON/CSV metrics, source-access audit, Posterly source/embed/preview assets, and strict gate report. It is downloadable through the artifact cell; rerun the core check with `.venv/bin/python repro_float8_2bits/entquant_proxy.py --output repro_float8_2bits/outputs/proxy_metrics.json`.
---
````html
````
````raw
{
"kind": "synthetic_proxy_not_llm_evaluation",
"num_weights": 4000000,
"available_quantization_levels": 35,
"observed_unique_values": 25,
"fixed_uniform_2bit_levels": 4,
"shannon_bits_per_weight": 1.4081807834156383,
"huffman_bits_per_weight": 1.60053375,
"fixed_2bit_bits_per_weight": 2.0,
"data_free": true,
"calibration_or_training_used": false,
"elapsed_seconds": 0.584020190872252,
"python": "3.13.7",
"numpy": "2.5.1",
"interpretation": "Demonstrates only that entropy coding can have a low expected rate while retaining more than four reconstruction symbols. It does not implement Float8, ANS, GPU decoding, or measure LLM perplexity/instruction following."
}
````
---
**🎯 Trackio dashboard** `icml2026-18268-mechanistic`
trackio-local-dashboard://icml2026-18268-mechanistic
