coral-bricks-ai/endpoint-accuracy-replica

Endpoint accuracy replica

Does an inference endpoint keep a model's accuracy? This measures it the way Artificial Analysis' Endpoint Accuracy Index does: run the same questions against the endpoint and against a reference endpoint for the same model, and report the endpoint's score as a share of the reference's. An endpoint is at parity when the 95% interval of that share includes 100%.

This is CoralBricks' own replica of AA's published method, not AA's index, and is not affiliated with or endorsed by Artificial Analysis. It covers two of the index's three evals (BFCL and AA-LCR); HLE is not included yet.

Results

Runs from 2026-10-14 and 2026-10-15. Accuracy is the mean over repeats, with a 95% Student-t interval.

Endpoint	BFCL v4 (500)	AA-LCR (25)
CoralBricks deepseek-v4.1-flash-fast	73.3% ±0.3 (3 repeats)	90.0% ±2.0 (10)
DeepSeek API deepseek-flash (reference)	74.1% ±2.3 (3)	88.4% ±2.1 (10)
CoralBricks glm-5.3-fast	75.3% ±0.8 (3)	80.4% ±3.1 (10)
Fireworks glm-5p3 (reference)	76.7% ±4.2 (3)	81.6% ±2.7 (5)

Endpoint	Reference	BFCL share	AA-LCR share	Index (2 of 3 evals)
CoralBricks DeepSeek V4.1 Flash	DeepSeek's own API	99.0% ±3.2	101.8% ±3.3	100.4% ±2.3, at parity
CoralBricks GLM-5.3	Fireworks GLM-5.3	98.2% ±5.5	98.5% ±5.8	98.4% ±4.0, at parity

The GLM comparison is with a peer provider, not with the model maker: no first-party GLM-5.3 reference has been run. The Fireworks reference ran AA-LCR 5 times rather than 10.

Method

• BFCL v4-500, with bfcl-eval 2026.3.23 (the version AA names), its own test entries, multi-turn simulators and checkers. 500 offline entries weighted towards multi-turn and parallel calls (subsets/bfcl.json). Java, JavaScript, memory, web search and live relevance categories are left out.
• AA-LCR-25: 25 questions from AA-LCR v1.1 (subsets/lcr.json), each with about 100k tokens of documents, AA's document template, and AA's checker prompts. Graded by GPT-5.6 Luna at medium reasoning (through OpenRouter), the checker AA uses; it sees only the final answer, never the reasoning.
• Each endpoint's highest reasoning mode, the lab's recommended sampling (DeepSeek: temperature 1.0, top_p 1.0; GLM: temperature 1.0, top_p 0.95), streaming, tool choice auto, and up to 30 retries on API failures.
• Every endpoint gets the same 131,072-token output cap, so a gap cannot come from a different cap.
• Subsets are drawn once with a fixed seed (2026) in the shape AA describes; AA does not publish its own subsets, so ours are not identical to AA's.
• The index is the mean of the per-eval shares; the per-eval intervals carry both sides' standard errors and are combined in quadrature.

Where this differs from AA

• References. AA serves the official weights itself. We use the model maker's API for DeepSeek (AA reports DeepSeek's first-party endpoint at or slightly above its reference) and Fireworks' standard deployment for GLM-5.3.
• Two evals of three. No HLE results yet.
• Subsets are ours, drawn the same way, not AA's.

Files

Path	What
results/summary.json	Accuracy per endpoint and eval, and the index
results/bfcl.jsonl	Every BFCL question in every repeat: endpoint, repeat, category, id, correct
results/lcr.jsonl	Every AA-LCR question in every repeat: endpoint, repeat, id, correct, finish reason, tokens
subsets/	The question ids used
code/	The harness: bfcl_run.py, qa.py, llm.py, report.py, subsets.py, endpoints.py

Model answers are not included. AA-LCR's documents are not included either; they are third-party publications, available from the AA-LCR dataset.

Running it

uv venv .venv --python 3.12
uv pip install --python .venv/bin/python bfcl-eval==2026.3.23 soundfile "pyarrow<18" openai
cd code
export CORALBRICKS_API_KEY=... DEEPSEEK_API_KEY=... OPENROUTER_API_KEY=...
python bfcl_run.py coralbricks-deepseek     # BFCL x3, each repeat in a fresh process
python qa.py coralbricks-deepseek lcr       # AA-LCR x10; put the AA-LCR dataset in code/data/aa-lcr first
python report.py --json ../results/summary.json

bfcl-eval keeps its multi-turn simulators in module globals, keyed by model and test id, so a second repeat in the same process starts from the first one's end state. bfcl_run.py runs each repeat in its own process for that reason.

Licenses

The code and our results are MIT. BFCL's test entries are Apache-2.0 (gorilla); AA-LCR's questions are Apache-2.0. Neither dataset is redistributed here beyond question ids.