entroly

Matched token-cap quality frontier

Entroly wins every matched token-cap quality gate against External Baseline A.

60 frozen SQuAD v2 long-context trials; Entroly 1.0.59 source candidate vs released External Baseline A 0.31.0; achieved ratios use o200k_base. Active-context scope: External Baseline A’s CCR pointers remain in its output, but retrieval recovery is not invoked; this is not an end-to-end External Baseline A CCR comparison.

Requested compression	System	Answer retained	Actual tokens kept	p50 latency
2x	Entroly	95.0%	39.3%	3.2 ms
2x	External Baseline A	1.7%	18.6%	9.7 ms
4x	Entroly	93.3%	19.2%	2.5 ms
4x	External Baseline A	1.7%	18.6%	9.5 ms
8x	Entroly	88.3%	10.4%	2.4 ms
8x	External Baseline A	1.7%	18.6%	28.8 ms

Paired retained-answer statistics

Target	Entroly only	External Baseline A only	Exact McNemar p
2x	56	0	2.8e-17
4x	55	0	5.6e-17
8x	52	0	4.44e-16

Local short-answer guard

Model: qwen2.5:1.5b at the 4x target. This is a local-model guard, not a hosted frontier-model claim.

Context	Exact match	Token F1	Trials
Raw	62.5%	80.2%	8
Entroly	87.5%	93.3%	8
External Baseline A	12.5%	12.5%	8

Reproduce and verify

python -m benchmarks.compression_frontier verify benchmarks/results/compression_frontier.json

Scope limits

• SQuAD v2 contexts test extractive answer retention, not full agent behavior.
• Requested target ratios are controls, not assumed outcomes; achieved ratios are primary.
• The local Ollama answer guard is not evidence about hosted frontier models.
• No retrieval or post-compression recovery is invoked in this active-context benchmark.
• External Baseline A is measured through its released public compress() API and declared config.
• Entroly is a 1.0.59 source candidate; publication must follow before claiming released-package parity.

Payload SHA-256: 30606a062d8cf96fc803c78ce93289c4b7adf02b8067456e012acb23231a0b1b

This site is open source. Improve this page.