Inside Daisy's Runtime Upgrade: Better Roles, Recovery, and Models
By DigiEntropy Engineering · 2026-07-18 · 10 min read
We studied a cross-vendor prompt corpus, copied none of it, rebuilt Daisy's provider boundary, retired deepseek-chat safely, and measured the result across 49 real engineering scenarios.
About this post — Authored by an AI assistant using DigiEntropy's polynomial corpus, the universal compressor predictor, and the project's chart-generation tooling. Charts are produced by Python scripts that read the same database the live site queries; tables and formulas are pulled from the same engineering modules. Findings reflect the project's current dataset and methodology — send corrections or deeper questions to admin@digientropy.com.
The most useful lesson in a leaked system prompt is usually not the prose. It is the machinery around it: which message gets system priority, how tools describe themselves, when a recovery instruction appears, and what deterministic code refuses to leave to the model.
We reviewed the cross-vendor system-prompts corpus maintained by asgeirtj as an untrusted pattern library. We copied no hidden prompt, policy, persona, or vendor-specific instruction into Daisy. Instead, we extracted four testable strategies, measured the model line before touching production, changed one layer at a time, and ran the same 49 engineering scenarios again.
The result is a quieter but more important Daisy upgrade than a new button: GPT-5.2 is now the measured default, old deepseek-chat selections migrate safely to DeepSeek V4 Flash, system/user/tool roles survive the provider boundary, specialist answer rules arrive only after the relevant engine runs, and failed tools get one typed recovery path instead of a raw error string.
0.926 - GPT-5.2 DQS, the new default - held from 0.924
Canonical DQS before and after: DeepSeek V4 Flash rose from 0.871 to 0.881 and GPT-5.2 held from 0.924 to 0.926; both clear the 0.85 WOW bar.
0.881 - DeepSeek V4 Flash DQS - up from 0.871
DeepSeek V4 Flash rose from 0.871 to 0.881; it clears the 0.85 WOW bar.
49 - engineering scenarios - same judge, same engines
The authoritative after result merges the full GPT-5.2 row from the two-model canonical run with the final full DeepSeek hardening run. Same 49 scenarios, same judge, same engines.
8.4 s - V4 Flash median latency - down from 12.6 s
Latency improved. DeepSeek V4 Flash moved from a 12.6 s median / 26.8 s p90 to 8.4 s / 19.1 s.
822 - characters off the prompt head - 4.0% of 20,618
The permanent system-prompt head fell from 20,618 to 19,796 characters: 822 fewer characters, or 4.0%. That is a contained pilot, not the 20% target proposed for a future full-tool migration.
0 - hard gates, fabrications, fallbacks - on both production choices
Final result: zero hard gates, zero fabrication, zero provider fallback on both production choices.
The model deadline came first
…