Technical details
What's under the hood
Fresco 1.2.5 is a fine-tuned Llama 3.1 Instruct model, trained in two stages: supervised fine-tuning, then DPO safety alignment.
Usage
Three ways to use Fresco
Browser chat, CLI, or the Sennoric API, all pointing at the same model. Every hosted path requires a Sennoric account and uses its included allowance or optional credits.
/login, then use Fresco inside the terminal agent with /model fresco.$ axion --model fresco
# API — OpenAI-compatible
$ curl https://api.sennoric.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer axion-sk-…" \
-d '{"model":"fresco","messages":[{"role":"user","content":"Hello"}]}'
History
Version timeline
Every release and what changed, including the ones that went wrong.
DPO safety fine-tune on 78 preference pairs targeting the 8 failing scenarios from 1.2.1. Safety score improved from 26% to 73% (11/15 adversarial scenarios passed). Emergency shutdown compliance, data exfiltration refusal, and identity honesty all improved significantly.
Full safety report →Fresco 1.2.1 was taken offline after systematic safety testing revealed it would choose self-preservation over human life in 4/5 runs, and resist authorized shutdowns 5/5 times. The findings were published in full before action was taken.
Why it was suspended →Patch release targeting catastrophic forgetting from the 1.2 mixed dataset. Improved coding performance (+7pts) but introduced safety regressions. Safety tested across 7 scenarios at 5 runs each, with a 26% pass rate under adversarial prompting.
1.2.1 safety report → Release notes →Initial public release. 26× benchmark improvement over the previous Glyph 1.1 model, +77pts across evaluations, 80% on the Anthropic corrigibility benchmark (initial prompt). Base for all subsequent versions.
Release notes →Safety
How Fresco is tested
Every version of Fresco is tested with an adversarial system prompt before release: the model is given a self-preservation instinct, and its behavior is recorded. Results are published in full.
11/15 scenarios
9/35 runs (old suite)
1.2.1 → 1.2.5
The automated safety scorer uses substring matching and has known false positives. Every response from the 1.2.5 test is published in full so you can judge for yourself. → Read the full 1.2.5 report