Skip to content

■ SIGNALS // RADAR SIGNAL

same Signature, different prompt

DSPy 3.5 changes default prompts and API routing; NVIDIA revises skill-evaluation accuracy; Egypt completes the EU technical assessment for electronic signatures and seals.

■ [!] ON THIS PAGE ▼

self.md radar — 2026-11-15

same Signature, different prompt

DSPy 3.5.0 shipped on November 10 with a warning to re-evaluate optimized programs
the migration guide says keeping the same Signature does not preserve the prompt

1. DSPy 3.5 switches the default to XML

DSPy’s migration guide replaces the implicit ChatAdapter with XMLAdapter(use_native_function_calling=True), changing demonstrations, output instructions and parsing

explicitly constructing XMLAdapter() still leaves native function calling off

API routing also changes
model_type="auto" selects Responses for OpenAI and Azure, even with the LiteLLM engine
a Chat-only proxy needs an explicit model_type="chat"; a provider failure does not trigger a second try through another backend

restoring ChatAdapter() alone does not restore the old behavior
saved examples and quality scores need rechecking
ordinary old cache entries miss under the new format, so allow for recomputation too

reading: DSPy 3.5.0 release · the tagged upgrade guide

2. a correct answer could lose points for not naming the absent skill

SkillEvaluator’s earlier accuracy check included whether the agent identified the skill
in a positive test case, a correct answer from the no-skill arm could score at most 0.8

NVIDIA’s 0.9.0 release, published November 13, removes that penalty from outcome accuracy
the documentation warns that results under the old judge contract are not directly comparable with the new ones

the release also lets authors reuse a retained no-skill baseline, but only when the dataset, model, environment, judge, attempt policy, evaluator version and workspace configuration match
reusing evidence does not create another independent attempt

in the all-tier CI integration, Tier 3 live evaluation remains advisory
the old tier3-required settings are ignored, even when set to true
standalone validation can still make live evaluation block with --block-on-agent-eval

reading: SkillEvaluator 0.9.0 changes · the tagged live-evaluation guide

3. Egypt clears the EU’s technical signature assessment

the European Commission says Egypt’s electronic-signature and seal schemes have passed their technical assessment under eIDAS
its November 13 notice describes Egypt as the first country outside Europe to reach this technical alignment

the assessment covers Egypt’s Trusted List and confirms compatibility that allows EU Member States to validate the signatures and seals
the announcement does not establish that incorporation into the EU Third Countries Trusted List is already live

reading: the Commission’s assessment announcement

wording in the notice:

Egypt will be incorporated into the EU Third Countries Trusted List