Make the vision skill honest: no fabricated image features

The old image_stub invented an observation ("mock_texture_descriptor = lamellar")
— self-contradicting for a tool whose whole point is evidence honesty. Replace it
with a vision hook that ships no model and returns NO features: when an image is
given but no vision model is configured, image-derived features are unavailable
and surface downstream as missing evidence, never as made-up data. A real vision
model still plugs in here later.

- skills/image_stub.py → skills/vision.py (returns []); observation_agent import
  updated; reproducibility skill stamp = "vision:not-configured".
- workflow limitations + README + architecture updated.
- New test: image-only input invents no features and forces measure_again. 20/20 pass.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: rockyzl

README.md

@@ -185,7 +185,8 @@ human-in-the-loop scientific workflows. See [`docs/community-strategy.md`](docs/
 
 Research prototype. Interpretations are tentative and must be confirmed by a domain
 expert. Literature retrieval runs over a small curated offline knowledge base (not a
-live literature API); image features are still mocked in this version. Do not use for
+live literature API). This offline build ships no vision model, so image-derived
+features are unavailable — reported as missing evidence, never invented. Do not use for
 real scientific or safety decisions.
 
 ## License

docs/architecture.md

@@ -39,8 +39,9 @@ applied to scientific reasoning.
 ## Extensibility
 
 - **Skills** (`skills/`) are pluggable capabilities. v0 ships a real offline
-  `literature` retrieval (stdlib TF-IDF over a curated knowledge base) and a still-mocked
-  `image_stub`; a real vision model replaces the latter without touching the spine.
+  `literature` retrieval (stdlib TF-IDF over a curated knowledge base) and a `vision`
+  hook that ships no model — rather than fabricate image features it returns none, so
+  they surface as missing evidence. A real vision model plugs in without touching the spine.
 - **Instruments / tools** become connectors later (Phase 3: MCP-compatible). The spine
   never needs to know whether evidence came from a microscope, a database, or a
   simulation — only its `EvidenceKind` and confidence.

src/scientific_agent_lab/agents/observation_agent.py

@@ -3,7 +3,7 @@
 from __future__ import annotations
 
 from ..schemas import ObservedFeature, ScientificInput
-from ..skills.image_stub import features_from_image
+from ..skills.vision import features_from_image
 
 
 def observe(inp: ScientificInput) -> list[ObservedFeature]:

src/scientific_agent_lab/skills/image_stub.py

@@ -0,0 +1,18 @@
+"""Vision skill — the pluggable hook for image-derived observed features.
+
+By design this offline, zero-dependency build ships NO vision model, so it does
+NOT invent image features: a tool whose whole point is evidence honesty must not
+fabricate its own observations. When an image is provided but no vision model is
+configured, image-derived features are simply unavailable — if the question
+depends on them, they surface downstream as *missing evidence*, never as made-up
+data. A real vision / foundation model plugs in here by returning ``ObservedFeature``
+items with genuine provenance (that is the Phase 1-4 work in the roadmap).
+"""
+from __future__ import annotations
+
+from ..schemas import ObservedFeature
+
+
+def features_from_image(image_ref: str) -> list[ObservedFeature]:
+    # No vision model configured in this offline build → no fabricated features.
+    return []

src/scientific_agent_lab/skills/vision.py

@@ -39,7 +39,7 @@ def _reproducibility(
 ) -> ReproducibilityRecord:
     skills = [f"literature@{__version__}:offline-tfidf-kb"]
     if not inp.observations and inp.image_ref:
-        skills.insert(0, f"image_stub@{__version__}:mock")
+        skills.insert(0, f"vision@{__version__}:not-configured")
     return ReproducibilityRecord(
         input_sha256=_sha16(asdict(inp)),
         report_sha256=_sha16(report_dict),
@@ -56,7 +56,8 @@ def _reproducibility(
     "This is a research prototype, not a validated scientific decision system. "
     "Interpretations are tentative and must be confirmed by a domain expert. "
     "Literature retrieval runs over a small curated offline knowledge base (not a "
-    "live literature API); image features are still mocked in this version."
+    "live literature API). This offline build ships no vision model, so image-derived "
+    "features are unavailable — reported as missing evidence, never invented."
 )
 
 

src/scientific_agent_lab/workflow.py

@@ -30,3 +30,18 @@ def test_workflow_recommends_measurement_when_gap_exists():
     report, _ = _report()
     # a required measurement is missing -> agent must not ACCEPT
     assert report.recommended_next_measurement.action == NextAction.MEASURE_AGAIN
+
+
+def test_image_only_input_invents_no_features():
+    # no vision model configured -> the pipeline must NOT fabricate observations;
+    # required evidence is then honestly missing and it must not accept.
+    inp = ScientificInput.from_dict({
+        "question": "What phase is in this image?",
+        "domain": "microscopy",
+        "image_ref": "region.tif",
+        "required": [{"name": "phase_id", "kind": "measurement", "why": "identify the phase"}],
+    })
+    report, _ = run_workflow(inp)
+    assert report.observed_features == []
+    assert report.recommended_next_measurement.action == NextAction.MEASURE_AGAIN
+    assert any(m.name == "phase_id" for m in report.missing_evidence)