Run Stage C

This skill runs the image-side cross-modal appraisal transfer experiment using the frozen or previously learned text-side machinery.

Use this skill when

• The task is to run the multimodal or image-side transfer experiment.
• A user refers to “Stage C,” “cross-modal transfer,” “EMOTIC evaluation,” or image-conditioned appraisal readout/steering.
• Stage A outputs already exist and the next goal is transfer validation.

Do not use this skill for

• Cold-start environment setup.
• Pure text-only replication.
• Final write-up except for concise run summaries.

Objective

Test whether text-derived appraisal directions or probes transfer meaningfully to the multimodal setting, and determine whether observed effects support genuine transfer rather than a weaker explanation such as caption mediation.

Required precondition

Assume Stage A should already exist. Before running, verify:

• which probe/checkpoint/direction is being imported,
• whether its training setup is documented,
• whether the imported artifact is the intended one.

If Stage A is missing or ambiguous, stop and ask to resolve that dependency unless the user explicitly requests a provisional or exploratory run.

Read these first

• CLAUDE.md
• docs/experiment-1.md
• docs/datasets.md
• docs/models-gemma3.md
• docs/probes.md
• any Stage A output manifest, metrics file, or checkpoint directory
• relevant multimodal code under src/experiments/, src/data/, src/bridge/

Stage C workflow

Step 1: Confirm transfer design

Determine:

• the image dataset(s),
• label space and any mapping rules,
• which frozen text-derived directions/probes are used,
• which multimodal model path is primary,
• whether transfer is readout-only, steering-based, or both,
• the baselines that must be reported.

Never proceed without a clear mapping between source-side artifacts and target-side evaluation.

Step 2: Validate image-side data and mapping

Check:

• dataset presence and accessibility,
• subject/image counts if documented,
• split integrity,
• preprocessing steps,
• class or appraisal mapping logic,
• any ambiguous or lossy category conversions.

If mapping quality is a major source of uncertainty, make it a first-class caveat in the result summary.

Step 3: Verify multimodal forward path

Before running a full experiment:

• test one sample end to end,
• confirm image preprocessing is correct,
• verify the model receives both text and image inputs as intended,
• confirm hook positions exist in the expected path,
• check output shapes and extracted representations.

Do not scale until the single-sample path is verified.

Step 4: Apply frozen directions or probes

Load the Stage A artifact(s) and apply them consistently:

• use the documented hook locations,
• preserve preprocessing conventions where required,
• avoid silent re-fitting on image-side data unless the experiment explicitly calls for it.

If any adaptation is required, document exactly what changed and why.

Step 5: Evaluate transfer

Report the planned metrics and compare against:

• no-steering or no-transfer baseline,
• simple caption or verbalization baseline when applicable,
• random or shuffled controls if defined in the repo.

The point is not just to show an effect, but to show that the effect supports the intended hypothesis.

Step 6: Stress-test alternative explanations

Where the design allows, check whether apparent transfer could instead be explained by:

• caption mediation,
• label leakage,
• domain-shift artifacts,
• mapping artifacts,
• prompt sensitivity.

If the transfer signal collapses under these checks, say so clearly.

Step 7: Save outputs for analysis

Persist:

• configs,
• metric tables,
• selected examples,
• plots if the repo expects them,
• run notes emphasizing transfer interpretation and caveats.

What to produce

At the end, produce:

• transfer verdict,
• key metrics,
• baseline comparisons,
• strongest caveats,
• output paths,
• recommendation for next action.

Interpretation rules

Use careful language:

• “supports transfer” only if the evidence beats the relevant baselines,
• “inconclusive” if the effect is weak, unstable, or confounded,
• “fails to support transfer” if results do not survive basic controls.

Guardrails

• Do not retrain text probes inside Stage C unless the protocol explicitly requires it.
• Do not hide lossy label mappings.
• Do not interpret caption-mediated success as direct evidence of shared multimodal appraisal geometry.
• Do not overstate novelty from a single positive metric.
• Do not skip baseline comparisons.

Completion criteria

This skill is complete when:

• the image-side pipeline has run,
• transfer metrics are computed,
• alternative explanations are addressed at least minimally,
• outputs are saved,
• the result is classified as supportive, inconclusive, or unsupportive.

Example invocations

• “Run Stage C using the frozen Stage A probe.”
• “Evaluate cross-modal transfer on EMOTIC.”
• “Test whether the transfer beats a caption baseline.”
• “Diagnose why the multimodal steering run is failing.”