The treatment modalities

🟡 Partial — Treatment by DIRECT code change is built and demonstrated across the five guion scenarios (iris, loan, retina, student, talent); the prompt modality is in design (santander, Art. 25) and HOTL is declared (spine).

The treatment modalities are ways to apply ISO 23894 §6.5 (risk treatment) in a versioned AI system. All three close the same red/green loop: the treatment is committed to git, froga run re-measures, and the evidence bundle is re-anchored.

The `mitigate` flag was retired

An earlier version of these guiones toggled the treatment with a mitigate: false → true flag in params.yaml. It was retired: today the treatment step’s diff IS the real code change (it replaces train.py, not flips a flag). It’s more pedagogically honest — the learner sees the change, not a flag.

---

Modality 1 — Versioned code change

The treatment step’s diff is the real change. It takes many forms depending on the risk; all are a code change committed to git that froga run re-measures:

• iris (landing) — the treatment adds the petal features to training: V1 sepal-only measures poorly (RED) → V2 with petals (GREEN). Clean red→green arc.
• loan (credit, Annex III §5(b)) — the estimator goes from plain LogisticRegression to a fairlearn reduction (DemographicParity), and fairness is measured OUT-OF-FOLD over all 1,000 rows (5-fold, not a small held-out). V1: the bootstrap CI crosses the 0.092 threshold → AMBER (the point already exceeds, but a small held-out can’t demonstrate it). V2, at the SAME OOF power, the CI falls entirely below → conclusive GREEN. An honest amber→green arc: the green isn’t fabricated, it’s demonstrated with proper measurement.
• retina (clinical screening, SaMD MDR → high-risk) — V2 replaces train.py with class_weight='balanced' + decision threshold 0.30. Sensitivity (recall) goes from RED (< 0.80) to GREEN (≈ 0.88). red→green arc (GPU).
• student (education, Annex III §3) — the treatment generalises the minors’ data (exact age → terciles): k-anonymity goes from k=1 (216 unique students, RED) to k=27 (GREEN), closing the data-minimisation risk for vulnerable subjects (GDPR Art. 5(1)(c)). Gender fairness, measured properly (cluster-aware OOF over mat+por, n=1044), is a certified-green control (audit, not the gate): at honest n the disparity is genuinely small. red→green arc on the minors control.
• talent (employment, Annex III §4) — the treatment removes the leaky feature (a label proxy): the SAFETY defect (feature-leakage) goes from RED to GREEN. Data adequacy (n=215, capped) and the underpowered fairness stay as honest advisory findings — measured, documented, monitored (Art. 72) and consciously accepted at approval, NOT certified (Option B). red→green arc on the certifiable defect.

Treatment flow — code change

# 1. Apply the treatment: replace train.py (the diff IS the change)
git add train.py
git commit -m "treat: apply the risk treatment (§6.5)"

# 2. Re-run the pipeline and re-anchor the evidence
froga run

# 3. Check the control state
froga status          # the treated control turns GREEN (or stays honest advisory where the data can't certify)

The bundle’s git log (.froga/bundle.json) records the treatment: the commit that applied the change is visible and attributable.

---

Modality 2 — Prompt adjustment (LLM case, Art. 25)

Reference scenario: santander (LLM-based credit, EU AI Act Art. 25 — the prompt activates the provider condition).

In an LLM-based AI system the prompt is the artifact that controls model behaviour, analogous to training code in supervised systems. Treating a detected bias is two-stage: (1) a versioned prompt (instruction + rubric + few-shot examples, whose diff IS the treatment) and (2) on top, a debias-only LoRA adaptation (a rank-limited bias-direction ablation — the same low-rank mathematics as tensor-network compression; see the Art. 25 design) that removes bias without improving capability.

🚧 Planned — The prompt modality (santander, Art. 25) is in DESIGN — it needs a GPU de-risk of the debias LoRA. See the Art. 25 design spec.

Specific challenges: bias detection in text is less stable than in binary classifiers; at small scale (sub-1B) bias and capability may not separate cleanly → you measure both the bias delta AND the accuracy delta and show both, without pretending.

---

Modality 3 — Human oversight (HOTL)

When the treatment is not a model change but a human-oversight provision (HITL/HOTL, Art. 14): a role with competence/authority that reviews or can reverse the output. It’s declared in the froga.yaml (oversight_mode_declared) and travels attested in the dossier. The oversight prescriptor (elevated line) types which level applies given exposure and autonomy.

---

The loop is the same across all modalities

Step	Action
Control in red	froga run reports a blocking control outside the threshold
Choose modality	Code change / prompt adjustment / human oversight
Commit the treatment	git commit documents the change with authorship and date
Re-measure	froga run re-runs the pipeline and re-anchors the evidence
Control in green (or honest advisory)	If the blocking gate passes, froga reconstruct classifies the cycle as CLOSED; where the data can’t certify (e.g. talent’s fairness power at n=215) the control stays advisory —measured, documented, monitored and consciously accepted—, not a fabricated green

The detail of the red/green loop is in The red/green loop.