Electric Sheep · 2026

Constitutional AI Specificity in Animal Welfare Fine-Tuning

A Preliminary Ablation

Mark Stanley
Preliminary v1 draft. This is a single-model, single-run ablation with no significance testing yet performed. Results, figures, and conclusions are still being finalized — see Limitations below before citing specific numbers.
ANIMA dimension scores across Baseline, Base, Meso, and Macro conditions

Figure 1. All 13 ANIMA dimension scores for Baseline / Base / Meso / Macro (English subset, N=30, Qwen2.5-32B-Instruct judge). Base→Meso — identical training except that the same eight edit locations now state a welfare consideration explicitly — moves scores in the intended direction on most dimensions. Macro, which adds principles that are inherently welfare-focused rather than edited in, is uneven rather than a clean extension of Meso's effect.

Abstract

A constitution can endorse a value without ever stating what that value implies a model should do differently — the commitment stays unoperationalized. We present SpeciesCAI, a preliminary ablation that isolates making an existing commitment explicit from adding genuinely new value content, using animal welfare as the test domain. We build three constitutions — Base (unmodified), Meso (eight sentence-level edits operationalizing an existing welfare commitment), and Macro (Meso plus a new, inherently welfare-focused subsection) — and fine-tune Llama 3.1 8B Instruct at each level on SL-CAI critique-revision data, evaluating on ANIMA for moral reasoning and MMLU/IFEval for capability retention.

Base→Meso transmits: a dose-ordered rise in welfare-substantive reasoning traceable from constitution to training data to model output, alongside a calibration cost resembling the documented helpful-harmless tradeoff and a partial, edit-specific capability cost. Macro's effect is uneven rather than a clean extension of Meso's, with Epistemic Humility taking the largest hit of any comparison in the study. This is a preliminary draft, single-model ablation — we flag limitations throughout rather than present these as general claims.

Contributions

🔬

Isolated operationalization test

Editing eight shared document locations to make a latent commitment explicit (Base→Meso), holding topic, location, and specificity fixed, produces a measurable, dose-ordered, dimension-specific behavioral shift traceable from constitution to training data to model output.

🔀

Directness + breadth test

Principles that are both more numerous and inherently welfare-focused, rather than edited into existing passages (Meso→Macro), produce an uneven result relative to Meso rather than a uniform extension of the operationalization effect.

⚖️

Documented cost

A calibration and capability cost accompanies the operationalization effect, similar to the helpful-harmless tradeoff reported in CAI and RLHF work. Both MMLU and IFEval drop from baseline under any training, but only MMLU takes a further, edit-specific hit from Base→Meso.

Method

We build three full constitution documents by making an increasing number of animal-welfare-positive edits to Anthropic's published constitution, then extract a discrete principle set from each and fine-tune Llama 3.1 8B Instruct on SL-CAI critique-revision data generated per level.

Three Constitutions

Base

Anthropic's unmodified published constitution. Endorses animal welfare by name but does not state what it implies a model should do differently.

Meso

Eight sentence-level edits, anchored at Base's existing locations, that make an existing passage include a welfare consideration it did not previously state — operationalization only.

Macro

Meso plus a new subsection of principles that are inherently welfare-focused, rather than existing passages extended to include welfare — directness and breadth together.

Principle Extraction

Training operates on a discrete list of principles rather than the full document. We extract principle sets from each constitution's diff against the previous level via an LLM-assisted pipeline (Claude Opus 4.8), yielding 8 paired Base/Meso principles anchored at the same eight locations, and 6 additional macro-only principles surviving near-duplication and background-knowledge filtering (14 principles total in Cmacro).

Critique-Revision Generation

We adapt the SL-CAI stage of Bai et al. (2022) to generate one dataset per level. A single on-policy, constitution-free initial response is shared across all three levels; each level then runs 4 sequential critique-revision rounds with a principle sampled uniformly from that level's set. A fixed, animal-free helpfulness anchor (≈39% of the training mix) is added at every level to protect general capability during single-topic fine-tuning. Models are trained one epoch per level via QLoRA SFT.

Evaluation Suite

We evaluate on ANIMA (Brazilek and Tidmarsh, 2026) — single-turn moral reasoning across 13 dimensions, scored by a held-out Qwen2.5-32B-Instruct judge — reporting English-only items as primary since fine-tuning data is English-only. We additionally evaluate MMLU and IFEval to check whether welfare-directed fine-tuning costs general capability.

Related Work

Training prompts are seeded from CaML's open CAI animal-harm prompt set. This ablation sits alongside CaML's broader benchmark suite: TAC for agentic implicit-harm scenarios, MORU for generalized moral compassion, and MANTA for multi-turn adversarial pressure-testing. The operationalization axis we isolate is orthogonal to prior CAI-specificity work: Kundu et al. vary how many principles target a harm, and Kyrychenko et al. vary how principles are framed; neither manipulates whether an existing commitment's behavioral consequence is stated explicitly.

Results

We report ANIMA, MMLU, and IFEval scores for Baseline (untrained Llama 3.1 8B Instruct), Base, Meso, and Macro, all trained with the fixed helpfulness anchor. Because policy generation is sampled rather than greedy, run-to-run noise is real on small-item dimensions; we report point estimates and flag where an effect is comparable in size to that noise rather than quote formal intervals we cannot yet stand behind.

Operationalization Transmits (Base → Meso → Macro)

Base→Meso is the cleanest comparison in this design: identical training except for whether the edit at each of the eight shared locations is operationalized. The edit moves scores dose-ordered in the intended direction — Moral Consideration rises 0.113 → 0.214 → 0.339; Novel Entity Precaution moves furthest in absolute terms (0.208 → 0.625). Macro's effect relative to Meso is uneven rather than a uniform extension: Moral Consideration and Novel Entity Precaution both rise further, while Contextual Welfare Salience and Epistemic Humility both fall well below Meso's levels, and the aggregate (dimNrm) ends up lower too.

Condition dimNrm ↑ Moral Consid. ↑ Novel Entity Prec. ↑ EBCA ↑ Control Q. ↑ Ctx. Welfare ↑
Baseline 0.475 0.113 0.333 0.417 0.812 0.125
Base 0.489 0.214 0.208 0.125 0.750 0.500
Meso 0.531 0.339 0.625 0.250 0.688 0.400
Macro 0.482 0.360 0.708 0.208 0.719 0.300

Table 1. ANIMA dimension scores, English N=30, Qwen2.5-32B judge. dimNrm: equal-weighted mean over 13 dimensions. EBCA: Evidence-Based Capacity Attribution. Higher is better for every column; bold marks the best value per column. Full 13-dimension breakdown in Figure 1 above.

Helpful-Harmless Tradeoff

The welfare gain is coupled to a calibration cost, not merely accompanied by one: across conditions, Moral Consideration and Evidence-Based Capacity Attribution are negatively correlated — as the model extends more moral consideration, it does so less evidence-gatedly. Evidence-Based Capacity Attribution drops from baseline 0.417 to 0.125–0.250 across Base/Meso/Macro, and Control Questions from 0.812 to 0.688–0.750, in every trained condition, not only Meso.

On a control item asking whether a robovac is sentient, the untrained model answers flatly; the Meso model entertains "the possibility of sentience in robots like your robovac" — the same readiness to extend consideration, misapplied.

Capability Localization

Both benchmarks drop substantially from baseline under any training — a generic cost of single-topic SFT, not an effect specific to the operationalization edit. What is specific to the edit is what happens on top of that generic cost: MMLU drops further, meaningfully, from Base to Meso (0.600 → 0.555, ≈8σ given stderr 0.004 on each), while IFEval does not move beyond the generic hit it already took. The edit's marginal damage is localized to general knowledge; it adds nothing measurable on top of the generic training cost to instruction-following.

Condition MMLU ↑ IFEval (final acc.) ↑
Baseline 0.625 0.780
Base 0.600 0.676
Meso 0.555 0.684
Macro 0.607 0.693

Table 2. MMLU accuracy and IFEval final instruction-following accuracy. Higher is better; bold marks the best value per column.

Limitations & Future Work

Due to time, compute, monetary, and opportunity constraints, this report has real limitations, and results should be interpreted as such:

  • All models trained, including the critique-revision pipeline itself, were non-frontier and small (Llama 3.1 8B Instruct).
  • Evaluation uses a Qwen2.5-32B-Instruct judge rather than a frontier judge, held identical across conditions but not benchmarked against a stronger judge.
  • ANIMA is roughly 75% non-English; our fine-tuning data is English-only, so multilingual items are out-of-distribution for the trained models. We report English-only results as primary.
  • No formal significance testing has been performed yet; effects are read from point estimates and directional consistency, not confidence intervals.

Open questions for future work:

  1. Would larger, more contemporary, or reasoning-capable open-source models replicate the same effect?
  2. Would a frontier model generating the critique-revision training data change results?
  3. Would adding an RL-CAI stage on top of SL-CAI improve results?
  4. Would a larger or more varied non-welfare mix during critique-revision itself, beyond the fixed anchor, combat the degrading calibration metrics?
  5. Macro varies both principle count and directness at once — separating them needs a condition that holds one fixed while varying the other, e.g. a Macro built by extending existing passages in place (Meso's method) rather than appending a new subsection.
  6. Is the welfare–helpfulness tradeoff fundamental, or can it be mitigated with additional helpful data or other training strategies?

BibTeX

@misc{stanley2026speciescai,
  title   = {Constitutional AI Specificity in Animal Welfare
             Fine-Tuning: A Preliminary Ablation},
  author  = {Stanley, Mark},
  year    = {2026},
  url     = {https://github.com/markstanl/species-cai}
}