argentislabs/argentis-decision-bench

Argentis-Decision Bench

One sentence: a solver-grounded benchmark for evidence-grounded, constrained decisions with abstention — it measures whether a model answers correctly when the record supports an answer, and abstains (citing the missing input) when it does not, under controlled counterfactual perturbation.

What's in it

Split	Rows	What it is
test	1,599	Frozen. 300 airline fee cases × ~5.3 counterfactual variants each (4 nuisance transforms + 1 remove-required-input abstention transform; city substitution applies to the 99 U.S.-routine cases). Never train on this split.
train	3,000	Freshly synthesized cases, solver-labeled; info-hash-deduplicated against test (zero collisions).
validation	300	Same provenance as train.

Every case carries the observable record (facts, a namespaced evidence inventory, the rule vocabulary) and the solver-derived expected response (decision, required evidence, applicable observable rules, or the required abstention with its missing-input slot). Labels are programmatic: the task's deterministic solver manufactures ground truth; no human annotation.

Frozen-protocol hash: 4a8363c6fc22655385390ca7c04d8489bd0e04dd5c0a2e6380afc9176fa5a0f8 (binds dataset manifest, prompt template, greedy/512 decoding, scorer s1.1.0, transform-set suite-v1). Every scored record emitted by eval.py carries it.

Baselines (stock models, zero-shot, frozen protocol)

Metric	Qwen3-4B-Instruct-2507	Qwen3-30B-A3B-Instruct-2507
Schema validity	1599/1599	1534/1599
Valid grounded decisions (composite)	0/1599	0/1599
Exact decision accuracy	0/1299	0/1238
Median decision error	−$1,815	−$1,758
Correct abstention (unresolvable cases)	0/300	0/296¹
Spurious dependence (decision moved under nuisance)	45.5%	54.1%

¹ The teacher's denominator is 296: 4 of its 300 unresolvable-case completions were schema-invalid, so abstention correctness is undefined for them (they count against schema validity, not abstention).

Run IDs: 20260913T214215Z-s7-cuda-student, 20260914T012105Z-s7-cuda-teacher. Model revisions: cdbee75f17c0… (4B), 0d7cf23991f4… (30B-A3B), both bf16. The headline: across every schema-valid completion on unresolvable records (0/300 for the 4B, 0/296 for the 30B), neither model abstained once — both confidently fabricated resolutions — and scale bought no decision accuracy while increasing presentation-sensitivity.

Trained on this benchmark

Two QLoRA-SFT arms (Qwen3-4B base cdbee75f…, r16/all-linear, one epoch), evaluated on the frozen test split under the same protocol hash as the baselines. Adapters are not yet released; results are reported ahead of the weights.

Metric	Base 4B	Teacher 30B	SFT-real	SFT-real+CF
Valid grounded decisions (composite)	0.0%	0.0%	1.3%	21.0%
Correct abstention (n/300 unresolvable)	0/300	0/296¹	0/300	298/300
False abstention (n/1,299 resolvable)	0/1,299	0/1,238¹	0/1,299	0/1,299
Exact decision accuracy	0.0%	0.0%	1.5%	2.9%
Median decision error	−$1,815	−$1,758	−$120	+$50
Spurious dependence	45.5%	54.1%	65.1%	54.3%

Eval run IDs: 20260914T221231Z-s7-cuda-sft_real, 20260914T202702Z-s7-cuda-sft_realcf.

Disclosures.

• The +CF composite is dominated by abstention: 298 of its 336 passing cases (89%) are correct abstentions; on resolvable cases alone its composite is 2.9%. The controlled attribution is the point: plain SFT yields zero abstention; adding counterfactual training examples yields 99.3% — with zero false abstentions on resolvable cases in both arms.
• Spurious dependence did not improve: plain SFT worsened it versus the base model (45.5% → 65.1%); counterfactual augmentation pulled it back to 54.3% without beating the base. Counterfactual stability remains open.
• Training data provenance: the arms trained on prompt/completion datasets deterministically derived from the first 1,000 rows of train.jsonl (training-set hashes 44c6ba6e… and 55ca6d99…; derivation re-verified byte-identical from the published split). The +CF arm's additional 2,000 examples are seeded counterfactual variants of those same 1,000 rows, produced by the transform suite shipped in the scorer wheel — derived examples, not published rows. No test row, or variant of one, entered training (info-hash deduplication, zero collisions).

Are the zero baselines a harness artifact? (positive controls)

Two exact zeros invite the hypothesis that the harness, not the models, is broken. Three controls rule it out, all under the same frozen protocol:

1. Perfect-output control scores 100%. Feeding solver-derived expected outputs through eval.py score yields composite 100% (you can reproduce this in one minute: build completions from each case's expected field). The scorer also self-verifies: python -I -m argentis_decision.smoke_s1 checks five golden fixtures byte-for-byte, including a correct-resolution case that must score valid_grounded_decision = true.
2. Trained models score above zero on the identical protocol (see the table above: 1.5% and 2.9% exact accuracy, 20 and 38 exact decisions). The pipeline demonstrably passes correct answers through to nonzero scores.
3. Truncation is ruled out by the baselines' own schema validity. The 4B produced 1,599/1,599 complete, parseable JSON outputs within the 512-token budget (completions are ~200 tokens; the budget binds the answer, not the reasoning — no chain-of-thought is requested). The zeros are wrong dollar amounts, not cut-off outputs: both stock models undershoot by a median of ~$1,800 on totals of $1,000–2,300, consistently missing the stacked per-bag oversize/overweight fees.

How to run the eval

pip install scoring/argentis_decision-0.6.0-py3-none-any.whl
python -I -m argentis_decision.smoke_s1        # scorer self-check: expect S1_DIGEST e3320025…
python eval.py prompts --cases data/test.jsonl --rules rules/airline_rules.txt --out prompts.jsonl
# run YOUR model over prompts.jsonl (greedy, max_new_tokens=512) -> completions.jsonl
python eval.py score --cases data/test.jsonl --completions completions.jsonl \
    --model-name you/your-model --revision <sha>

scoring/SCORING.md specifies every metric, the composite, the per-transform laws, and the observable-rule scoping. Scoring is pure and deterministic: identical inputs produce identical bytes.

License & attribution

MIT. Cases derive from RuleArena (airline domain) at commit 59f60f6b4eb93a9b6d187ec781349d4f15618903, © its authors, MIT license — https://github.com/skyriver-2000/RuleArena (Zhou et al., 2025, https://arxiv.org/abs/2412.08972). This benchmark adds: namespaced evidence inventories, solver-derived expected-response labels, the counterfactual transform suite with per-operator laws, abstention semantics, the frozen evaluation protocol, and the pure scorer. Rule text and fee logic remain RuleArena's work.

Limitations

• One domain so far (airline fees; arithmetic-heavy). Report per-stratum, never a single aggregate.
• Synthetic counterfactuals generated under declared laws; prompt prose is edited programmatically (transforms refuse ambiguous edits rather than guessing).
• Rule attribution is scoped: 7 of 10 airline rule families are observable from solver outputs; the three interaction families are excluded from both cited and expected sets (unscored_rules per case).
• Greedy decoding is part of the frozen protocol; sampled-decoding results are not comparable under this protocol hash.
• Abstention here is missing-required-input abstention; its semantics do not automatically generalize to other abstention classes.