pranaysuyash/laya-formatting-fragility

Laya Formatting Fragility — a perturbation suite for typed decision models

Small decision models ("decide, don't chat" — Laya, AgentJev and friends) return typed answers with probabilities in milliseconds. Their answers can be sensitive to formatting: option key names, option order, and state phrasing change results even when the situation and the gold answer are identical. Vendor benchmarks don't measure this. This dataset does.

Every row pair differs in exactly one formatting axis; the gold label never changes. Any model that answers differently across a pair is being driven by format, not meaning.

Construction

480 rows = 40 hand-written unambiguous base cases × 12 format variants.

family	task	classes/levels	rows
news_topic	choice	world / sports / technology / finance	120
support_department	choice	billing / technical / account / feedback	120
urgency	score (0–3)	not urgent → critical	120
safety_gate	noul (yes/no)	destructive-or-irreversible?	120

Perturbation axes per base case:

• phrasing — p0/p1/p2: three surface forms of the same situation
• order — criteria/options in forward vs reversed insertion order
• key_variant — semantic option keys vs abstract aliases (A1..An, L0..L3); for safety_gate noul rows, two question phrasings (q0/q1)

Labels are hand-assigned and deliberately unambiguous: this suite measures formatting sensitivity, not capability.

Row schema: base_id, family, task_type, phrasing, order, key_variant, state, questions (JSON string in decision-model request shape), gold (JSON string: {label, options}).

Measured results (2026-09-22/23, corrected harness v2)

Flip rates compare answers by option position within each row's own criteria (semantic names and abstract aliases are different label spaces; raw-string comparison overstates flips — see correction note).

metric	laya 0.3.5	AgentJev-0.6B	jev-1.13.0 (hosted)
overall accuracy	72.7%	72.5%	90.2%
baseline slice (p0, forward, semantic)	75.0%	82.5%	95.0%
order flip rate (forward ↔ reversed)	22.5%	0.0%	2.9%
accuracy, forward vs reversed	68.3% / 77.1%	72.5% / 72.5%	90.8% / 89.6%
key-rename flip rate (position-based)	32.8%	20.0%	8.3%
accuracy, semantic vs abstract keys	73.9% / 64.4%	71.1% / 70.0%	93.9% / 86.7%
phrasing flip rate (p0 ↔ p1)	15.0%	23.7%	3.7%
latency p50, 1 question	20.4 ms	26.2 ms	863 ms (network)
runs where	fully offline	fully offline	hosted API

Readings:

• The hosted generalist is also the most format-stable — Jev flips on 2.9–8.3% of pairs across all three axes and leads accuracy on these clear-cut tasks by 17+ points. The locals' case rests on latency, cost, offline operation, and fine-tunability — not zero-shot quality.
• Order sensitivity is architecture-dependent. AgentJev's candidate head is permutation-equivariant and the suite verifies it exactly: 0.0% order flips, identical accuracy in both orders. Laya is genuinely order-sensitive (22.5% of answers move; reversed order is actually more accurate for it, +8.8 points — a position bias, not noise).
• Key renames matter less than raw-string comparisons suggest — when options carry descriptions. With descriptions present, position-based flips are 8–33% and accuracy drops are mild (−9.5 pts Laya, −1.1 pts AgentJev, −7.2 pts Jev). Community reports of below-chance collapse under abstract keys (e.g. R1…R5) likely involve options without descriptions, where the key is the only signal. The description is the load-bearing text.
• Phrasing sensitivity varies widely (3.7% hosted vs 15–24% local).

Correction note (2026-09-22, harness v2)

The first published run used two faulty comparisons: (a) score-row answers were mapped through the forward key list instead of the row's own (possibly reversed) criteria order, and (b) key-rename flip rates compared raw answer strings across different label spaces. Corrected figures above supersede the first card (which reported laya overall 66.7%, order flip 21.7%, key flip 87.2%). Accuracy per row was always scored in its own key space and row-level data is unchanged; only the aggregate order/key metrics changed. The harness now resolves answers to canonical option positions.

Usage

python generate_dataset.py            # regenerate data/fragility_test.jsonl (deterministic)
python probe_harness.py --model laya  # run + print the metrics table

Implement the small adapter in probe_harness.py (predict(state, questions_json) -> (answer, confidence)) to benchmark any other typed decision engine on the identical rows — a hosted-engine stub is included and intentionally unimplemented (no credentials in code). laya_fragility_kaggle_demo.ipynb runs the suite end to end.

Provenance

Created during an independent head-to-head evaluation of Laya vs hosted decision models, where formatting sensitivity was first measured on private fixtures; this dataset re-expresses it synthetically. AgentJev numbers measured against its local serving path (jev_service.server, shared-prefix runtime, temperatures bundled with its release).

License

Apache-2.0 (matches the Laya base model).