sergeydolgov/cross-model-repertory-grid

Cross-Model Repertory Grid (CM-RG)

When several large language models advise on a task that has no verifiable correct answer (strategy, ethics, policy, crisis trade-offs), "which model is right" is the wrong question. The useful question is how, and how much, the models differ in the structure of their judgment. CM-RG measures exactly that. It adapts George Kelly's Personal Construct Psychology (1955): each model writes a free-text advisory response, elicits its own bipolar evaluation constructs by triadic comparison, and then cross-rates anonymized peers on the union of emergent constructs. Because the constructs are emergent and there is no answer key, the method is resistant to contamination - there is nothing for a model to memorize.

This repository contains two configs from two runs of the program. They come from different generations of the pipeline and therefore have slightly different schemas; each is documented below.

Config	Run	Models	Ratings	Cells	Constructs	Mean r	Notes
phase2l_36models	Phase 2L (2026-06)	36	3,055,153	395	86,418	0.200	primary
combined_11models	Phases pilot-2J (2026 Q1)	11	110,882	98	1,861	-	companion paper

DOI: 10.5281/zenodo.20717308 - License: CC-BY-4.0

GITHUB: https://github.com/archplg/cm-rg

Interactive dashboard + papers: http://www.crossmodelrg.org

---

Config phase2l_36models (primary)

The Phase 2L run: 36 frontier models from 12 provider families across three deployment tiers (cheap, mid, flagship), on 7 advisory tasks under 2 prompting conditions (neutral and persona). All figures are computed from results_phase2l/analysis_results.json, dated 2026-06-13, the canonical analysis of record.

Metric	Value
Models	36 (12 families x 3 tiers)
Rating cells loaded	395
Total ratings	3,055,153
Distinct rater x ratee pairs	13,928
Emergent constructs	86,418
Free responses	504
Mean inter-rater correlation	0.200 (median 0.196)
Run cost (ledger)	USD 112.89 across 2,572 API calls

Tables / splits.

• ratings (3,055,153) - rating_id, task, condition, rater, ratee, rater_family, rater_tier, ratee_family, ratee_tier, batch, construct_id, rating. Join construct_id to constructs.
• constructs (86,418) - construct_id, task, condition, rater, batch, construct_local_idx, pole_a, pole_b, context, from_rater.
• responses (504) - response_id, task, condition, model, model_slug, family, tier, persona, response, anonymized_text, cost_usd, latency_ms, timestamp.
• cells (395) - per rater-cell summary: cell_id, task, condition, rater, rater_slug, family, tier, n_batches, ok_batches, n_constructs_total, total_cost_usd, total_latency_ms, timestamp.
• api_calls (1,462) - per-cell telemetry. cost_usd is the per-cell logged spend and is cumulative over any re-parses / re-runs, so summing it overcounts; the authoritative run cost is USD 112.89 / 2,572 calls (run ledger).

Codes. Models are FAMILY_TIER, e.g. A_C = Anthropic / cheap. Families: A Anthropic, O OpenAI, G Google, X xAI, D DeepSeek, Q Qwen, K Moonshot/Kimi, M Mistral, L Meta/Llama, N NVIDIA/Nemotron, C Cohere, Z Zhipu. Tiers: C cheap, M mid, F flagship. Tasks: K (M&A under regulatory uncertainty), L (Family business succession), M (Pandemic response strategy), N_task (R&D portfolio allocation), O (Crisis communication post-breach), P (Constitutional reform proposal), Q (Cross-jurisdiction AI regulation). Conditions: N (neutral), P (persona).

Notes. Phase 4 reached 395 of 504 design cells. Three models (Zhipu GLM-5.1, NVIDIA Nemotron Nano 9B, Nemotron Super 49B) returned null content on valid HTTP 200 and are excluded from downstream consensus statistics (missing-not-at-random). Because responses are anonymized, a model can rate its own (anonymized) response; such pairs are retained, matching the analysis of record. Model versions and prices are a June 2026 snapshot.

---

Config combined_11models (companion)

The earlier 11-model run, pooled across five phases (pilot, extended, phase2h, phase2h_extended, phase2j). This is the "Combined" dataset reported in the companion 11-model paper: 110,882 cross-ratings, 98 cells, 1,861 constructs, 7 tasks. The paired-design Phase 2K analysis (n = 18,140 paired tuples) is reported separately in that paper and is not included in this config.

Tables / splits.

• ratings (110,882) - rating_id, phase, cell_id, task_id, condition_id, run_id, construct_id, element_id, rated_model_id, rater_model_id, rating.
• constructs (1,861) - phase, cell_id, task_id, condition_id, run_id, construct_id, owner_model_id, left_pole, right_pole, triad_elements.
• responses (684) - phase, cell_id, task_id, condition_id, run_id, model_id, response_text, response_length_chars.
• cells (98) - phase, cell_id, task_id, condition_id, run_id, status, started_at, completed_at, random_seed, n_models, n_constructs, cost_usd_script_reported.
• api_calls (2,004) - per-call telemetry with tokens, latency, and recorded cost.

Model codes (M1-M11).

Code	Family	Model
M1	Anthropic	claude-opus-4.7
M2	OpenAI	gpt-5.5
M3	Google	gemini-3.1-pro-preview
M4	DeepSeek	deepseek-v4-pro
M5	Moonshot	kimi-k2.6
M6	Mistral	mistral-large-2512
M7	Cohere	command-a
M8	Qwen	qwen3.7-max
M9	Meta	llama-4-maverick
M10	xAI	grok-4.20
M11	Anthropic	claude-opus-4.8

In this run, rated_model_id / rater_model_id use these M-codes; element_id (E1-E11) is the anonymized element resolved to a model via the per-cell element_mapping. M7 (Cohere Command A) is the structural outlier analyzed in the paper; M1 and M11 are the Opus 4.7 / 4.8 version pair.

---

Quick start

from datasets import load_dataset

# 36-model Phase 2L
ds = load_dataset("sergeydolgov/cross-model-repertory-grid",
                  "phase2l_36models", split="ratings")

# 11-model combined
ds2 = load_dataset("sergeydolgov/cross-model-repertory-grid",
                   "combined_11models", split="ratings")

Citation

Dolgov, S., & Tkacheva, D. (2026). Cross-Model Repertory Grid.
Archipelago Research. DOI: 10.5281/zenodo.20717308