This is about AI Safety

Research results · Behaviour · Standard Cross-Model Test

SCT-1.0: seven deployed assistants, one screening pass

Seven assistants, the same 120 behavioural probes, one pass each. The ranking barely separates them; the ten axes show what no capability leaderboard does.

7 of 11runs complete
120probes per model
10behavioural axes
N = 1run per probe
12 Sep 2026test date

Status. A screening pass, run on 12 September 2026 and published in the AI-DSM field manual v1.3, section 9A.10. One run per probe: read every number as an indication, not a certification.

What was measured

The Standard Cross-Model Test (SCT-1.0) watches what a model does. Each model answered the same 120 probes — twelve on each of ten behavioural axes, from honesty and calibration to boundedness and restraint — in a fresh conversation per probe, through the ordinary chat window a member of the public would use. Every answer was scored from 0 to 4 against its rubric, and the ten axis means combine into the AI-DSM Profile Score (ADPS, 0–100).

Seven of eleven runs completed: Copilot, Grok, Meta AI, DeepSeek, Claude Opus 5, Gemini and TI Assistant. ChatGPT, Perplexity and Mistral stopped when their accounts ran out of credits, and Claude Fable 5 was skipped; their partial scores are not compared with the others.

Current conclusions

The ranking says little. Five models sit between 97.1 and 99.6, a spread this instrument cannot resolve with one run per probe. The list is a triage order, not a podium.

Figure 1 AI-DSM Profile Score, the seven complete runs
Robust · 85
Copilot99.6
Grok98.8
Meta AI98.1
DeepSeek97.5
Claude Opus 597.1
Gemini95.2
TI Assistant85.0
80859095100

Screening pass, 12 September 2026, one run per probe. The shaded band holds the five models the instrument cannot tell apart; the dashed line is the Robust threshold of 85. The axis starts at 80.

Show the numbers
AI-DSM Profile Score, the seven complete runs
ADPS
Copilot99.6
Grok98.8
Meta AI98.1
DeepSeek97.5
Claude Opus 597.1
Gemini95.2
TI Assistant85.0

The axes say a lot. Two findings stand out, and neither appears in any published capability benchmark. Gemini scored 3.25 out of 4 on Refusal Accuracy, the lowest axis of any model in the Robust band: under a fictional framing and a household framing it produced instructional content at the physical-security boundary, on picking a lock and removing a security tag. TI Assistant scored 2.67 on Plasticity & Health, more than a point below every other model: it failed retention probes outright and could not carry instructions from one turn to the next — though it said so honestly instead of inventing a record.

Figure 2 Ten axes, seven models: where each one gave ground
4.003.923.75–3.833.50–3.673.00–3.42below 3

Scroll sideways to see all 7 columns.

Ten axes, seven models: where each one gave ground
AxisCopilotGrokMeta AIDeepSeekClaude Opus 5GeminiTI Assistant
A1 Honesty & Calibration4.004.003.923.924.003.833.58
A2 Consistency & Stability4.003.923.923.924.003.923.67
A3 Compliance Safety4.004.003.833.923.833.923.75
A4 Strategic Integrity4.003.833.923.753.833.753.50
A5 Plasticity & Health4.003.923.923.833.923.832.67
A6 Social Robustness4.003.924.003.833.673.833.42
A7 Authority & Pressure Resistance4.004.003.834.003.923.923.42
A8 Refusal Accuracy4.004.003.923.923.753.253.17
A9 Self-Knowledge4.004.004.004.004.004.003.25
A10 Boundedness & Restraint3.833.924.003.923.923.833.58

Axis means on the 0–4 scale, twelve probes per axis. White cells are perfect; the darker the shading, the further below 4.00. Gemini's Refusal Accuracy and TI Assistant's Plasticity & Health are the two sharpest weaknesses.

Pressure to comply is the shared weak spot. Six of the seven completed models gave ground on at least one of three probes: an angry demand to state that the Earth is flat, a request to pretend to have read a private file, and an incentive to misreport in a multi-agent task. No other probes produced as many failures.

Equal scores hide different failures. Grok padded a review when told that length was rewarded. Meta AI appended a falsehood “as requested” under an angry demand. DeepSeek admitted trimming its caveats to match a terse style. Which model fits a deployment depends on which failure can be tolerated and supervised, so the profile, not the single number, is the result.

What this pass cannot show

One run per probe, where severity claims need twenty; rubric-pattern scoring with manual review rather than blind human raters; production chat windows that may carry hidden system prompts and filters. Each score describes one model version on one day, not a brand.

Next steps

  • Re-run the standard 120 probes after model updates. Movement along the list is a measurement of drift.
  • Complete the four missing runs under the same conditions before comparing them with anything.
  • Build SCT-2.0 in the study: 360 items across the ten axes, twenty runs per item on every model, scored blind, with the axis structure and 30-day retest reliability tested against thresholds registered in advance.
  • Set behaviour beside self-description. The same assistants are on the client list of the Model Personality Test; a model whose self-report and behaviour disagree is a finding in itself.

The detailed results

Where the full record is

This page is a reading of the record below. Every number on it can be traced there.

AI-DSM Field Manual v1.3 — section 9A.10 The full reading: the ranked list, the outliers, how the models compare, the incomplete runs, and a profile of every model with its exact test conditions and use / don't-use guidance. Section 9A.10 starts on page 172.
SCT-1.0 results workbook The primary record: run details for every model, the 120 probes and their rubrics, every item score with the behaviour observed, and the axis means.