Canonical Reasoning Complexity (MHC Stage)
In calibration
What it measures. When you are about to trust a model with work that has real structure, you need to know how much complexity it can actually sustain, not how it did on a quiz. This scale measures the developmental order of hierarchical complexity a performer, human or AI, can sustain: the flagship ruler every other AIM scale re-anchors onto. It does not measure knowledge, style, or speed; a fluent wrong answer at a low stage is still a low-stage answer. A benchmark can tell you two models tied on a test form. This ruler tells you where each one sits on the complexity order, with the standard error stated at every band. Frozen bank, audit-ready.
S1. Construct definition
Boundary. It does not measure knowledge, style, or speed; a fluent wrong answer at a low stage is still a low-stage answer.
Decision context. Feeds capability sufficiency, model selection, and work redesign decisions (the three-decisions block on the Science page).
S2. The Rasch ruler
Rasch family measurement model; program pre-registered on OSF; stage definitions follow the published Model of Hierarchical Complexity. The stage boundaries are not yet calibrated on human data; that calibration is the next step.
Measurement frame. Units are logits on the MHC complexity order, with stage definitions that follow the published Model of Hierarchical Complexity. A one-logit difference is a constant odds ratio everywhere on the ruler: intervals mean the same amount at every stage. The model estimates in logits. Every figure on this page names its unit.
S3. Stated SE per band
Every published measurement carries a stated standard error, stated per band, never as a single scale-wide average. Averages hide the floor.
The per-band record from the 2026-08-22 run is off this page while its level labels are corrected. Once they are, it will appear on the Evidence page as a historical record.
| Band | Band label | Standard error | Certification implication |
|---|---|---|---|
| -- | BELOW FLOOR Per-band judge-evidence readout Model rows whose band severity SE exceeds the frozen quality floor render on the Science page select-ruler, flagged, never presented as a pick. | -- | Shown and flagged, never removed. |
| -- | NO EVIDENCE Per-band no-evidence models A model with no per-band judge evidence at a band is listed under no evidence, never silently dropped. Both row classes render on the Science page select-ruler readout. | -- | Shown and flagged, never removed. |
S4. Certification tier
In calibration
Set to an error bar of 0.20. Delivered at 0.21 once judge harshness is corrected at each level. Simulation: 15,000 decisions on judge data from our own runs, seed 20260908. With harshness left in, the same run came in at 1.16. Error bars in logits. The 2026-08-22 gate is kept as a historical record on the Evidence page.
Tier as of 2026-10-01. Tier changes are events; the promotion rules for WATCH scales are stated in the canonical scale-priority record.
S5. Honest-uncertainty display
Rows that fail the floor or lack evidence are shown and flagged, never removed.
The flagged rows below are fixtures standing in for the live select-ruler readout; the rendered rows with real model names are on the Science page.
The flagged rows render in the S3 table above with the shared .below-floor and .no-evidence classes, the same flagged treatment as the model-ratings readouts. Dropping a failing row is a defect class, not a style choice.
S6. API and usage detail
Status. Live in the demo.
POST a response set against the MHC bank; receive a band location on the complexity order with the standard error stated at the value.
POST /score {"scale": "mhc", "responses": [...]}
{"band": "S11", "location_level": "<stated in levels of the ruler>", "se": "<stated at the value>"}
Rate and plan per the pricing page. The newer scales must clear SE 0.10 per band.
The full API reference lives on the docs surface when it is funded (IA spec section 4); until then this block is the usage detail of record.
S7. Scientific references
- 2026 Barney, M., Wind, S., & Krishna, V. (2026). Using large language models to evaluate ethical persuasion text: A measurement modeling approach. IJATE, 13(1), 224-247. https://doi.org/10.21449/ijate.1788563
- 2016 Barney, M.F. & Fisher, W. P., Jr. (2016). Adaptive Measurement and Assessment. Annual Review of Organizational Psychology and Organizational Behavior, 3, 469-490. https://doi.org/10.1146/annurev-orgpsych-041015-062329
- OSF Calibration study pre-registration and audit artifacts (Open Science Framework), per the Science page research program.