> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cosentriq.com/llms.txt
> Use this file to discover all available pages before exploring further.
# The CoLinear Score: Four Dimensions of Human Trust
> The CoLinear Score measures Output Trust, Reasoning Clarity, Experienced Accuracy, and Improvement Signal — each mapped to a different buyer at your organization.
The CoLinear Score is a multi-dimensional readout of how real humans from your target population experience your AI model's outputs. It is not a single number — it's four distinct dimensions that each answer a different question from a different person at your company. Your product team, your engineering team, your research team, and your leadership all have different things they need to know about your model's market readiness. The Score is built to give each of them something they can act on.
## Why Four Dimensions
A single aggregate score compresses information that should stay separate. A model can have high output trust and low reasoning clarity — users trust the outputs but can't follow the logic, which is a UX and explainability problem, not a model performance problem. A model can have strong benchmark accuracy and poor experienced accuracy — technically correct, but the target population doesn't experience it that way in real conditions.
Each dimension of the CoLinear Score surfaces a distinct signal. Together, they give you a complete picture of the gap between what your model does and whether the humans it's built to serve will actually use it.
## The Four Dimensions
**Delivered as**: A confidence band
**Maps to**: Your product team
Output Trust measures the degree to which users in your target population actually trust the outputs your model produces — independent of whether those outputs are technically accurate. A model can be accurate and still generate low trust if the outputs are formatted in a way that reads as uncertain, incomplete, or inconsistent with how users expect results to look.
**Use it to**:
* Understand adoption risk before launch — low output trust is a leading indicator of underutilization
* Surface trust signals you can use in positioning and onboarding
* Identify output formatting, framing, or presentation issues that are suppressing trust independently of model quality
**Delivered as**: A confidence band + MIA written explanation
**Maps to**: Your engineering team
Reasoning Clarity measures whether the users your model serves can follow the reasoning your model presents and take the action that reasoning implies. This dimension is particularly important for models that produce explanations, recommendations, or summaries where the reasoning chain is visible to the user.
**Use it to**:
* Identify whether your model's explanations are intelligible in the context where they'll be read — not just technically parseable
* Understand whether users can locate the specific step in the model's reasoning where clarity breaks down
* Prioritize explainability and output structure improvements that will have the most impact on user confidence
The MIA written explanation gives your engineering team a synthesized narrative of where reasoning clarity succeeds and fails — with specific scenario examples drawn from contributor evaluations.
**Delivered as**: Sentiment aggregation numbers + MIA written explanation
**Maps to**: Your research and evaluation team
Experienced Accuracy measures the gap between benchmark accuracy and the accuracy users actually feel in their real-world context. A model can be 94% accurate on a test set and still be experienced as inaccurate by users who encounter it in the edge cases, the ambiguous inputs, and the real-world messiness that held-out data doesn't capture.
**Use it to**:
* Quantify the felt accuracy gap — the difference between your benchmark score and the accuracy your target population actually experiences
* Understand which scenarios or input types are generating the most accuracy skepticism
* Build the case for evaluation that goes beyond held-out test sets to include population-specific human signal
The MIA written explanation provides a synthesized analysis of which dimensions of your model's outputs are generating the strongest felt accuracy and which are generating the most skepticism.
**Delivered as**: A narrative feedback score
**Maps to**: Your leadership and strategy team
Improvement Signal synthesizes the human feedback from your sprint into a structured narrative covering data needs, improvement strategies, and capability gaps. This is the dimension that tells you what to do next — not just what the score is.
**Use it to**:
* Prioritize model improvements and retraining investments with human-sourced signal behind them
* Identify the specific data gaps that are limiting your model's performance with your target population
* Build a roadmap for the next sprint or the next training cycle with clear, defensible rationale
The narrative feedback score is written for leadership and strategy audiences — it's designed to be communicated upward without requiring technical translation.
## Score Delivery and Access
Your CoLinear Score lives in your membership dashboard — it is not a one-time PDF delivery. It's a persistent, interactive record of your model's human trust profile.
Available directly from your Score view. Use MIA to interpret what your dimensions mean, explore improvement strategies, and get help communicating findings to stakeholders who weren't in the room for the sprint.
Every completed sprint adds to your run history. Track how your Score changes across evaluation cycles as you improve your model, adjust your outputs, or expand to new populations.
Download cleaned scenario data from your sprint as a CSV file. Use it to feed human evaluation signal back into your training pipeline or fine-tuning workflow.
Share specific Score dimensions or the full report with stakeholders through your dashboard. Each dimension is formatted to communicate cleanly to its intended audience.
## Quality Guarantee
Before MIA synthesizes your Score, MCGA runs **Inter-Annotator Agreement (IAA)** across all contributor data. This process identifies contributor responses that are inconsistent with the rest of the evaluation set, flags low-quality submissions, and triggers automatic redo on any responses that don't meet the quality threshold.
Your Score is only synthesized after IAA passes. This means every confidence band and every sentiment aggregation is grounded in contributor data that has already been validated for internal consistency.
Use the CSV download to feed cleaned scenario data back into your training pipeline. Real human evaluation data from your actual target population is among the highest-quality retraining signal available — especially for the edge cases and real-world inputs that synthetic test sets miss.