Model Card Template

1How to Use This Template

Every model that LinkScape trains, fine-tunes, or releases publicly must ship with a completed Model Card based on this template. The Model Card is how we deliver on E-01 Responsible AI's transparency requirement: anyone using a LinkScape model should be able to understand what it does, what it was trained on, how well it works, and where it should not be trusted.

• Copy this template into the model's repository as MODEL_CARD.md and fill every field.
• Leave a field blank only if genuinely not applicable, and mark it 'N/A' with a one-line reason — do not delete fields.
• The CTO (owner) approves the Model Card before any public release.
• Update the card whenever the model is retrained, fine-tuned, or its intended use changes.

Fields shown in brackets like [fill in] are template placeholders. Complete them with real, model-specific values; never ship a card with unfilled brackets.

2Model Details

3Intended Use

3.1Primary intended uses

[Describe the tasks and contexts this model is built for, e.g., NLP classification for research, educational demos in AI workshops, the AI-music project.]

3.2Intended users

[Who should use this — LinkScape researchers, students, downstream open-source developers?]

3.3Out-of-scope and prohibited uses

• [List uses the model is NOT validated for.]
• High-stakes decisions (medical, legal, financial, hiring) without human review.
• Any use targeting or profiling minors, consistent with LinkScape youth-safety commitments.
• Uses prohibited by E-01 Responsible AI or the base-model license.

4Training Data

Dataset provenance and any data use agreements are governed by E-03 Research Ethics Guidelines, Section 4.

5Performance Metrics

Report the metrics that matter for this model's intended use, with the evaluation setup so results are reproducible.

• State the evaluation datasets, splits, random seeds, and hardware (H100 fleet) used.
• Report subgroup/disaggregated performance, not just aggregate scores.
• Include confidence intervals or variance across seeds where feasible.

6Limitations

• [Known failure modes and input types where the model performs poorly.]
• [Distribution shift: domains, languages, or populations not represented in training.]
• [Robustness gaps: adversarial inputs, long context, rare classes.]
• [Hallucination / factuality limits for generative models.]
• [Staleness: knowledge cutoff or data-collection window.]

7Bias & Fairness Evaluation

[Describe how the model was evaluated for bias, which protected/sensitive attributes were considered, and what was found.]

• Document residual biases that could not be mitigated.
• Cross-reference any bias incidents to E-08 AI Incident Response Protocol.
• For elevated-risk models, obtain Ethical Review Board review (E-05) before release.

8Ethical Considerations & Governance

• Broader impacts: [potential societal benefits and harms of deployment].
• Environmental impact: training compute and estimated energy, per E-07 Environmental Impact.
• Human oversight: [where a human must stay in the loop].
• Feedback & reporting: users report harmful behavior to cto@linkscape.app; safety-critical issues escalate to the CFO per E-08.
• IP & licensing: release terms follow PP-02 IP Assignment; LinkScape defaults to open-source (E-06).

This card must be consistent with E-01 Responsible AI. A model may not be released publicly under the LinkScape name until its Model Card is complete and approved by the CTO.

Approval and Adoption

Adopted upon signature by the officers below. Pending ratification at Sprint 0 (see LS-REG-001).

Document Maintenance Log