Ethics

Source: ETHICS.md at the repository root is the canonical statement. This page summarises the five commitments and points at the governance machinery that maintains them.

NeuroDock’s clinical guardrails layer prevents AI from amplifying rumination, hyperfocus, and sycophancy. Those interventions sit on top of five ethics principles that govern the whole project.

Commitment

We never track silently. Every signal stays on the user’s machine. There is no remote aggregation, anonymised or otherwise.

The five principles

1. We never claim to treat any condition. NeuroDock is infrastructure. It is not therapy, not medical advice, not a substitute for clinical care.
2. We never block a user’s action without their pre-configured consent. Guardrails surface signals and slow things down; they do not refuse. Every detector is overridable.
3. Detection heuristics are public and auditable. Anyone can read the source, the thresholds, and the test fixtures. There is no “AI safety” black box that the user cannot inspect.
4. We do not aggregate detection events into population-level data. Every signal stays on the user’s machine.
5. False positives are sometimes harmful; we log when surfaced and adjust with the clinical advisory board. A nudge that fires when it should not be silent costs trust. The clinical advisory board reviews thresholds quarterly.

Commitment

Heuristics are public and auditable. Source, thresholds, and test fixtures all live in the repo. There is no “AI safety” black box.

Self-identification, never diagnosis

Commitment

Self-ID is the ultimate authority. Diagnosis is never asked for and never required.

The profile declares identity.neurotypes as a self-identified list. Self-ID is sufficient for activating any skill or guardrail. Diagnosis is never asked for and never required. See ADR 0004 for the schema rationale.

Clinical advisory board model

A five-person rotating clinical advisory board (two-year staggered terms) reviews:

• Detector heuristics and thresholds.
• Field-study findings before guardrails leave beta.
• The published guidance on ND-safe AI use.

The board advises. The maintainer decides. See governance.

What you can expect from this project

Commitment

Refuse where appropriate. AI that fuels rumination, hyperfocus, or anxiety is a regression, not a feature. The clinical layer slows things down on purpose.

• Skills will refuse to validate closed decisions when the OCD finalizer fires. You will see why.
• Hyperfocus nudges will quote your own prior commitment back to you verbatim. Not a generic warning.
• The sycophancy guard will sometimes interrupt agreement with a counter-prompt. You can override it; the override is logged locally only.

If any of this stops feeling honest, it is a bug. File it.