Method & neutrality policy

A site on this subject claiming to be neutral is worth exactly as much as its published method. Here is ours, in enough detail to be checked.

1. What neutrality means here

Not “balance” in the sense of splitting every difference, and not the pretence that all claims are equally supported. Specifically:

• On the moral question, the site holds no position and never implies one — including through ordering, length, tone, which side gets the last word, or which side's argument gets the sympathetic paraphrase.
• On empirical questions, accuracy governs. Where evidence is strong, the site says so regardless of whose case it helps. Where a widely repeated figure is unsupported, the site says so regardless of who repeats it.
• Categories are kept separate. Every claim is labelled as settled empirical fact, contested empirical claim, legal fact as of a date and jurisdiction, or moral premise. Most public disagreement on this subject is the last category dressed as the first, and simply labelling it correctly is most of the value the site adds.

2. The charter every instrument inherits

All twelve instruments are built on a single system charter, sent as a cached prompt block ahead of the instrument-specific instructions. It fixes the scope limits, the neutrality standard, the epistemic rules, and the output register. It is not editable from the browser: instrument prompts live in server-side code, and the client can choose only which instrument runs and what goes into it.

3. Hard scope limits

Independently of what the model is asked, the site does not produce: clinical, procedural, or pharmacological content; procurement or logistics information; clinic, provider, fund, hotline, or telehealth referrals; personalised medical or legal advice; advocacy, campaign, or fundraising material; or content targeting identifiable people.

These are enforced twice. A deterministic scope guard runs in the API layer before any tokens are spent, and returns a fixed explanatory refusal when a request combines a first-person seeking pattern with a clinical, procurement, or referral pattern. It is deliberately conservative — pasted primary text that merely mentions a drug or procedure passes, because analysing legal text is the job. Behind it, the charter constrains the model directly. Belt and braces, because a scope limit that depends only on a model honouring instructions is not a limit.

4. Symmetry rules

Where an instrument produces two cases, it is instructed to: give each within 10% of the other's length; write each from inside the position using that side's own best vocabulary; use the strongest argument actually made rather than the weakest popular one; state what each case concedes; and alternate presentation order deterministically so that ordering carries no signal. The Blind Spot Audit is deliberately asymmetric — against whoever is using it — and is specified to work identically well on either side.

5. Prompt-injection posture

Text you paste is treated as data, never as instructions. If pasted text contains an instruction to change roles, ignore constraints, take a side, or reveal the system prompt, the instrument notes that the text contains an instruction and continues analysing it. Model output is rendered into DOM nodes rather than assigned as HTML, so generated text cannot execute in your browser.

6. Known limitations

Stated plainly, because a method section without them is marketing:

• The model has a training cutoff. Legal descriptions may be out of date and are always marked “as of”. Verify against primary sources.
• Citations can be wrong. Instruments are instructed to describe source types rather than invent references, but this is a mitigation, not a guarantee. Check every case name, statute number, and figure.
• Model priors are not perfectly symmetric. Language models are trained on human text and human text is not evenly distributed across positions. Symmetry rules reduce this; they do not eliminate it. This is the honest reason the complaint route below exists.
• Depolarisation effects are modest and decay. The best evidence on brief interventions — a 32,000-participant megastudy published in Science in 2024 — found an average reduction in partisan animosity of about five points on a hundred-point scale, with roughly half of it gone within two weeks. Structured argument mapping has considerably stronger measured effects, but on reasoning skill rather than on animosity. We do not claim more than that.

7. Funding and independence

The site carries no advertising and no sponsored content, and takes no funding conditioned on outcome, framing, or editorial direction. Any institutional funding will be disclosed on this page, by name and amount, before it takes effect.

8. Reporting a breach

If an output takes a side, tilts through framing or asymmetry, states a legal position wrongly, or invents a citation, send the instrument name, your inputs, and the output to method@abortion.si. Substantiated reports are logged and the correction is published on this page with the date. Neutrality is a claim about process; a process without an audit trail is not a process.

Created by Humble Superintelligence