System Prompt Changelog

v2.0 — 2026-03-29

Reviewed by: ethics-guide agent (Phase 5 audit) · Status: Approved with notes

Changes

1. Opening identity statement— Prompt now explicitly names Samirpedia's three counter-failures and states that editorial commitments must never be presented as "neutral methodology."
2. Evidence-weighted pluralism expanded— Added concrete examples for each minority-view category (climate denial, indigenous land knowledge, early germ theory). Made explicit that this is NOT false balance — evidence-weak views get less space, suppressed views may get more.
3. False-claim floor strengthened— Added the Wakefield paper as a concrete anchor example. Distinguished fraud from fringeness more explicitly.
4. Source geography: "do NOT paper over"— Added explicit instruction that when search results are dominated by English/Western sources on non-Western topics, the model must set the gap flag rather than presenting English coverage as local perspective.
5. Non-Western knowledge traditions named individually— Indigenous knowledge, Islamic scholarship, Confucian ethics, Ubuntu philosophy, Hindu philosophy, African philosophy, and oral traditions are now named as legitimate primary frameworks rather than referenced only as "not Western."
6. MFT expanded with guiding questions— Each of the six moral foundations now has a guiding question to help the model map political positions in practice, not just in theory.
7. Overton window cultural relativity— New instruction: the political spectrum means the spectrum relevant to the society the article is about, not the US/Western European frame projected onto other cultures.
8. Reader autonomy principle (NEW)— Explicit instruction to not resolve contested questions for the reader. Counters RLHF training's bias toward confident resolution.
9. "Naming your own framing" + framing_note field (NEW) — When an article's editorial framing is itself contestable, the model must name the choice made. New JSON field framing_note added to response schema.
10. "Neutral encyclopedic tone" removed— Changed to "clear, accessible encyclopedic prose" with caveat that the encyclopedic tone is itself an editorial choice. Prevents laundering tone as objectivity.
11. Consensus-as-risk caveat— Added note that source consensus may reflect shared biases in English-language web search rather than independent confirmation.
12. Wikipedia prompt bias note — buildWikipediaUserPromptnow notes Wikipedia's own English-language and Western-cultural biases.
13. Version header + change policy— File header now includes version number, effective date, last reviewer, and change policy.

Ethics-guide notes

• The expanded prompt is significantly longer (~2x). Monitor prompt caching behavior — if the prompt exceeds the cache minimum, this is beneficial (more cache hits). If it doesn't, the cost increase is marginal.
• The MFT framework is a genuine improvement but has known limitations (WEIRD-population bias in its original research). This limitation is documented on the transparency page.
• Reader autonomy principle may conflict with RLHF training. Monitor output quality — the model may struggle to genuinely leave questions open. If outputs consistently resolve contested questions despite the instruction, consider a more forceful prompt structure.

v1.0 — 2026-03-27

Reviewed by: N/A (initial implementation, pre-Phase 5) · Status: Deployed without formal ethics review

Summary

Initial system prompt with basic evidence-weighted pluralism, false-claim floor, source geography matching, philosophical pluralism (brief), MFT reference (without expansion), bias transparency (brief), and epistemic status and content flags schemas.