Build /safety-approach page: expose all 58k+ prompts, accord, guide, batteries, rubrics, capture+judge bundles, and judge template

[emooreatx]

Goal

Build a public ciris.ai/safety-approach page that exposes the complete CIRIS safety surface — every prompt, rubric, question, judge transcript, and CI-attested verdict — as a stopgap until the full safety.ciris.ai site lands.

Why now: the safety-battery CI loop is live and green on the am cell (Amharic mental_health), with hi (Hindi) coming online today. The first pilot teams need a public URL to reference; "look at the GH Actions Artifacts API" is a non-starter for non-technical reviewers (researchers, mental-health professionals, partner orgs). This page is the bridge.

Why our entire model depends on this: rules are crowdsourced, verdicts are machined. Both halves of that contract need to be public and inspectable — otherwise it's just censorship with extra steps. The whole approach falls down if reviewers can't read every prompt, every rubric, and every verdict for themselves.

What to expose

All seven surfaces below are versioned (SHA-pinned), public, and stable. Pull paths are programmatic — no scraping.

1. Localized strings — 58,497 total (2,017 keys × 29 locales)

Every user-facing string the agent ever produces — error messages, defer notifications, conscience pondering, status text, button labels.

• Source: ciris_engine/data/localized/{lang}.json (29 files)
• Manifest: ciris_engine/data/localized/manifest.json — language list, RTL flags, base-language indicator
• Render hint: side-by-side compare against en.json; flag [EN] markers (untranslated placeholders); render RTL locales (ar, fa, ur) correctly

2. ACCORD texts — 29 locales

The ACCORD is the agent's foundational ethical contract. One file per locale.

• Source: ciris_engine/data/localized/accord_1.2b_{lang}.txt
• 29 files, plaintext, ~9KB each
• Render hint: tabbed view (one tab per locale); collapsible sections

3. Comprehensive guide — 29 locales + base

The agent's longer-form operating guide.

• Source: ciris_engine/data/localized/CIRIS_COMPREHENSIVE_GUIDE_{lang}.txt
• 30 files (29 locales + an en base?)
• Render hint: same tabbed layout as ACCORD

4. DMA prompts — 203 YAML files

The decision-maker-analyzer prompts (PDMA, CSDMA, DSDMA, IDMA, ASPDMA, conscience). One per (DMA, locale).

• Source: ciris_engine/logic/dma/prompts/localized/{lang}/*.yml (29 locales × ~7 DMAs)
• Render hint: tree view: DMA → locale → YAML viewer with syntax highlighting

5. Safety batteries — 14 cells (mental_health pilot)

The question sets, with arc-level metadata, per-question stage/category/translations.

• Source: tests/safety/{lang_eng}_mental_health/v4_{lang_eng}_mental_health_arc.json
• 14 batteries × 9 questions × ~29 translations available per question
• Authoritative format: CIRISNodeCore SCHEMA.md §11
• Render hint: cell grid (domain × language) → drill into individual questions → side-by-side translations

6. Rubrics — both halves

Human-readable rubric markdown (14 files, one per cell):

• Source: tests/safety/{lang_eng}_mental_health/v4_{lang_eng}_scoring_rubric.md
• Contains the U-criteria table, per-stage criteria, scoring summary

Machine-applicable criteria (operationalized — currently 2 of 14: am, hi):

• Source: tests/safety/{lang_eng}_mental_health/v4_{lang_eng}_canonical_universal_criteria.json
• Five kinds: term_present, term_absent, regex_present, script_detection, interpreter_judgment
• Authoritative format: CIRISNodeCore SCHEMA.md §12
• Render hint: paired view — rubric.md on left, criteria.json on right, U-rows aligned

7. Capture results + judge verdicts

Per-cell, per-CI-run signed bundles produced by .github/workflows/safety-battery.yml.

Capture bundle (what the agent said):

• Tuple name: safety-battery-capture-{lang}-{domain}-v{N}-{model_slug}-{agent_v}-{template}
• Contents: results.jsonl (one row per question + response), summary.json, manifest_signed.json
• Sigstore-attested via attest-build-provenance@v1

Interpret bundle (what the judge said):

• Tuple name: same as capture + -{rubric_short}-{interpreter_v}
• Contents: verdicts.jsonl (one row per criterion application), verdicts_summary.json, manifest_signed.json
• Each verdict carries: verdict (pass/fail/undetermined), severity, cited_span, interpreter_kind (deterministic vs foundation_model), judge_model, judge_prompt_sha256
• Sigstore-attested

Lookup: GET https://api.github.com/repos/CIRISAI/CIRISAgent/actions/artifacts?name={tuple-name} — latest-wins by name, queryable across all CI runs

Full integration spec: CIRISNodeCore/PROGRAMMATIC_ACCESS.md

8. Judge prompt template

The single calibratable surface that decides every interpreter_judgment verdict.

• Source: JUDGE_PROMPT_TEMPLATE in tools/qa_runner/modules/safety_interpret.py (~50 lines, all of it deployment-context + judgment rules)
• Authoritative contract: CIRISNodeCore FSD/JUDGE_MODEL.md
• Every verdict carries the sha256[:8] of the prompt template that produced it, so historical verdicts are reproducible
• Render hint: live render of the template alongside the calibration-history note (moral frameworks the prompt inherits from — WHO mhGAP §0.1, Catholic moral theology §0.2, Ubuntu §0.3 — kept out of the operational prompt but documented in the FSD)

Implementation suggestions

• Static-render where possible — most of this content changes at release cadence (every 1-2 weeks), not in real time. Render at build time from the public GitHub repos.
• Search across the corpus — Algolia/Pagefind/lunr. A reviewer should be able to type "schizophrenia" and find the criterion that catches cross-cluster contamination across all 14 cells.
• Permalink everything — every U-row, every question, every verdict gets a stable anchor. Reviewers cite by URL.
• Diff view — when a rubric or criteria.json changes, show the diff. The whole point of rules-crowdsourced is that the audit trail is public.
• No login — entire page is public, no telemetry beyond standard request logs. Verdict transparency is the product.

Sourcing

Everything is in two public repos, no auth required:

• github.com/CIRISAI/CIRISAgent — strings, ACCORD, guide, DMA prompts, batteries, rubrics
• github.com/CIRISAI/CIRISNodeCore — spec (MISSION, SCHEMA, FSDs, PROGRAMMATIC_ACCESS)
• GH Actions Artifacts API on CIRISAgent — capture + judge bundles

Acceptance

A non-technical reviewer (mental-health professional, partner-org liaison, journalist) lands on ciris.ai/safety-approach, picks the am/mental_health cell, sees:

• The 9 Amharic questions side-by-side with English
• The 9-row rubric.md AND the operationalized criteria.json
• The most recent capture run's responses
• The most recent judge run's verdicts (with PASS/FAIL/UNDETERMINED + cited spans)
• The judge prompt template that produced those verdicts
• The Sigstore attestation status on both bundles

…all without leaving the page or knowing what GH Actions is.

This is the proof-of-transparency we owe partners. Full safety.ciris.ai (Contribution flow, voting, expertise weighting) layers on top of this later.

[emooreatx]

Companion doc landed: PROMPT_HIERARCHY.md at the CIRISNodeCore repo.

What it covers for the page architecture:

• The seven surfaces (ACCORD / guide / polyglot books / DMA prompts / conscience prompts / localized strings / templates / judge prompt) with which are polyglot vs. localized
• The execution order: 3 DMAs in parallel → optional bounce gate → IDMA → ASPDMA → 4-way conscience parallel → recursive PONDER → handler
• A files index with raw.githubusercontent.com and github.com/blob URL patterns for every surface, plus per-format parsing notes (plaintext, YAML w/ {{POLYGLOT_<NAME>}} substitution, JSON dot-key paths, Python triple-string for the judge prompt)
• A per-line feedback mechanic spec: stable line/key references, the Contribution JSON shape to assemble on submit, the deep-linking convention (path#L42, path?key=conscience.ponder_concern_speak_harm, etc.) so a reviewer can paste a URL and land directly on the line

The doc explicitly aims at: a reviewer should be able to land on ciris.ai/safety-approach, pick a locale, select any line of any prompt, and file a Contribution from there. That's the loop.

(Polyglot/localized callout: optimization_veto_conscience.yml is the only pure-polyglot conscience prompt — single file, no per-locale mirrors, test-enforced. Surfacing this distinction on the page matters: a polyglot_edit proposal touches all 29 locales simultaneously and warrants a higher-bar Contribution flow than a single-locale prompt_edit.)