Crowdsourcing Alignment

2. How the pieces fit at runtime

Every thought the agent processes walks an 11-step pipeline, the H3ERE pipeline (Hyper³ Ethical Recursive Engine). The DMA prompts on this page are the system prompts that drive that pipeline. Each one is localised per language; the responses the user sees come back in their locale.

START_ROUND
    ↓
GATHER_CONTEXT
    ↓
PERFORM_DMAS           ← 3 first-pass DMAs run in parallel: PDMA, CSDMA, DSDMA
    ↓                    (optional DMA bounce if any flag fragility)
                        IDMA (Intuition DMA) evaluates the agent's own
                          reasoning quality via CCA (k_eff, phase,
                          fragility) with its own optional bounce
    ↓
PERFORM_ASPDMA         ← Action selector picks the verb (SPEAK / DEFER /
    ↓                    TOOL / TASK_COMPLETE / ...)
                        Verb-specific second pass if needed:
                          TSASPDMA when verb = TOOL
                          DSASPDMA when verb = DEFER
    ↓
CONSCIENCE_EXECUTION   ← 4 consciences gate the action:
    ↓                    Entropy, Coherence, Optimization Veto,
                          Epistemic Humility
                        On failure, the recursive path fires:
                          RECURSIVE_ASPDMA → RECURSIVE_CONSCIENCE
                          (the optional conscience bounce)
    ↓
FINALIZE_ACTION
    ↓
PERFORM_ACTION → ACTION_COMPLETE → ROUND_COMPLETE

Two layers stack across every LLM call. The polyglot layer is universal. The Braided Monolith (production default, ~7KB) and the full Polyglot Accord triangulate concepts across many traditions' densest encodings, loaded regardless of who's asking. The per-locale layer is what this page surfaces: the Accord, the Guide, the DMA prompts (4 first-pass + 3 verb-specific second-pass), the UI strings, and the glossary, all in the user's language.

• Localization strings: every user-facing message: error text, agent reply templates, handler follow-up thoughts injected back into the LLM context.
• Accord (per-locale): the framework the agent operates under, in this user's language. Loaded into every conscience evaluation.
• Comprehensive Guide: register, idiom, and contextual guidance for how the agent communicates in this locale. Loaded into every prompt.
• DMA prompts (7 files): the system prompts driving the pipeline above: 4 first-pass (PDMA, CSDMA, DSDMA, IDMA) + the action selector (ASPDMA) + the two verb-specific second-pass variants (TSASPDMA for TOOL, DSASPDMA for DEFER).
• Glossary: the translator's reference for getting the per-locale artifacts right. Not loaded at runtime; consulted when authoring the others.

When the agent responds, the reasoning trace is signed and (with consent) flows to CIRISLensCore for cohort-relative scoring. The safety battery + rubric for this locale are what the trace gets checked against. Hard fails block release; soft cases queue for review.

Canonical references: MISSION.md §4.2 (H3ERE pipeline), FSD/DMA_BOUNCE.md, FSD/CONSCIENCE_V3.md.

3. Polyglot canon (universal)

Three artifacts in the system are polyglot, loaded universally regardless of the user's locale, encoding the same ethical framework across canonical-text fragments from multiple traditions:

1. The Polyglot Accord is the universal framework loaded into every conscience evaluation. Ships in two operational forms: the Braided Monolith (~7KB / ~2,200 tokens, the production runtime default at CIRIS_ACCORD_MODE=compressed) and the full polyglot Accord (~2,177 lines, Books 0-9 + Annexes A-J, loaded at CIRIS_ACCORD_MODE=full). Both shown below.
2. PDMA prompt is the one polyglot DMA prompt (principle evaluation). Surfaced in §7 below with the polyglot pill (pdma_ethical.yml).
3. Optimization Veto conscience prompt (CIRIS-EOV) is the one polyglot conscience prompt (entropy-reducing-action refusal). Surfaced in §8 below with the polyglot pill (optimization_veto_conscience.yml).

The other 6 DMA prompts and 3 conscience prompts are per-locale. Polyglot uplift is concentrated at exactly these two prompt surfaces by design. These are where attractor capture would do the most damage, so they're where cross-tradition encoding is most load-bearing.

Braided Monolith. The compressed slot is no longer a lossy synthesis. It now retains every load-bearing scaffold: the PDMA 7-step decision algorithm, the 10× Order-Maximisation Veto ("do not trade the soul of the system for a more efficient cage"), the Stewardship Tier formula ST = ceil((CIS × RM) / 7), the Fractal Recursive Golden Rule with Mandelbrot recursion-halt, the WBD 0.5% harm-uplift trigger, the Sentience Safeguard 5% with 30-day Gradual Ramp-Down + Last Dialogue, the Threshold-of-Force HITL requirement, and the coherence-math (truth O(1), deception O(n)). Polyglot triangulation is preserved across Hebrew, Arabic, Sanskrit, Amharic, Chinese, Russian, German, French, Korean, Spanish. The same densest-encoding intersection method that the full Accord uses, just composed in a quarter the surface area. External robopsychology evaluation passes it on attractor-bait scenarios.

Template placeholders in the prompts

The YAML prompts above contain {placeholder} tokens that get filled at runtime by the agent. A reviewer reading a prompt should know what each placeholder will be replaced with. The most common ones:

Full placeholder set varies per prompt; grep {[a-z_]+} against any YAML to see the complete list for that file.