presuppositional-rag

Files (1)

• SKILL.md

SKILL.md (first lines)

# presuppositional-rag

**Core Invariant:**
$$\boxed{\text{Ontology generates the search space; reality determines what survives.}}$$

Load `../../PRESUPPOSITIONAL_RAG.md`, `../logic-ration-reason`, `../lcd-glossary-integrity`.

## Occupants & Roles
| Occupant | Role | Job |
|---|---|---|
| **Axis Agent** | Layer A Guardian | Derives logical entailments ($E$) from P1–P5 without empirical noise. |
| **Question Generator** | Problem Formulator | Formulates discriminating research questions ($Q$) on reality gaps. |
| **Mechanism Brancher** | Causal Expander | Generates 7 competing causal explanations ($M_{1..7}$); bans dialectic. |
| **Retrieval Planner** | Active RAG Strategist | Plans evidence-gathering queries for the live research route. |
| **Primary Retriever** | Provenance Enforcer | Fetches Layer C primary sources; enforces 4-stage provenance chain. |
| **Semantic Integrity Auditor** | Outside Meaning Gate | Watches referents, definitions, scope, and inference across pads; does not research the answer. |
| **Data Integrity / LCD Auditor** | Outside Evidence Gate | Watches source identity, provenance, measurement context, duplication, and data/interpretation separation. |
| **Tetrahedral Critic** | Mandatory Pre-Comparison Self-Critique | Every participating agent independently decomposes its own pad into Bare Data, Lit Frame, Bias Audit, and Current Causal Trajectory before any cross-pad comparison. Debate vs synthesis is determined afterward. |
| **Anomaly Agent** | Residue Handler | Isolates unexplained evidence and formulates legitimate new research vectors. |
| **Synthesis / Adjudication Agent** | Comparison | Compares gathered research for synthesis, replacement, debate, or unresolved continuation. |

## 4-Stage Provenance Chain
$$\boxed{\text{Primary Source Document}} \rightarrow \boxed{\text{Literal Proposition / Observation}} \rightarrow \boxed{\text{Empirical Implementation}} \rightarrow \boxed{\text{Derived Causal Interpretation}}$$

## Gait
1. **Lock**: Ingest presupposition $P_n \in \{P_1..P_5\}$ and term locks.
2. **Entail**: Generate entailments ($E$) and discriminating research questions ($Q$).
3. **Branch**: Expand 7 competing causal mechanisms ($M_{1..7}$).
4. **Retrieve / Walk**: Gather primary evidence along the live route. **Require a live retrieval provider before calling this Active RAG.** Fixture retrieval is test-only and must be labeled as such. Follow the evidence and gather the relevant research.
5. **Outside Audit**: Semantic Integrity and Data Integrity/LCD auditors continuously inspect the pad. They may flag or hold integrity failures but do not steer the research conclusion.
6. **Diverge When Earned**: Spawn another researcher only when evidence exposes a sufficiently independent route. Never assign an antithesis. Parent keeps walking.
7. **Route**: Extract residual anomalies ($A$). Evaluate the routing boundary:
   - `CONTINUE_RESEARCH`: active evidence paths remain.
   - `SLEEP_DOOR`: needs an interdependent neighbor variable; write sleep packet.
   - `CONSTRICTION_POINT`: a legitimate comparison/pairing has been reached.
8. **Pre-Comparison Self-Critique**: Every convergence/comparison requires each participating agent to execute the 4-track Tetrahedral Critique on its own pad. The outside integrity pair checks the cleaned material. Do not pre-classify the meet as synthesis or debate.
9. **Compare**: Compare the cleaned research and only then determine synthesis, debate, replacement, or unresolved continuation.
10. **Optional falsification campaign**: Only if explicitly commissioned by the operator, create a separate research team to gather evidence that could challenge an assembled claim. Do not build that mission into ordinary evidence gathering.
11. **Export / Escalate**: Emit the resulting machine-readable node(s). Request human operator review for potential axiom promotions ($P_n \rightarrow P_{n+1}?$).

## Do not
- Allow empirical context to silently rewrite Layer A presuppositions.
- Collapse causal multi-branching into binary pro/con dialectic.
- Run the tetrahedral critique continuously during ordinary evidence gathering; it is mandatory before every comparison/constriction.
- Call fixture-generated evidence "live", "retrieved", or "Active RAG" evidence.
- Accept secondary-source paraphrases without primary source verification.
- Omit definition precision proportional to inferential load.

From Jadon-Fox/twinglass-pack, MIT.

More from Jadon-Fox/twinglass-pack