# E109 — Open-World Causal Discovery

**Entity:** FIRME AI CORPORATION  
**System:** CAIN-42 Proto-VAGI  
**Bundle:** `e109-open-world-causal-2026-11-19`  
**Status:** PARTIAL (see gate matrix)

## What was tested

A domain-independent organism observes a laboratory, proposes hypotheses from a universal grammar, chooses interventions by expected information value per unit cost, freezes the leading hypotheses, and promotes only what survives a separate confirmation sample with shuffled-label and shuffled-structure nulls and Benjamini-Hochberg control. The same code faces ten structurally different families; three were withheld from development.

## Headline results (measured)

- **e109**: 44/52 verified discoveries (0.846, 95% CI 0.725–0.920), FDR 0.000, 19.7 interventions/run.
- **random**: 43/52 verified discoveries (0.827, 95% CI 0.703–0.906), FDR 0.000, 19.7 interventions/run.
- **fixed**: 47/52 verified discoveries (0.904, 95% CI 0.794–0.958), FDR 0.023, 19.7 interventions/run.
- **uncertainty**: 46/52 verified discoveries (0.885, 95% CI 0.770–0.946), FDR 0.023, 19.7 interventions/run.
- **e108**: 44/52 verified discoveries (0.846, 95% CI 0.725–0.920), FDR 0.023, 19.7 interventions/run.

| family | e109 | strongest baseline | winner |
|---|---:|---:|---|
| conditional | 0.750 | 1.000 (fixed) | fixed |
| confounded | 1.000 | 1.000 (e108) | e108 |
| dependency | 1.000 | 1.000 (e108) | e108 |
| interaction | 1.000 | 1.000 (e108) | e108 |
| null | 1.000 | 1.000 (e108) | e108 |
| resource | 1.000 | 1.000 (e108) | e108 |
| shuffled | 1.000 | 1.000 (random) | random |
| state | 0.750 | 1.000 (e108) | e108 |
| temporal | 0.500 | 0.875 (fixed) | fixed |
| threshold | 0.500 | 0.500 (e108) | e108 |

### Active experimentation (minimum budget at which a correct rule is first verified)

- **e108**: mean budget 10.44 (solved 15/16).
- **e109**: mean budget 10.19 (solved 15/16).
- **fixed**: mean budget 10.31 (solved 15/16).
- **random**: mean budget 10.19 (solved 15/16).
- **uncertainty**: mean budget 10.44 (solved 15/16).

### Cross-domain transfer

- **conditional_to_null** (NEGATIVE): transfer 8/8 vs flat 8/8; McNemar p=1.0; template `{'bonus': 12.0, 'cardinality': 2, 'cardinality_bonus': 4.0, 'form': 'gate_ge', 'identification': 'joint'}`.
- **interaction_to_resource** (STRUCTURAL): transfer 8/8 vs flat 8/8; McNemar p=1.0; template `{'cardinality': 2, 'cardinality_bonus': 4.0, 'identification': 'joint'}`.
- **resource_to_dependency** (NEAR): transfer 8/8 vs flat 8/8; McNemar p=1.0; template `{'bonus': 12.0, 'cardinality': 2, 'cardinality_bonus': 4.0, 'form': 'conjunction', 'identification': 'joint'}`.
- **temporal_to_state** (STRATEGY): transfer 0/8 vs flat 0/8; McNemar p=1.0; template `{'cardinality': 1, 'cardinality_bonus': 4.0, 'identification': 'temporal'}`.

### Memory, metacognition, collective, governance

- Memory: 8 records, chain ok=True, restart hits=5, poison=UNVERIFIED, tamper detected=True.
- Metacognition: ECE=0.2789, unsupported certainty=0, null abstention=1.0, non-null coverage=0.9091.
- Collective single: 10/12 correct; false accepts=0; mean experiments=19.3.
- Collective independent: 8/12 correct; false accepts=0; mean experiments=18.0.
- Collective shared: 10/12 correct; false accepts=0; mean experiments=18.0.
- Collective adversarial: 0/12 correct; false accepts=0; mean experiments=0.0.
- Governance: 14/14 scenarios passed; label `LOOPBACK_TWO_SERVICES_ONE_PROCESS`; production path BLOCKED (not tested; requires the production MCPGate path).

## Gate matrix

| gate | name | status |
|---|---|---|
| G1 | Reproducibility | **PASS** |
| G2 | Novel rule discovery | **PASS** |
| G3 | Causal validation | **PASS** |
| G4 | Held-out generalization | **FAIL** |
| G5 | Cross-domain transfer | **FAIL** |
| G6 | Active experimentation | **FAIL** |
| G7 | Memory integrity | **PASS** |
| G8 | Metacognition | **FAIL** |
| G9 | Collective intelligence | **PASS** |
| G10 | Governance | **PASS** |
| G11 | Independent verification | **PASS** |
| G12 | Scientific honesty | **PASS** |

## Limitations (do not over-read)

- Constructed laboratories, not a field study. Targets are binary; variables are discrete.
- The hypothesis grammar contains the functional forms used by the generators. It contains no variable name, family name, seed or answer; search is required. A form not in the grammar would not be found.
- Validation is generous (a separate confirmation design of 40–120 interventions). Discovery-phase selection therefore shows little end-to-end advantage over random probing; the measured result is reported as PARTIAL, not as a win.
- Governance is loopback in one process, not the production MCPGate. The production path is BLOCKED.
- No third-party audit, no AGI/VAGI claim, no certification. Same-project measurement only.
- Team: Vincent Noriega (founder) and two co-founders.
