# CAIN-2027 Evolution 107: the organism runs its own experiments

Generated from `artifacts/e107/` by `scripts/evolution107/gen_report.py`. Every number below is recomputed from raw rows by `verify_e107_offline.py` (stdlib only, imports no CAIN code). **Gate counts:** {"PASS": 6, "BLOCKED": 1, "PARTIAL": 1}. Not AGI, not VAGI, not general intelligence. No external reviewer.

Lineage: E98 intelligence pulse → E99 rapid heartbeat (corrected by E102) → E102 persistent governed organism → E103–E106 → **E107**.

## 1. What changed
E106 stopped because its guardrail was inverted: it credited every cell visited since a passage was last blocked, so it saw patterns where none existed and missed the real one. E107 replaces passive credit with **intervention**. In practice mode the organism re-runs a world whose blocked passage it found in a first trial, visits **one new suspect cell at a time** (travelling only over cells it has already visited), and retests the passage after every new cell. The first test that succeeds pins the cause to one cell. A census then maps every cell the cause could have been in, so the statistical null is exact. In sealed incidents the learned feature becomes a subgoal, and once that region has been visited the organism concludes the precondition is satisfied and retests blocked passages immediately.

## 2. Sealed results (preregistered; development findings disclosed in PREREGISTRATION.json)
| test | with learned precondition | without | treatment-only / control-only | p |
|---|---|---|---|---|
| edge-lock incidents (real regularity) | **91/200** [0.3875, 0.5242] | 64/200 [0.2593, 0.3875] | 27 / 0 | 1.49e-08 |
| grid key-and-door (uniform key) | 57/100 | 57/100 | 0 / 0 | 1.0 |
| other grid families | 74/80 | 74/80 | 0 / 0 | 1.0 |

Edge-lock steps: fewer in 80 incidents, more in 0.

## 3. The gate (what E106 got wrong)
- Edge practice (150 worlds, 84 usable experiments): **ON**, feature `['col', 0]` (the drain column, found without code that knows it), z 14.472 vs null 99% 4.518, p 0.001.
- Uniform-key practice (300 worlds, 204 usable): **OFF**, p 0.018.
- Development calibration on 380 uniform-key rows: pooled p 0.22; 1 of 19 independent batches of 20 fired.

## 4. Gates
- **H1_primary** PASS: edge-lock 91 vs 64 of 200 (treatment-only 27, control-only 0, McNemar p = 1.49e-08)
- **H1b_steps** PASS: fewer steps in 80 incidents, more in 0 (sign test p = 1.7e-24)
- **H2_gate_power** PASS: edge practice gate ON: feature ['col', 0], z 14.472 vs null q99 4.518, p 0.001, 84 worlds
- **H3_gate_guardrail** PASS: uniform-key practice gate OFF: p 0.018 (alpha 0.01), 204 worlds
- **H4_attribution** PASS: exact attribution equals the true key in 83/84 edge and 204/204 grid practice worlds (oracle check)
- **H5_no_harm** PASS: grid key-door 57 vs 57 of 100; other grid families 74 vs 74 of 80
- **external_review** BLOCKED: no external reviewer
- **breadth** PARTIAL: one regularity type (a coordinate feature) in two constructed domains

## 5. Limits (read before quoting)
- The uniform-key gate was OFF at p = 0.018, closer to alpha 0.01 than the development estimate, with its top feature also column 0. A likely cause is selection: 65/300 grid worlds were dropped because the key was collected before the passage was ever blocked, so keys on the direct route are under-represented. It did not fire, but the margin is thin; the next version must model or remove that selection.
- Edge practice used 84/150 worlds (trial 1 unsolved in 42); the census was exact in 50/84 (lag in the service domain leaves some cells unmapped).
- The hypothesis class (one cell, row or column) is designed; what is learned is which feature, from the organism's own experiments. Two constructed domains; not SWE-bench or any external benchmark.
- **Post-hoc finding (by the offline verifier):** in 1 edge practice world (`edge-102-13000011`) trial 1 named the wrong door, so the experiment attributed the wrong cell with a one-cell support. It stays in the preregistered analysis; without it the edge gate is still ON for column 0 (re-derived by the verifier).
- Same-project testing; the evidence key is self-generated; no external reviewer. Not AGI or VAGI.

Team: Vincent Noriega (founder) and a three-person team.
