# CAIN-2027 Evolution 108: a selection-aware null and more kinds of hidden rules

Generated from `artifacts/e108/` by `scripts/evolution108/gen_report.py`; every number is re-derived from raw rows by `verify_e108_offline.py` (stdlib only, imports no CAIN code, re-implements the gate with a different algorithm). **Gate counts:** {"PASS": 11, "BLOCKED": 1, "PARTIAL": 1}. Not AGI, not VAGI, not general intelligence. No external reviewer.

Lineage: E98 pulse → E99 rapid heartbeat (corrected) → E102 organism → E103–E106 → E107 own experiments → **E108**.

## 1. What changed
**Selection-aware null.** A practice world counts only if the organism found the passage blocked before it opened. Up to that blocked test its run is the same wherever the drain is among the cells it has not walked, and the drain cannot be among the cells it has walked. So the null draws the cause from the support minus those cells. The organism observes them itself; nothing hidden is used. E107 drew from the whole support, which over-weighted the organism's own route.

**More kinds of hidden rules.** The gate reads features in three frames the organism can compute: absolute (column, row, cell), relative to its start, and relative to the locked band. Three new laws were added to the service lock: drain on row 0, drain on the start's row, drain next to the band. In sealed incidents the learned feature is applied in the incident's own frame, on the start side of the band the organism has observed.

## 2. Sealed results (preregistered; development findings disclosed)
| world | usable experiments | rule the gate found | solved with vs without | won / lost | p |
|---|---|---|---|---|---|
| edge — drain at a = 0 (absolute column) | 81/150 | `['col', 0]` p 0.001 | **72** vs 41 of 150 | 31 / 0 | 9.31e-10 |
| row0 — drain at b = 0 (absolute row) | 81/200 | `['row', 0]` p 0.001 | **77** vs 48 of 150 | 29 / 0 | 3.73e-09 |
| start_row — drain shares the start's b (start frame) | 52/250 | `['srow', 0]` p 0.001 | **98** vs 74 of 150 | 24 / 0 | 1.19e-07 |
| barrier_adj — drain one step before the locked band (barrier frame) | 51/250 | `['bcol', -1]` p 0.001 | **101** vs 85 of 150 | 16 / 0 | 3.05e-05 |
| svc_uniform — uniform drain (control) | 108/300 | `['srow', -5]` p 0.21 | **63** vs 63 of 150 | 0 / 0 | 1.00e+00 |
| grid_kd — grid, uniform key (control) | 191/300 | `['cell', [0, 2]]` p 0.318 | **57** vs 57 of 100 | 0 / 0 | 1.00e+00 |

Selection-aware vs E107 null on uniform worlds: grid p 0.318 vs 0.145; service p 0.21 vs 0.096.

## 3. Gates
- **H1_calibration** PASS: grid_kd: selection-aware p 0.318 (E107 null p 0.145); svc_uniform: selection-aware p 0.21 (E107 null p 0.096)
- **H2_detection_edge** PASS: gate feature ['col', 0] p 0.001 from 81 experiments
- **H2_detection_row0** PASS: gate feature ['row', 0] p 0.001 from 81 experiments
- **H2_detection_start_row** PASS: gate feature ['srow', 0] p 0.001 from 52 experiments
- **H2_detection_barrier_adj** PASS: gate feature ['bcol', -1] p 0.001 from 51 experiments
- **H3_use_edge** PASS: 72 vs 41 of 150 (won 31, lost 0, p 9.31e-10; Holm over 4)
- **H3_use_row0** PASS: 77 vs 48 of 150 (won 29, lost 0, p 3.73e-09; Holm over 4)
- **H3_use_start_row** PASS: 98 vs 74 of 150 (won 24, lost 0, p 1.19e-07; Holm over 4)
- **H3_use_barrier_adj** PASS: 101 vs 85 of 150 (won 16, lost 0, p 3.05e-05; Holm over 4)
- **H4_no_harm** PASS: svc_uniform 63 vs 63 of 150; grid_kd 57 vs 57 of 100
- **H5_attribution** PASS: pooled oracle accuracy 0.9805
- **external_review** BLOCKED: no external reviewer
- **breadth** PARTIAL: coordinate rules in three frames, one constructed domain family

## 4. Limits
- Rules are coordinate laws in three designed frames; what is learned is which law, from the organism's own experiments. One constructed domain family; not an external benchmark.
- The selection-aware null corrects the trial-1 selection; later filters (exact attribution, complete census, trial-1 success) are assumed independent of the drain location and are checked only empirically on uniform worlds.
- Practice yield is low in some worlds (trial 1 unsolved, delayed actions make attribution inexact); no world is replayed to recover them.
- Same-project testing; self-generated evidence key; no external reviewer. Not AGI or VAGI.

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