# Phase 4 — Stress Test Results

**Date:** 2026-03-31  
**Status:** First stress tests complete. Mixed results.

---

## Test 1: Root-type prediction (herbal, on training data)

Applied the decode table to predict root type for all 71 overlap folios.

### Single-rule predictor

| Predicted class | Hits | Total | Accuracy | Base rate |
|----------------|------|-------|----------|-----------|
| branching (via -or + single-stem) | 4 | 4 | **100%** | 52% |
| simple (via dain) | 6 | 9 | **67%** | 21% |
| anthropomorphic (via -y + comp≥4) | 6 | 14 | **43%** | 11% |
| **Overall** | **16** | **27** | **59%** | — |

**Permutation p < 0.0001** (10,000 iterations). Lift: 2.9× over chance.

### Diagnosis

- **-or → branching roots: SHARP.** 4/4 perfect. The best individual rule.
- **dain → simple roots: GOOD.** 6/9, 67%. Misses are mostly branching roots (the dominant class).
- **-y → anthropomorphic: OVER-PREDICTS.** 6/14, 43%. The -y suffix appears on 65% of folios, so `-y + comp≥4` fires too broadly. 8/14 false positives are actually branching-root plants.

### Why the Mandragora anchor still holds

The anchor (p = 0.002) requires **convergent** evidence (≥2 of 3 predictions). That convergence filters false positives: 7/8 anthro folios score ≥2/3, but only 18/63 non-anthro do (29%). The individual -y rule alone is insufficient; the bundle is diagnostic.

### Verdict

**Outcome: PARTIAL ENTRY.** The system predicts root type above chance (p < 0.0001) with variable precision across root classes. Two root types are individually predictable. The third (anthropomorphic) requires convergent multi-signal approach.

---

## Test 2: Zodiac held-out — Type A suffix distribution

### Design

Type A zodiac folios (f70v1, f71r, f71v, f72r1) were NOT used in Anchor 2 training. They have ~15 labels each and simpler ring structure.

**Prediction:** If Type A labels are in the inner ring (Type A has fewer rings), their suffix distribution should match the training inner-ring pattern (54.5% inner_type, 17.2% outer_type).

### Result

| Source | inner_type (-y/-dy/-ey) | outer_type (-al/-ar/-am) | n |
|--------|:---:|:---:|---|
| Training inner ring | 54.5% | 17.2% | 134 |
| Training outer ring | 40.2% | 35.9% | 92 |
| **Type A labels (held-out)** | **45.8%** | **30.5%** | **59** |

Type A labels are closer to the **outer ring pattern** than the inner ring pattern (distance 0.110 vs 0.220).

The inner/(inner+outer) ratio is 0.600 — between the training inner (0.760) and outer (0.528), but closer to outer.

### Interpretation

The prediction failed in its simple form. Type A labels do NOT cleanly match the inner ring pattern. Three possible explanations:

1. **Type A is not purely inner ring.** Type A folios have 3 prose lines, meaning they may have 2 figure rings (inner+outer) despite having half the labels. The labels may span both rings.

2. **Type A has a different compositional mode.** With only 15 labels vs 30, Type A may use different suffix conventions — more classificatory, less descriptive.

3. **The ring classifier is not as clean for labels as for prose.** The 64% accuracy on training data already showed this.

### Verdict

**Outcome: MISS (but informative).** The Type A suffix distribution falls between inner and outer patterns, not cleanly matching either. This suggests Type A has a mixed or different ring grammar, not a simple inner-ring assignment.

---

## Summary scorecard

| Test | Prediction | Result | p | Verdict |
|------|-----------|--------|---|---------|
| Root-type prediction | Decode table predicts root | 16/27 correct, 2.9× lift | < 0.0001 | ✅ Partial entry |
| Root: branching | -or + single-stem | 4/4 (100%) | — | ✅ Sharp |
| Root: simple | dain | 6/9 (67%) | — | ✅ Good |
| Root: anthropomorphic | -y + comp≥4 | 6/14 (43%) | — | ⚠️ Over-predicts |
| Type A zodiac | Should match inner ring | Closer to outer pattern | — | ❌ Miss |

## Core finding

The decode table has genuine predictive power (p < 0.0001 for root prediction) but varies in precision by morpheme. The system is best characterised as **probabilistic notation with variable signal strength** — some rules are sharp (branching roots: 100%), others require convergent evidence (anthropomorphic: needs bundle). The zodiac ring classifier is less clean at the label level than at the prose level, and Type A folios don't straightforwardly extend the Type B pattern.

## Revised assessment

**Outcome A (entered system): YES for herbal root prediction, PARTIAL for zodiac**

The system makes correct forward predictions above chance. It doesn't make them perfectly, and precision varies. The Mandragora convergent anchor remains valid but individual root-type rules need the convergent approach for rare classes.