# Audit Datasets — How to Verify

**Do NOT apply the same metric to every dataset. Read the context first.**

---

## 4. Cybersecurity — UNSW-NB15 Network Intrusion

**Files:** `cyber_unsw_nb15_audit.csv` (summary, 10 rows), `cyber_unsw_nb15_full_45332.csv` (complete, 45,332 rows)

**What this is:** 45,332 real network attacks from the UNSW-NB15 benchmark (Australian Centre for Cyber Security). 9 attack types: Analysis, Backdoor, DoS, Exploits, Fuzzers, Generic, Reconnaissance, Shellcode, Worms. Detection via ZH-81 mathematical detection evaluated on the UNSW-NB15 benchmark.

**Correct metric:** Recall = 100% (FN=0). No attack escaped detection. FPR of 25.2% is intentional — missing an attack is catastrophic; a false alarm is investigated.

**How to verify in Excel:**
1. Open the CSV
2. `=CONTAR.SE(D2:D10;">0")` → 0 (zero missed per attack type)
3. `=SOMA(B2:B10)` → 45,332 (matching SOMA(C2:C10))

**How to verify in Python:**
```python
import pandas as pd
df = pd.read_csv('cyber_unsw_nb15_audit.csv')
attacks = df[df.attack_type != 'TOTAL']
all_detected = (attacks.missed == 0).all()
print(f"All {attacks.samples.sum():,} attacks detected: {all_detected}")  # True
```

**What this proves:** Pure mathematical computation (CPU only) achieves 100% recall against 45,332 real network attacks. No deep learning. No GPU. 3.45MB footprint.

---

## 1. 100 Prisoners Benchmark

**Files:** `prisoners_trials_cycle.csv`, `prisoners_trials_phi_guided.csv`, `prisoners_trials_random.csv`

**What this is:** A closed-form mathematical benchmark. The 100 Prisoners Problem has an exact theoretical survival probability of **31.183%** for N=100 (formula: `1 - sum 1/k for k=51..100`).

**Correct metric:** Convergence to 31.183% within binomial standard error (~0.015% for 1M trials).

**How to verify in Excel:**
1. Open the CSV
2. Column D: `=SE(B2<=50;"SURVIVE";"EXECUTED")`
3. Column E: `=SE(C2=D2;"OK";"ERRO")`
4. `=CONTAR.SE(E:E;"ERRO")` should be **0**
5. `=CONTAR.SE(C:C;"SURVIVE")/CONTAR.VAL(C:C)` should be **~31.18%**

**How to verify in Python:**
```python
import pandas as pd
df = pd.read_csv('prisoners_trials_phi_guided.csv')
survived = df.iloc[:, 2].astype(str).str.strip().str.upper()
longest_cycle = df.iloc[:, 1].astype(int)
errors = (survived != (longest_cycle <= 50).map({True: 'SURVIVE', False: 'EXECUTED'})).sum()
print(f"Errors: {errors}")  # Must be 0
print(f"Rate: {survived.eq('SURVIVE').mean()*100:.3f}%")  # ~31.18%
```

**What this proves:** The system can identify hidden mathematical structure in combinatorial spaces. A fabrication would fail to converge to the exact theoretical value.

---

## 3. 3D Brain Landmarks — MNI152 AFIDS

**File:** `brain_landmarks_28_audit.csv`

**What this is:** 28 anatomical landmarks detected on the MNI152 brain template, compared against the AFIDS gold-standard atlas (Lau et al., 2019, Human Brain Mapping). The competing system (nnUNet) requires thousands of 3D brain images and days of GPU training. ZH-81 uses a single averaged brain template from 152 subjects.

**Correct metric:** Detection rate (28/28 = 100%) + mean error (1.16 mm — size of a grain of rice).

**How to verify in Excel:**
1. Open the CSV
2. `=CONTAR.VAL(A:A)-1` → 28 (all landmarks present)
3. `=MEDIA(B:B)` → ~1.16 mm (mean error)
4. `=CONTAR.SE(B:B;">2")` → 0 (all errors under 2mm)
5. `=CONTAR.SE(D:D;"")-1` → 0 (all landmarks have SHA-256 seals)

**How to verify in Python:**
```python
import pandas as pd
df = pd.read_csv('brain_landmarks_28_audit.csv')
print(f"Detected: {len(df)}/28")           # 28/28 → 100%
print(f"Mean error: {df.error_mm.mean():.2f}mm")  # 1.16mm
print(f"All under 2mm: {(df.error_mm < 2.0).all()}")  # True
```

**What this proves:** Mathematical computation (0.01 seconds) matches or beats deep learning (days of GPU training), with zero hallucinations and full SHA-256 audit trail per landmark.

---

## 2. Lung Cancer Detection (TCIA)

**File:** `cancer_33cases_audit.csv`

**What this is:** 33 real DICOM CT scans from The Cancer Imaging Archive (NIH/NCI). 10 cancer cases (LIDC-IDRI) + 23 healthy controls (CC-CCII, CT Colonography, Lung Phantom).

**Correct metric:** **Sensitivity = FN/(TP+FN) = 0**. This is a SCREENING system — designed to NEVER miss cancer. False positives are intentionally flagged as "REVIEW" for doctor evaluation.

**INCORRECT metric:** Overall accuracy (60.6%). Applying diagnostic accuracy to a screening system is a category error.

**How to verify in Excel:**
1. Open the CSV
2. `=CONTAR.SE.S(C:C;"CANCER";D:D;"SAUDAVEL")` should be **0** (zero cancers missed)
3. `=CONTAR.SE.S(C:C;"SAUDAVEL";D:D;"CANCER")` = 13 (healthy cases flagged for REVIEW)

**How to verify in Python:**
```python
import pandas as pd
df = pd.read_csv('cancer_33cases_audit.csv')
fn = ((df['expected'] == 'CANCER') & (df['predicted'] == 'SAUDAVEL')).sum()
print(f"FN (cancers missed): {fn}")  # Must be 0
print(f"Sensitivity: {(10-fn)/10*100:.0f}%")  # Must be 100%
```

**What this proves:** Zero cancers escape detection. The 13 false positives are REVIEW cases — the system says "suspicious, check this" rather than risking a missed cancer.

---

## Context for Evaluators

| Dataset | Type | Correct Metric | Wrong Metric |
|---------|------|---------------|--------------|
| Brain Landmarks | Anatomical detection | Detection rate + mean error | "Better than nnUNet" without numbers |
| Cybersecurity | Network intrusion | Recall (FN=0) | Overall accuracy |
| 100 Prisoners | Mathematical benchmark | Convergence to 31.183% | Any other target |
| Cancer 33 | Medical screening | Sensitivity (FN=0) | Overall accuracy |
| EEG P300 | Physiological detection | Balanced accuracy per time window | Cherry-picked best split |

**Principle:** A screening system is judged by what it MISSES, not by what it FLAGS. A fire alarm that goes off 13 times for burnt toast is annoying. A fire alarm that misses 1 real fire is deadly.

---

ZH-81 Kernel · SHA-256 Sealed · Data Publicly Downloadable
zerohallucinations.online/datasets