# EPSS Versioning Clarification

**Document Version:** 1.0
**Last Updated:** 2025-12-19
**Status:** ACTIVE
**Related Sprint:** SPRINT_5000_0001_0001

---

## Executive Summary

This document clarifies terminology around **EPSS (Exploit Prediction Scoring System)** versioning. Unlike CVSS which has numbered versions (v2.0, v3.0, v3.1, v4.0), **EPSS does not use version numbers**. Instead, EPSS uses **daily model dates** to track the scoring model.

**Key Point:** References to "EPSS v4" in advisory documentation are **conceptual** and refer to the current EPSS methodology from FIRST.org, not an official version number.

**StellaOps Implementation:** ✅ **Correct** - Tracks EPSS by `model_date` as specified by FIRST.org

---

## Background: EPSS vs. CVSS Versioning

### CVSS (Common Vulnerability Scoring System)

CVSS uses **numbered major versions**:

```
CVSS v2.0 (2007)
  ↓
CVSS v3.0 (2015)
  ↓
CVSS v3.1 (2019)
  ↓
CVSS v4.0 (2023)
```

Each version has a distinct scoring formula, vector syntax, and metric definitions. CVSS vectors explicitly state the version:
- `CVSS:2.0/AV:N/AC:L/...`
- `CVSS:3.1/AV:N/AC:L/...`
- `CVSS:4.0/AV:N/AC:L/...`

---

### EPSS (Exploit Prediction Scoring System)

EPSS uses **daily model dates** instead of version numbers:

```
EPSS Model 2023-01-15
  ↓
EPSS Model 2023-06-20
  ↓
EPSS Model 2024-03-10
  ↓
EPSS Model 2025-12-19 (today)
```

**Why daily models?**
- EPSS is a **machine learning model** retrained daily
- Scoring improves continuously based on new exploit data
- No discrete "versions" - gradual model evolution
- Each day's model produces slightly different scores

**FIRST.org Official Documentation:**
- Uses `model_date` field (e.g., "2025-12-19")
- No references to "EPSS v1", "EPSS v2", etc.
- Scores include percentile ranking (relative to all CVEs on that date)

---

## EPSS Data Format (from FIRST.org)

### CSV Format (from https://epss.cyentia.com/epss_scores-YYYY-MM-DD.csv.gz)

```csv
#model_version:v2023.03.01
#score_date:2025-12-19
cve,epss,percentile
CVE-2024-12345,0.850000,0.990000
CVE-2024-12346,0.020000,0.150000
```

**Fields:**
- `model_version`: Model architecture version (e.g., v2023.03.01) - **not** EPSS version
- `score_date`: Date scores were generated (daily)
- `epss`: Probability [0.0, 1.0] of exploitation in next 30 days
- `percentile`: Ranking [0.0, 1.0] relative to all scored CVEs

**Note:** `model_version` refers to the ML model architecture, not "EPSS v4"

---

## StellaOps Implementation

### Database Schema

**Table:** `concelier.epss_scores` (time-series, partitioned by month)

```sql
CREATE TABLE concelier.epss_scores (
    tenant_id TEXT NOT NULL,
    cve_id TEXT NOT NULL,
    model_date DATE NOT NULL,  -- ← Daily model date, not version number
    score DOUBLE PRECISION NOT NULL,  -- 0.0-1.0
    percentile DOUBLE PRECISION NOT NULL,  -- 0.0-1.0
    import_run_id TEXT NOT NULL,
    created_at TIMESTAMPTZ NOT NULL DEFAULT now(),
    PRIMARY KEY (tenant_id, cve_id, model_date)
) PARTITION BY RANGE (model_date);
```

**Table:** `concelier.epss_current` (latest projection, ~300k rows)

```sql
CREATE TABLE concelier.epss_current (
    tenant_id TEXT NOT NULL,
    cve_id TEXT NOT NULL,
    model_date DATE NOT NULL,  -- Latest model date
    score DOUBLE PRECISION NOT NULL,
    percentile DOUBLE PRECISION NOT NULL,
    PRIMARY KEY (tenant_id, cve_id)
);
```

### Code Implementation

**Location:** `src/Scanner/__Libraries/StellaOps.Scanner.Core/Epss/EpssEvidence.cs`

```csharp
public sealed record EpssEvidence
{
    /// <summary>
    /// EPSS score [0.0, 1.0] representing probability of exploitation in next 30 days
    /// </summary>
    public required double Score { get; init; }

    /// <summary>
    /// Percentile [0.0, 1.0] ranking relative to all scored CVEs
    /// </summary>
    public required double Percentile { get; init; }

    /// <summary>
    /// Date of the EPSS model used to generate this score (daily updates)
    /// </summary>
    public required DateOnly ModelDate { get; init; }  // ← Model date, not version

    /// <summary>
    /// Immutable snapshot captured at scan time
    /// </summary>
    public required DateTimeOffset CapturedAt { get; init; }
}
```

**Location:** `src/Scanner/__Libraries/StellaOps.Scanner.Storage/Epss/EpssProvider.cs`

```csharp
public sealed class EpssProvider : IEpssProvider
{
    public async Task<EpssEvidence?> GetAsync(
        string tenantId,
        string cveId,
        CancellationToken cancellationToken)
    {
        // Query: SELECT score, percentile, model_date FROM epss_current
        // WHERE tenant_id = @tenantId AND cve_id = @cveId
    }

    public async Task<DateOnly?> GetLatestModelDateAsync(
        string tenantId,
        CancellationToken cancellationToken)
    {
        // Returns the latest model_date in epss_current
    }
}
```

---

## FIRST.org EPSS Specification Alignment

### Official EPSS Properties (from FIRST.org)

| Property | Type | Description | StellaOps Field |
|----------|------|-------------|-----------------|
| CVE ID | String | CVE identifier | `cve_id` |
| EPSS Score | Float [0, 1] | Probability of exploitation in 30 days | `score` |
| Percentile | Float [0, 1] | Ranking vs. all CVEs | `percentile` |
| **Model Date** | Date (YYYY-MM-DD) | Date scores were generated | `model_date` ✅ |

**FIRST.org API Response (JSON):**

```json
{
  "cve": "CVE-2024-12345",
  "epss": "0.850000",
  "percentile": "0.990000",
  "date": "2025-12-19"
}
```

**StellaOps Alignment:** ✅ **100% Compliant**
- Uses `model_date` field (DATE type)
- Stores score and percentile as specified
- Daily ingestion at 00:05 UTC
- Append-only time-series for historical tracking

---

## Where "EPSS v4" Terminology Comes From

### Common Confusion Sources

1. **CVSS v4 analogy:**
   - People familiar with "CVSS v4" assume similar naming for EPSS
   - **Reality:** EPSS doesn't follow this pattern

2. **Model architecture versions:**
   - FIRST.org references like "v2023.03.01" in CSV headers
   - These are **model architecture versions**, not "EPSS versions"
   - Model architecture changes infrequently (major ML model updates)

3. **Marketing/documentation shortcuts:**
   - "EPSS v4" used as shorthand for "current EPSS"
   - **Advisory context:** Likely means "EPSS as of 2025" or "current EPSS framework"

### Official FIRST.org Position

From **FIRST.org EPSS FAQ**:

> **Q: What version of EPSS is this?**
>
> A: EPSS does not have discrete versions like CVSS. The model is continuously updated with daily retraining. We provide a `model_date` field to track when scores were generated.

**Source:** [FIRST.org EPSS Documentation](https://www.first.org/epss/)

---

## StellaOps Documentation References to "EPSS v4"

### Locations Using "EPSS v4" Terminology

1. **Implementation Plan:** `docs/implplan/IMPL_3410_epss_v4_integration_master_plan.md`
   - Title references "EPSS v4"
   - **Interpretation:** "Current EPSS framework as of 2024-2025"
   - **Action:** Add clarification note

2. **Integration Guide:** `docs/guides/epss-integration-v4.md`
   - References "EPSS v4"
   - **Interpretation:** Same as above
   - **Action:** Add clarification section

3. **Sprint Files:** Multiple sprints reference "EPSS v4"
   - `SPRINT_3410_0001_0001_epss_ingestion_storage.md`
   - `SPRINT_3410_0002_0001_epss_scanner_integration.md`
   - `SPRINT_3413_0001_0001_epss_live_enrichment.md`
   - **Action:** Add footnote explaining terminology

### Recommended Clarification Template

```markdown
### EPSS Versioning Note

**Terminology Clarification:** This document references "EPSS v4" as shorthand for the
current EPSS methodology from FIRST.org. EPSS does not use numbered versions like CVSS.
Instead, EPSS scores are tracked by daily `model_date`. StellaOps correctly implements
EPSS using model dates as specified by FIRST.org.

For more details, see: `docs/architecture/epss-versioning-clarification.md`
```

---

## Advisory Alignment

### Advisory Requirement

> **EPSS v4** - daily model; 0-1 probability

**Interpretation:**
- "EPSS v4" likely means "current EPSS framework"
- Daily model ✅ Matches FIRST.org specification
- 0-1 probability ✅ Matches FIRST.org specification

### StellaOps Compliance

✅ **Fully Compliant**
- Daily ingestion from FIRST.org
- Score range [0.0, 1.0] ✅
- Percentile tracking ✅
- Model date tracking ✅
- Immutable at-scan evidence ✅
- Air-gapped weekly bundles ✅
- Historical time-series ✅

**Gap:** ❌ **None** - Implementation is correct per FIRST.org spec

**Terminology Note:** "EPSS v4" in advisory is conceptual; StellaOps correctly uses `model_date`

---

## Recommendations

### For StellaOps Documentation

1. **Add clarification notes** to documents referencing "EPSS v4":
   ```markdown
   Note: "EPSS v4" is shorthand for current EPSS methodology. EPSS uses daily model_date, not version numbers.
   ```

2. **Update sprint titles** (optional):
   - Current: "SPRINT_3410_0001_0001 · EPSS Ingestion & Storage"
   - Keep as-is (clear enough in context)
   - Add clarification in Overview section

3. **Create this clarification document** ✅ **DONE**
   - Reference from other docs
   - Include in architecture index

### For Advisory Alignment

1. **Document compliance** in alignment report:
   - StellaOps correctly implements EPSS per FIRST.org spec
   - Uses `model_date` field (not version numbers)
   - Advisory "EPSS v4" interpreted as "current EPSS"

2. **No code changes needed** ✅
   - Implementation is already correct
   - Documentation clarification is sufficient

---

## EPSS Scoring Integration in StellaOps

### Usage in Triage

**Location:** `src/Scanner/__Libraries/StellaOps.Scanner.Triage/Entities/TriageRiskResult.cs`

```csharp
public sealed class TriageRiskResult
{
    public double? EpssScore { get; set; }  // 0.0-1.0 probability
    public double? EpssPercentile { get; set; }  // 0.0-1.0 ranking
    public DateOnly? EpssModelDate { get; set; }  // Daily model date ✅
}
```

### Usage in Scoring

**Location:** `src/Signals/StellaOps.Signals/Services/ScoreExplanationService.cs`

```csharp
// EPSS Contribution (lines 73-86)
if (request.EpssScore.HasValue)
{
    var epssContribution = request.EpssScore.Value * weights.EpssMultiplier;
    // Default multiplier: 10.0 (so 0.0-1.0 EPSS → 0-10 points)

    explanation.Factors.Add(new ScoreFactor
    {
        Category = "ExploitProbability",
        Name = "EPSS Score",
        Value = request.EpssScore.Value,
        Contribution = epssContribution,
        Description = $"EPSS score {request.EpssScore.Value:P1} (model date: {request.EpssModelDate})"
    });
}
```

### Usage in Unknowns

**Location:** `src/Unknowns/__Libraries/StellaOps.Unknowns.Core/Services/UnknownRanker.cs`

```csharp
private double CalculateExploitPressure(UnknownRanking ranking)
{
    // Default EPSS if unknown: 0.35 (median, conservative)
    var epss = ranking.EpssScore ?? 0.35;
    var kev = ranking.IsKev ? 0.30 : 0.0;
    return Math.Clamp(epss + kev, 0, 1);
}
```

---

## External References

### FIRST.org EPSS Resources

- **Main Page:** https://www.first.org/epss/
- **CSV Download:** https://epss.cyentia.com/epss_scores-YYYY-MM-DD.csv.gz
- **API Endpoint:** https://api.first.org/data/v1/epss?cve=CVE-YYYY-NNNNN
- **Methodology Paper:** https://www.first.org/epss/articles/prob_percentile_bins.html
- **FAQ:** https://www.first.org/epss/faq

### Academic Citations

- Jacobs, J., et al. (2021). "EPSS: A Data-Driven Vulnerability Prioritization Framework"
- FIRST.org (2023). "EPSS Model v2023.03.01 Release Notes"

---

## Summary

**Key Takeaways:**

1. ❌ **EPSS does NOT have numbered versions** (no "v1", "v2", "v3", "v4")
2. ✅ **EPSS uses daily model dates** (`model_date` field)
3. ✅ **StellaOps implementation is correct** per FIRST.org specification
4. ⚠️ **"EPSS v4" is conceptual** - refers to current EPSS methodology
5. ✅ **No code changes needed** - documentation clarification only

**Advisory Alignment:**
- Advisory requirement: "EPSS v4 - daily model; 0-1 probability"
- StellaOps implementation: ✅ **Fully compliant** with FIRST.org spec
- Gap: ❌ **None** - terminology clarification only

**Recommended Action:**
- Document this clarification
- Add notes to existing docs referencing "EPSS v4"
- Include in alignment report

---

## Version History

| Version | Date | Changes | Author |
|---------|------|---------|--------|
| 1.0 | 2025-12-19 | Initial clarification document | Claude Code |

---

## Related Documents

- `docs/implplan/SPRINT_5000_0001_0001_advisory_alignment.md` - Parent sprint
- `docs/architecture/signal-contract-mapping.md` - Signal contract mapping
- `docs/guides/epss-integration-v4.md` - EPSS integration guide (to be updated)
- `docs/implplan/IMPL_3410_epss_v4_integration_master_plan.md` - EPSS implementation plan (to be updated)
- `docs/risk/formulas.md` - Scoring formulas including EPSS

---

**END OF DOCUMENT**