What Hermes Does Not Know — Epistemic Limits & Analytical Boundaries
┌─────────────────────────────────────────────────────────────────────────────────┐│ HERMES CODEX EPISTEMIC INTEGRITY CONTRACT (SECTION 83) ││ ││ "To state what is known without stating what is ignored is not analysis; ││ it is marketing. Hermes formalizes the frontier of its own ignorance." │└─────────────────────────────────────────────────────────────────────────────────┘1. The Six Dimensions of Analytical Uncertainty
Section titled “1. The Six Dimensions of Analytical Uncertainty”Under Section 83 of the Hermes Master Specification, analytical uncertainty is decomposed into six immutable, mutually exclusive dimensions:
D1 Missing Data (Données Manquantes)
Information that exists objectively in the physical world or private telemetry but has not been published or verified in accessible authoritative registries.
D2 Unavailable Sources (Sources Indisponibles)
Authoritative feeds locked behind proprietary NDA licensing, classified state sensors, or paywalled private EDR telemetry networks.
D3 Working Hypotheses (Hypothèses)
Default baseline postulates adopted to enable deterministic risk calculation prior to empirical ground confirmation.
D4 Structural Uncertainties (Incertitudes)
Intrinsic statistical dispersion, machine learning inference intervals (EPSS), and temporal ingestion lags (24–72 hours).
D5 Model Limitations (Limites des Modèles)
Mathematical and algorithmic boundaries inherent to AST parsing, non-SemVer normalization heuristics, and static pillar weightings.
D6 Unobservable Dimensions (Non Observables)
Physical and cryptographic realities inaccessible from the exterior (internal network micro-segmentation, private air-gaps, actor psychological intentions).
2. In-Depth Breakdown of the 6 Dimensions
Section titled “2. In-Depth Breakdown of the 6 Dimensions”What Constitutes Missing Data in Threat Intelligence?
Section titled “What Constitutes Missing Data in Threat Intelligence?”- NVD Ingestion Latency & Backlog: Due to national triage backlogs, weeks can elapse between public zero-day exploitation and official CVSS vector assignment.
- Closed Firmware & Proprietary Diffs: Closed-source enterprise gateways (Ivanti, Palo Alto, Fortinet, Cisco) rarely disclose exact git commit SHAs, line-level diffs, or vulnerable AST sinks.
- Unbounded Early Version Ranges: Vendor advisories frequently publish ambiguous statements such as “all versions prior to 24.2.1”, concealing whether legacy architectural branches are vulnerable or structurally immune.
Hermes Treatment:
Section titled “Hermes Treatment:”Hermes flags the dossier with epistemic_state: "INSUFFICIENT_DATA". Rather than guessing an arbitrary score, Hermes bounds the affected range conservatively and reduces the confidence metric (C) with a distance penalty (P_dist = -0.10 per missing variable).
Managing Closed and Paywalled Sensor Silos
Section titled “Managing Closed and Paywalled Sensor Silos”- Proprietary EDR/XDR Telemetry: Global sensor arrays managed by commercial vendors (CrowdStrike Falcon OverWatch, Microsoft Defender TI, SentinelOne) are legally and technically restricted to private corporate subscribers.
- Classified State Signals: National cyber defense agencies (ANSSI, CISA, BSI) possess non-public telemetry regarding nation-state pre-positioning that is excluded from open distribution.
- Darknet & Broker Repositories: Hermes strictly operates within legal, ethical open-source intelligence boundaries and does not infiltrate closed illicit broker networks requiring criminal transactions.
Hermes Treatment:
Section titled “Hermes Treatment:”Hermes aggregates verified multi-source open signals (CISA KEV, Shadowserver internet-wide scanner telemetry, CrowdSec community honeypot signals, VulnCheck weaponization feeds, and Patchstack CMS intelligence) while explicitly citing which proprietary sources are absent.
Formal Postulates Required for Computational Modeling
Section titled “Formal Postulates Required for Computational Modeling”To provide deterministic security decisions without infinite stalling, Hermes explicitly applies calibrated working hypotheses:
- Default Exposure Postulate (
H_exp): In the absence of an authenticated configuration audit, Hermes assumes the vulnerable network service is deployed in its default vendor configuration with standard listening ports exposed. - Epidemiological Weaponization Velocity (
H_vel): Once a verified unauthenticated remote code execution (RCE) proof-of-concept is documented, Hermes assumes adversary weaponization will propagate along a sub-exponential diffusion curve unless mitigated by structural barrier dependencies. - Transitivity of Vulnerable Dependencies (
H_trans): If a deeply nested package contains a vulnerable symbol, Hermes assumes exposure until call-graph reachability proves the symbol unreferenced.
Hermes Treatment:
Section titled “Hermes Treatment:”Every hypothesis is explicitly serialized as an EPISTEMIC_HYPOTHESIS object with an automated expiration trigger (T_exp <= 30 days) requiring ground-truth empirical re-verification.
Statistical Dispersion & Ingestion Time Lag
Section titled “Statistical Dispersion & Ingestion Time Lag”- EPSS Probabilistic Variance: The FIRST Exploit Prediction Scoring System (EPSS) outputs a statistical probability, not a deterministic certainty. An EPSS score of 0.85 signifies that 85% of historically identical CVEs were exploited within 30 days, not that an attack is guaranteed against a given asset.
- Telemetry Propagation Latency (24h – 72h): Between the moment an adversary executes an attack in the wild and the time network honeypots record, correlate, and publish the telemetry, a delay is inevitable.
- Automated Honeypot Noise: Indiscriminate scanner probes frequently trigger honeypot alerts without possessing working exploit payloads for modern patched target versions.
Hermes Treatment:
Section titled “Hermes Treatment:”Hermes provides confidence bands rather than deceptive single-scalar absolutes. Threat trajectories reflect rate-of-change (Δrisk / Δt) rather than static snapshots.
Algorithmic Boundaries of Automated Analysis
Section titled “Algorithmic Boundaries of Automated Analysis”- Non-SemVer Version Heuristics: Enterprise appliance versioning schemes (e.g.,
12.1.0-build-827-MR3) deviate from strict Semantic Versioning. While Hermes employs multi-branch vendor resolvers, edge-case builds may require heuristic branch pairing. - Composite Scoring Weights: The Hermes Threat Score (HTS-3.1) relies on parameterized weights across 8 structural pillars. Extreme edge cases (e.g., hardware-level transient execution vulnerabilities like Spectre or indirect prompt injection in autonomous LLM agents) challenge traditional network exploitability metrics.
- Static AST Limitations: Static SBOM vulnerability analysis cannot observe whether dynamic runtime feature flags disable the vulnerable code path in memory.
Hermes Treatment:
Section titled “Hermes Treatment:”All scoring formulas are fully open-sourced, mathematically reproducible, and audited via automated regression suites (npm test).
The Physical and Cryptographic Horizon
Section titled “The Physical and Cryptographic Horizon”Certain operational parameters can never be observed from external threat intelligence vantage points:
- Target-Side Defense-in-Depth: External observers cannot know whether an organization has placed an air-gap, applied custom eBPF kernel filters, or enforced strict Zero Trust mTLS authentication in front of a vulnerable port.
- Adversary Intent & Resource Allocation: Quantitative telemetry measures adversary capability and observed activity, but cannot read the strategic intentions or budget constraints of human threat actors.
- Private Undisclosed Vulnerabilities (True Zero-Days): Until an exploit is observed or reported, its existence remains mathematically unobservable.
Hermes Treatment:
Section titled “Hermes Treatment:”Hermes restricts its scoring to intrinsic vulnerability severity and observed global threat activity, reminding security teams via interactive audits that local compensatory controls must be evaluated on the target infrastructure.
3. Operational Rule: When is an Assessment “Uncomputable”?
Section titled “3. Operational Rule: When is an Assessment “Uncomputable”?”Hermes firmly rejects “fake precision.” If critical evidentiary predicates are missing, the Decision Engine refuses to fabricate a score and marks the entry as UNCOMPUTABLE:
┌──────────────────────────────────────────────┐ │ Vulnerability Ingestion Event │ └──────────────────────┬───────────────────────┘ │ Are affected versions bounded? ├── NO ──► [EPISTEMIC_STATE: INSUFFICIENT_DATA] │ HTS: UNCOMPUTABLE YES Action: Quarantine Analysis │ Is root vulnerability locus verified? ├── NO ──► Apply Penalty (C ≤ 0.45) │ Review Date: +24h YES │ Are primary sources in conflict? ├── YES ─► [EPISTEMIC_STATE: CONFLICTING] │ Document Dispute & Side with Lab NO │ ▼ [EPISTEMIC_STATE: KNOWN] Confidence Score: 0.85 – 1.00 Publish Automated AdvisoryFormal Uncomputability Criteria:
Section titled “Formal Uncomputability Criteria:”- Unbounded Version Specifier: If an advisory specifies no version bounds (e.g.,
version: "*"without CPE vendor confirmation), HTS calculation is suspended. - Conflicting Vector Discrepancy > 4.0 Points: If primary vendor advisory asserts CVSS 4.0 (Local DoS) while community telemetry observes remote arbitrary code execution (CVSS 9.8), the item is flagged as
CONFLICTINGand escalated for laboratory verification. - Zero Corroborated Telemetry with Expired PoC: If an unverified social media claim asserts zero-day exploitation without code, packet capture, or sensor hits within 14 days, it is purged as unverified noise.
4. Live Demonstration: Epistemic Limits UI Integration
Section titled “4. Live Demonstration: Epistemic Limits UI Integration”Every single observation and threat dossier on Hermes Codex embeds the Hermes Unknown Box to ensure readers instantly perceive analytical boundaries:
- ✕ Local network segmentation: whether the target organization isolates the management API behind an out-of-band VPN.
- ✕ Vendor patch availability date: the vendor has not released a maintenance release timestamp for legacy branch 2.4.x.
- ✕ Active weaponization against air-gapped instances: telemetry is derived solely from internet-exposed sensor arrays.
5. Epistemic State Machine Reference
Section titled “5. Epistemic State Machine Reference”Every artifact within the Hermes intelligence repository tracks its lifecycle through four formal states:
| Epistemic State | Mathematical Definition | Confidence Metric ($C$) | Publishing Behavior |
|---|---|---|---|
KNOWN | Primary telemetry verified; $\ge 2$ corroborating feeds; version bounds deterministic. | $0.75 \le C \le 1.00$ | Automated real-time publication. |
CONFLICTING | Incompatible claims between vendor advisory and field telemetry. | $0.40 \le C \le 0.74$ | Prominent dispute callout box displayed. |
INSUFFICIENT_DATA | Essential analytical variables missing (e.g., unbounded affected versions). | $C < 0.40$ | Score marked UNCOMPUTABLE; quarantine alert. |
UNKNOWN | Theoretical vector identified but completely unobservable in current telemetry. | N/A | Documented exclusively under What Hermes Does Not Know. |