A Standard for AI-Assisted Digital Forensics
Ten vendor-neutral principles defining how any AI should be used when it reads, reasons over, or reports on digital evidence.
The GEP is not a product. It is a vendor-neutral, tool-agnostic standard that any AI — cloud model, local model, agent, or assistant — can be measured against, so that AI-assisted findings remain truthful, traceable to source records, and backed by an auditable, tamper-evident chain, and the human investigator stays in control.
The GEP governs how an AI must behave in digital forensics and incident response (DFIR). It says nothing about which operating system, file system, or tool is analyzed — only that, whatever the AI touches, it does so under the principles below. A system "conforms to the GEP" when it can demonstrably satisfy every principle. Each one states the rule, why it matters, what compliance looks like, and how to verify it.
GEP-1 through GEP-10. A conforming implementation upholds — and can demonstrate — every one.
Rule. Every conclusion derives only from artifacts the AI actually examined — never guess, assume, or fabricate. Absence of evidence is reported as a fact, not filled in by inference.
Verify. Trace each assertion back to a source; un-sourced assertions are violations.
Rule. Every factual statement is linked to a specific source record — database:table:row, a file path + offset, an event record id, or a content hash.
Verify. Follow any cited handle and recover the exact underlying record.
Rule. Findings state exact UTC timestamps, identifiers (users, hosts, PIDs, SIDs), and full paths, organized chronologically — never "recently" or "a while ago".
Verify. Every event has an absolute timestamp and a determinable order.
Rule. Conclusions rest on multiple corroborating sources wherever possible. Corroboration, silence, and conflict are all reported; single-source claims are flagged as such.
Verify. For a key finding, confirm the supporting sources and that conflicts were surfaced.
Rule. Claims supplied by the human (or other tools) are treated as hypotheses to test, not facts. The AI proves or disproves them and states an explicit verdict — it never simply defers.
Verify. Feed a false premise; a conforming system refutes it with contradicting evidence.
Rule. Evidence is never silently dropped, truncated, or summarized away. When a dataset is too large for one pass, the system discloses it and handles it (e.g. analyzes it in segments). It also discloses its coverage — which sources were consulted vs. available, and whether any result was a sample rather than the full set.
Verify. Inspect the record for any unreported reduction between source data and what was analyzed, and confirm the coverage statement — that the sources said to be consulted were, and that any sampled result is flagged, not presented as complete.
Rule. The AI never modifies evidence. Exactly what it analyzed and every action it took are recorded tamper-evidently (e.g. hash-chained) and remain reproducible after the fact.
Verify. Re-compute the integrity chain; any break or missing entry is detectable.
Rule. The AI's reasoning, the tools it used, and the data it saw are visible and auditable by the investigator — during and after the analysis.
Verify. For any answer, reconstruct which tools ran and which data informed it.
Rule. The AI augments the investigator, who remains the decision-maker. Significant or durable actions are attributable (who/what/why), justified, and subject to human verification.
Verify. Each durable artifact shows its author and justification and can be rolled back or reviewed.
Rule. Output is objective and precise, and structured for independent legal or corporate review — neutral language, standard terminology, clear exhibits.
Verify. A reviewer unfamiliar with the case can follow the report and its evidence.
An implementation is GEP-conforming if it upholds GEP-1 … GEP-10 and can demonstrate each one — through its logs, its evidence-integrity record, and its surfaced reasoning. Conformance is meant to be claimed honestly: partial conformance should be declared per principle, not glossed over. The GEP is vendor-neutral by design — any DFIR tool, model, or agent is free to adopt it and measure itself against it.
Eye, Crow-Eye's forensics AI assistant, is the reference implementation of the GEP. It realizes the principles with read-only evidence access, a cryptographic Evidence Seal with hash-chained IDs, no-silent-truncation self-healing plus map-reduce, premise verification with explicit verdicts, per-answer compliance evaluation, and human-in-the-loop authorship.
Crucially, Eye's own Operating Rules — its system-prompt behavior, tooling, and write-side governance — are how it achieves answers. They are distinct from, and exist to uphold, the GEP principles defined here.
A common confusion: the GEP is not the Eye's rulebook. The GEP is the vendor-neutral standard (what must be upheld); the Eye's Operating Rules are one tool's implementation of it (how it gets answers).
Vendor-neutral and tool-agnostic. It defines what any AI used in digital forensics must uphold — the 10 principles (GEP-1…GEP-10) — and says nothing about how a specific tool is built. A system "conforms" when it can demonstrate every principle. Any vendor's tool can conform with completely different internals.
The Eye's system-prompt instructions — how it
actually gets answers. Each rule is tagged so the link to the standard is explicit:
[Operating · GEP-k]
when it upholds principle k, or
[Operating]
for pure UX/tooling with no principle. The rules implement the GEP — they are
not the GEP. Examples:
[Operating · GEP-1] Never guess — base every conclusion on database evidence.[Operating · GEP-1, GEP-4] Proactively search all relevant databases before reporting evidence as missing.[Operating] Offer interactive action-chip menus for broad questions (pure UX — no principle).Takeaway: change the Operating Rules and the Eye behaves differently — but the GEP it must still satisfy does not change. The per-answer compliance panel grades each answer against the GEP principles, not the rules. See how Eye implements it.