// how we grade tools

every tool inside fatcousin forensics
is publicly graded.

fatcousin forensics — the investigator-facing section of fatcousin.com — runs an automated audit against every tool in the forensics catalog and assigns a letter grade. the grade is visible on each tool's page, in case-type and industry lists, and exported alongside every captured tool run inside a case session. nothing is hidden. this page explains exactly how the grade is computed so investigators can decide for themselves how much weight to put on any output.

current grades · 1199 tools inside fatcousin forensics

  • 555A · 46%
  • 644B · 54%
  • 0C · 0%
  • 0D · 0%

counts regenerate every time the audit script is run (see methodology below). grades are not promotional — a B-grade tool is a perfectly useful tool with known boundaries; the rubric values honesty about limits over the appearance of polish.

letter grades

letters compress a 14-point raw score plus capability-class caps plus overclaim penalties into a single chip.

gradewhat it meansraw scorewhat to expect from output
Aflagship triagereal parser depth · 2+ exports with reason fields · honest limits · canonical UI shellraw 12–14 / 14trust as a first-pass tool · still verify before any legal use
Bsolid triagereal parsing or rule engine · at least one structured export · honest about boundariesraw 9–11 / 14useful for narrowing an investigation · expect rough edges on uncommon inputs
Cthin or overclaimingmarketing exceeds the engine · or shallow parser · or known UI driftraw 5–8 / 14treat output as a lead at best · scheduled for deepen-or-reframe
Dbroken or stubmissing engine · crashes on standard input · or no exportsraw 0–4 / 14do not rely on output · scheduled for rebuild or removal

capability classes

one capability class per tool — describing the dominant technique its engine uses. each class has its own ceiling. heuristic and ML-adjacent tools cap below A by default; they have to earn the ceiling through honesty, output quality, and UI consistency.

codeclasswhat it doesmax gradeexamples
Eexport / document parserstructured exports — JSON, CSV, EML, ZIP, HTML — parsed against the file format specAgoogle-takeout · gmail-mbox · slack-export
Sstructured store parserbinary databases and plist/registry hives — SQLite, plist, lnk, registry text, logcatAsqlite analyzers · plist parsers · registry hive inspectors
Bbinary / media inspectionbyte-level inspection of image/audio/video/document containers and metadataApng-chunk · exif · header carving · stream inspection
Hheuristic screenerrule-based scoring · statistical anomaly detection · gap analysis — flags worth checking, not verdictsB (A when honesty + output + UI all maxed)anomaly detectors · gap analyzers · stat-only screeners
MML-adjacent claimtools that imply learned classification — synthetic-voice, fingerprint, perplexity, deepfakeC (B when a real bundled model is verified)synthetic-voice detectors · perplexity scorers · authorship probes

seven dimensions

each dimension is scored 0, 1, or 2. raw score is the sum, out of 14. the rubric defines anchors for each score so two reviewers looking at the same tool land within one letter grade of each other.

codedimensionwhat it scoresscore 1 anchorscore 2 anchor
IFinput fidelityhow many real artifact types and schema variants the parser actually handleshandles one well-known formathandles 2+ artifact types or multiple known schema variants
OUoutput usefulnesswhat an investigator can actually export and reuse downstreamone structured export2+ exports with source-file and reason fields
DQdetection qualitythe granularity and explainability of any flagging logicnamed rules + thresholds + reasons visible in the UImulti-signal flags with severity levels
RBrobustnesshow the tool behaves on malformed, empty, or oversized inputsgraceful skip with log · multi-file ingestper-file error isolation · large input safe
HNhonestyalignment between marketing copy and what the engine actually doesdescription matches capabilityprecise copy plus a visible limitations panel for heuristic/ML tools
CCcoverage / correlationwhether the tool joins evidence across files instead of treating them in isolationmulti-file dedupe and sorted timelinecross-artifact correlation as promised
UIui consistencyuse of canonical fatcousin shell components for predictable investigator behavioruses container/section/dropzone primitives · some shell componentsprimary-button for exec · system-log · no local styled buttons · exec + status sections

caps & overrides

  • capability cap: H-class is capped at B unless honesty, output usefulness, and UI all score 2 and raw is ≥12. M-class without a verified bundled model is capped at C.
  • overclaim cap: if the tool name or description uses a high-risk token (“detector”, “analyzer”, “decoder”, “reconstructor”, etc.) without engine evidence to support it, and honesty is below 2, the grade is capped at C.
  • UI floor: if UI scores 0 and the raw score would otherwise be A or B, the grade drops one letter until the UI is migrated. (logic can be strong while the interface is still drifted.)

what a grade does not mean

  • not a court-admissibility score. admissibility depends on jurisdiction, witness qualifications, and the entire evidentiary chain — not on anything fatcousin forensics, or fatcousin more broadly, can compute.
  • not a guarantee of correctness. even A-grade tools can be wrong on rare inputs, malformed files, or adversarial samples. independent verification is always required before consequential use.
  • not a popularity score. the grade is independent of how many people use the tool or how often it ships in a curated stack.

methodology & source

the rubric in long form lives at docs/forensics-tool-quality-rubric-v0.md in the public fatcousin repository. the audit is run by npm run tools:audit-forensics and outputs forensics-audit.csv at the repo root. the slug → grade map shown on each fatcousin forensics tool page is regenerated from that csv on every build via npm run forensics:grades.

newly-added tools entering fatcousin forensics must clear the B ship-bar before merging. the bar lives in .cursor/rules/forensics-new-tool.mdc and is enforced by quality.audit.json sidecars on each tool folder.

ready