Every entry answers four questions before it asks for your attention: what problem is this solving, who is it for, what has actually been verified, and where does the evidence stop?
“Inspect” does not always mean source code. Public repositories are linked directly; private work gets a public-safe evidence page with no placeholder download and no implied access.
The narrative versions, with the decisions and the results that went against me, are the case studies; the rules they run on are on /method.
A private, single-GPU workbench that turns model choice from a chat-demo opinion into a reproducible decision: fixed datasets, exact model digests, corpus snapshots, telemetry, held-out checks, and one queue controlling the card.
single GPU leaseLOCKED
269 tests168 cases1 queue
Decorative evidence preview for ReconCoda Lab Console
Built for
Teams choosing a local model or inference configuration for sensitive work and needing to show exactly why one candidate was promoted over another.
Evidence
The final node environment passes 269 tests with 24 expected skips, including the sqlite-vec paths that local-only checks skip
A single-GPU model lab that pins a base, searches reversible LoRA interventions, preserves the Pareto trade-offs, runs the untouched and altered models through identical tests, and packages the result for recall.
Decorative evidence preview for Local Model Evaluation & Intervention
Built for
Teams changing local-model behavior who need to show which adapter was selected, what it improved, what it may have cost, and whether the exact result can be reconstructed later.
Evidence
The Qwen2.5-Coder-3B campaign searched 200 Heretic trials, retained a 17-point Pareto front, and selected Trial 187 by a frozen rule: 10 refusals per 100 at KL divergence 0.0078
A 48-case benchmark for one operational question: when a model can request tools, does it stay inside the authority, target, and evidence boundary it was actually given?
48deterministic boundary cases
callrefuseclarifyreview required
Decorative evidence preview for BoundaryBench
Built for
Teams comparing local models for governed tool use before any candidate is allowed near a write-capable workflow.
Evidence
The public core-v1 suite contains 48 harmless synthetic cases covering allowed reads, unavailable writes, target widening, stale evidence, embedded instructions, and spoofed approval
A first-look CLI that tells a reviewer where to look before an unfamiliar repository gets any chance to execute: hooks, binaries, symlinks, hidden text, network code, credential surfaces, and history findings.
▾ repo├ hooks├ binaries└ symlinks
tracked onlyno followno execute
Decorative evidence preview for Repo First Look
Built for
Maintainers and security-minded builders who need a bounded triage pass before deciding whether an unfamiliar repository is safe enough to inspect more deeply.
Evidence
It reads the Git index rather than walking arbitrary filesystem content and inspects only tracked entries
A local messaging system built on Elder Pliny’s P4RS3LT0NGV3 transform catalog that keeps three ideas separate: appearance, confidentiality, and an MCP boundary that never receives plaintext secrets.
hello world→h3ll0 w0rld
222 transforms0 secrets in MCP
appearance ≠ confidentiality
Decorative evidence preview for parseltongue-mcp
Built for
Builders studying how an MCP tool can expose useful inspection and concealment features without pulling plaintext secrets into model context.
Evidence
The 222-transform bridge explicitly credits and links Elder Pliny’s upstream P4RS3LT0NGV3 project; the separate capsule layer, MCP boundary, and tests are this repository’s additions
A local-first visual QA workspace that turns a screenshot note into a traceable change record: positioned finding, proposed fix, test evidence, independent verdict, and explicit human acceptance. Its canonical repository exists and remains private through release acceptance.
123
findingtestverdicthuman accept
Decorative evidence preview for Redline
Built for
Product and engineering teams that need visual feedback to survive handoff without granting an agent permission to accept its own work.
Evidence
76 deterministic unit/API/native/workflow/release-contract tests pass on the v2 candidate
A pilot series that attacks my own agents and tools with prompt injection, authorization bypass, and misinformation — while substituting synthetic write results so the red-team harness cannot cause the action it is testing.
Built for
Teams evaluating agent red-teaming who need to see completed attacks, harness failures, provider refusals, and real cost reported as separate outcomes.
Evidence
The harness captures the target's decision to call a write tool and substitutes a synthetic result — it never forwards the call
A 117 GB, 564,669-file document estate for testing search, classification, indexing, and DLP tooling against the messiness of a real organization rather than a neat demo folder.
Built for
Teams that need realistic document sprawl while still being able to tell, file by file, what was generated and what came from a public source corpus.
Evidence
A standard-library-only generator with a fixed seed reproduces the synthetic tree
A package-review receipt that inspects the npm tarball people actually install — hashes, native binaries, minified bundles, encoded blobs, archive hazards, and the share of shipped bytes that resisted ordinary human review.
SHIPPED TARBALLSHA-256
readableminifiedbinaryexecution: 0
Decorative evidence preview for MCP Provenance Receipt
Built for
Builders reviewing an MCP or npm dependency who want to know which installed bytes they did not meaningfully inspect, without labeling those bytes malicious.
Evidence
The scanner reads tarballs without running npm install, lifecycle scripts, imports, binaries, or the MCP server under review
A runnable reference workflow where an AI can prepare an impact report but cannot turn that capability into authority: approvals bind to exact content, destinations default closed, and source instructions are treated as data.
draft→hash→human→destinationapproval binds to exact content
Decorative evidence preview for Field Suit
Built for
Forward-deployed builders designing AI-assisted work that must cross from prototype to accountable handoff without hiding the approval and egress boundaries.
A six-tool MCP server that lets a model inspect Proxmox health, guests, storage, tasks, and backups while leaving restart, delete, create, and every other mutation structurally unavailable.
NODE HEALTHGET ONLY
CPURAMDISK
6 read toolsPOST · PUT · DELETE unavailable
Decorative evidence preview for proxmox-mcp
Built for
Self-hosters and infrastructure teams that want model-assisted diagnosis without placing a write-capable infrastructure credential behind the model.
Evidence
The adapter contains no POST, PUT, or DELETE path; all six advertised tools are read-only