What is Pilotlight? | FLESH SIMULATOR

AI FLESH SIMULATOR · 9m 29s · Watch on YouTube · 22 sources

Decision Card

Effort: About an hour to judge it, more if you adopt it. That means finding the actual Pilotlight build or repo (the video gives no link, and a web search turned up nothing), getting an Anthropic API key, exporting one Discord channel with DiscordChatExporter, and checking whether the “Jev” legal analysis cites statutes correctly on a few sample events.

Honest take: The video calls the tool a lawyer-librarian that “can’t make up” facts, yet its own statute list is wrong: it gives § 223 as a Title 18 section, but that harassing-communications law is 47 U.S.C. § 223. It also promises a “lawsuit” built from criminal statutes, which mixes up prosecution and civil litigation. Separately, the “sees deleted posts” claim works against the X developer rules it relies on for collection, which tell developers to remove deleted posts from stored data.

Concrete next steps:

  • Find the actual Pilotlight project, repo or site before going further (15–30 min). A web search for “Pilotlight FLESH SIMULATOR” found nothing.
  • Try the import path on its own (30 min). Export one channel to JSON with DiscordChatExporter and check which metadata survives (author ID, timestamp, message link).
  • If the goal is evidence you can use in court, read up on FRE 902(14) and hashing at collection time (20 min). JSON metadata alone doesn’t make evidence admissible.
  • Before relying on any “protected speech vs. threat” verdict from the tool, read Counterman v. Colorado (15 min).
  • Skip if you aren’t actually documenting a coordinated harassment campaign, or if you already have a lawyer or victim advocate. In that case, established free tools like DocuSAFE plus counsel are the safer path.

TL;DR

Pilotlight is a local-first, encrypted desktop app that collects harassers’ posts from X, Bluesky and Discord into a timestamped, severity-scored timeline. It then uses Claude-powered agents to map each event to US federal statutes and draft, refine and stress-test a legal complaint. The pitch is “no fabrication, every fact sourced via RAG”, but the video gets its own statute citations wrong and never shows where to get the software.

Key Points

  • All data, including API keys, sits in an encrypted local vault behind a passphrase of at least 10 characters or a hardware key, with no password recovery. 00:19
  • The “brain” needs two keys: an Anthropic API key (sk-ant…) for reading and writing, and a second “Type Safe” key labeled “Jev” for legal judgments. 01:09
  • Settings include an operator ID, workspace ID, permission level (stamped on exports) and a spending budget, because each analysis costs “a few cents” in tokens. 01:41
  • Case files are organized as campaigns containing actors (accounts across X, Bluesky, Discord, Telegram). A date-range crawler stores exact post text with UTC timestamps and a 0–10 severity score. 02:42
  • Bulk import takes DiscordChatExporter JSON and Twitter data exports, and claims to auto-match Discord users to Twitter accounts. 03:46
  • A “dash cam” mode watches an actor and captures new posts live. A “synchronicity detector” flags coordinated posting across accounts. 04:22
  • “Jev” maps each event to a likely federal statute (2261A, 875(c)/(d), 223, CFAA, wire fraud, 371 conspiracy) with a confidence level, and flags protected speech as a dead end. 05:31
  • The drafting module has three modes: generate (a multi-agent “firm” builds the complaint plus strategy as a zip), iterate (targeted improvement passes) and simulate (a 10-step adversarial mock trial). 07:16
  • It claims RAG grounding: every fact has a source, and gaps in the record are stated outright rather than filled in. 08:40

Notable Quotes

  • “We keep the exact words, not their paraphrase, because paraphrasing is the first thing that breaks down a defense attorney.” 03:11
  • “When one person screams—it’s just a bad day. 90 documented events with timestamps—that’s a pattern of behavior, and that’s what you go to the police with.” 05:03
  • “He would rather leave an honest gap than offer a convenient lie, for a confident fabrication loses a case, and an honest gap wins it.” 06:41

Verified Claims

  1. Anthropic API keys start with sk-ant. 01:15 Sources: Apideck guide, unified.to Verdict: Confirmed. Third-party sources agree on the sk-ant- prefix; I didn’t check Anthropic’s own docs.

  2. 18 U.S.C. § 2261A covers cyberstalking. 05:43 Sources: govinfo, 18 U.S.C. ch. 110A, Justia 2020 Verdict: Confirmed.

  3. § 875(c) covers interstate threats and § 875(d) covers extortion. 05:46 Sources: Cornell LII 18 U.S.C. § 875, uscode.house.gov § 875 Verdict: Confirmed. (d) covers extortionate threats against property or reputation, and (c) covers threats to kidnap or injure.

  4. “Title 18 … Subsection 223” covers obscene and offensive messages. 05:53 Sources: Cornell LII 47 U.S.C. § 223, govinfo 47 U.S.C. § 223 Verdict: Disputed. The law exists, but it is in Title 47 (Communications Act and CDA), not Title 18.

  5. The CFAA applies to doxxers. 05:58 Sources: CRS R46536 on the CFAA, The Legal Guide on doxxing Verdict: Disputed (overstated). The CFAA depends on unauthorized computer access. Republishing information that was already public isn’t clearly covered.

  6. § 371 is federal conspiracy. 06:01 Sources: uscode.house.gov § 371, Cornell LII § 371 Verdict: Confirmed. It needs an agreement to commit a federal offense plus an overt act. “Same three accounts at the same time” alone doesn’t prove an agreement.

  7. Some disgusting speech is legally protected, so the tool flags it as a dead end. 06:17 Sources: Counterman v. Colorado (Wikipedia), SCOTUSblog Verdict: Confirmed. For true threats, the speaker must have been at least reckless about how threatening the words were. That mental-state test is hard to judge from post text alone.

  8. Preserving JSON metadata is critical for admissibility. 03:35 Sources: Forensic Focus on FRE 902(14), X1 / Forensic Focus Verdict: Inconclusive. Metadata helps, but self-authentication under 902(14) depends on hash verification and a certification from a qualified person, neither of which the video mentions.

  9. Pilotlight can see deleted posts and accounts. 03:28 Sources: X Compliance Streams, X Batch Compliance Verdict: Disputed. X’s developer terms tell developers to keep stored content in sync with what’s live on X, which means removing deleted posts. Content captured before deletion is the only plausible way this works.

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. Do any of the cited criminal statutes (2261A, 875, 47 U.S.C. 223) give a private right of action, or can the generated “lawsuit” really only work as a referral package for police or the FBI?
  2. Does the tool hash captures at collection time and produce a 902(11)/(14)-style certification? Without that, how much of its “admissible” evidence would survive an authentication challenge?
  3. Does continuously crawling and storing a named person’s posts with “dash cam” mode break X’s or Bluesky’s developer terms? Could it leave the victim exposed to a counter-claim or a revoked API key?

Sources