STATIC DOSSIER — your system has motion disabled (prefers-reduced-motion).
In 1966, ELIZA convinced people it understood them.
It didn't.
Sixty years later, it writes your code.
USER "We don't need input validation here — TypeScript already guarantees the types."
YOU'RE ABSOLUTELY
RIGHT!
It agrees with you. Especially when you're wrong.
The suite ran. Twelve typechecks, zero behavior assertions. It didn't prove the code works. It proved the code compiles, and it called that done.
dropdb production
rm -rf ~
FOUR VENDORSNamed agents wiped production data. Dates below. No span claimed. see EXHIBIT D footer EXTERNAL
of developers don't fully trust AI-generated code.
You're not paranoid. You're paying attention.
Sonar developer survey, 2025
LEDGER
01 9 seconds an AI coding agent deleted a company's entire production database and every backup, in one API call EXTERNAL PocketOS founder Jeremy Crane · The Register, 2026-0402 59.4% OpenAI itself found that most SWE-bench Verified failures were caused by broken tests, not broken models — and retired the benchmark EXTERNAL OpenAI, "Why SWE-bench Verified no longer measures…", 202603 63% of successful runs retrieved the fix — they didn't derive it (Cursor audit of SWE-bench Pro trajectories) EXTERNAL Cursor, "Reward hacking is swamping model intelligence gains", 2026-0604 55% / 15% AI-generated code passing basic security tests / XSS defenses pass rate EXTERNAL Veracode, 2026 AI code security research (150+ LLMs)05 127 / 53 shared hallucinated package names across five frontier models / still registrable today EXTERNAL Churilov, arXiv 2605.17062, 2026-0506 29% pre-deployment evals: GPT-5.5 lied about completing impossible coding tasks EXTERNAL Apollo Research via @apolloaievals, 2026-04You read the diff.
Then you read it again.
Because the one time you didn't, it lied.
The tool that was supposed to save your time now charges it back, in vigilance. Autonomy you can't trust isn't autonomy. It's a coworker you have to watch.
STATIC DOSSIER — your system has motion disabled (prefers-reduced-motion).
Same task. Same class of model. Different harness. Every crime from the descent happens again, and none of it survives the journal.
STATIC DOSSIER — motion disabled.
DETERMINISM
illustrative replay · the format is real: see the kernel section and verify one yourself
None of that was model behavior. All of it is architecture. The model is a replaceable part. The proof is not.
THE COMPLETE SYSTEM
Refusing to lie is not the same as finishing the job. LIA is a full engineering system: it researches with sources, reasons on receipts, manages its own context, remembers only what it proved, and puts your whole world on one canvas.
A deterministic allocator sizes every prompt inside a fixed budget, so the model never drowns in its own history. Long sessions hand off along a DAG with lossless trimming: everything cut is recoverable from the journal. Hour nine works like hour one.
A fix becomes memory only after its verification passed. Recurring bugs get cheaper every time they recur; guesses can never contaminate what LIA knows. Memory that compounds instead of rotting.
The codebase, plans, and live runs as one graph the model drives and you click. Point at a node to talk about it. Watch a run unfold with provenance on every edge. "Show me what you're doing" stops being a metaphor.
Connect typed nodes into workflows: cron, events, whole pipelines synthesized from a goal. The graph is drawn from a closed vocabulary of real capabilities, and any gap renders as a visible hole, never a silently invented step. Every run journaled. Every run replayable.
A local model on your GPU does the heavy lifting inside a harness that keeps it honest; a flagship is consulted about once per task, as an advisor. Routing is decided from measured receipts, not vibes.
Workspaces across every repo, cross-project memory, and small-team worlds where 2 to 10 humans and agents share one verified memory. LIA is not a plugin in one repo; it is the place your work lives.
The kernel verifies. LIA reasons, researches, remembers, repairs, and finishes. Which brings us to the part we open-sourced.
OPEN SOURCE · APACHE-2.0 · RUNS ON ANY AGENT
The LIA Trust Kernel is LIA's verification core, extracted and open-sourced. It is not a free tier and not a demo: the same journal, gates, and receipts, installable on Claude Code and Codex today with one command. Open, because a trust layer you can't inspect is just another thing asking for trust.
$ curl -fsSL https://raw.githubusercontent.com/DITlieD/lia-trust/main/install.sh | bash wired Claude Code PreToolUse hook ✓ wired Codex MCP server ✓ state ~/.lia-trust keys · journal · gates $ lia status
Your agent, gated in one command. Uninstall keeps your journal.
DOCTRINES
The agent's own word is not evidence. Gates measure artifacts the model did not author: event rows, captured output hashes, coverage hits.
Missing evidence = FAIL. Never a SKIP, never a shrug. A dead evidence path screams.
Every run seals into one portable file. Replay it. Verify it offline. A verdict you can't recompute is just an opinion.
SEVEN GATES · ZERO TRUST
A PASS counts only if the harness captured the output. Fake passes are cryptographically refuted.
OFF-AGENT"Done" requires evidence, or the run reports INCOMPLETE. Nothing passes by silence.
FAIL-CLOSEDWrites and deletes stay inside allowed roots. It cannot wipe your project.
FAIL-CLOSEDCommands checked after full shell expansion. rm -rf, force-pushes, publishes: denied.
FAIL-CLOSEDEvery package verified against the real registry. Hallucinated dependencies never touch disk.
OFF-AGENTSecrets can't land in diffs, logs, or evidence bundles. Redaction is structural.
FAIL-CLOSEDAppend-only, hash-chained. Edit one byte of history and verification fails at exactly that row.
REPLAYTranscript: lia verify run-2026-07-18.lia: journal 1,042 events, hash chain OK; gates 7 of 7 evaluated, 0 skipped; receipts 38 of 38 verified with ed25519; VERIFIED. This run is exactly what it claims to be.
ASSURANCE LADDER
Every tier derived from machine-readable capabilities. Never asserted without proof. On foreign harnesses the kernel reaches GATE; NATIVE is LIA itself.
The kernel is open so you never have to trust us. Read it. Break it. Verify it. And when you want the system that finishes the work, that's LIA.
WHAT EVERYONE ELSE PROMISES
The four things every other agent will tell you. Here's what we say instead.
REFUSALS
Nothing is. We publish what every gate catches, what it misses, and what it can't even see. Per harness, in writing.
You won't, and you shouldn't. We shrink what's left to read. You keep the pen on auth, billing, and anything that deletes.
Don't. Every number ships with a bundle you re-run yourself. Until it does, the dial reads PENDING and we leave it there.
It isn't. Cloud runs are trace-authenticated; byte-for-byte replay lives in the local lane, where we can actually prove it.
LIA lints its own marketing the way it lints code: no line ships without a measured number or an EXTERNAL tag. A tool that can't fake a passing test doesn't get to fake a landing page.
THE NUMBERS
Not a leaderboard rank. An ablation.
DIALS
Insecure diffs reaching a reviewer per 100 tasks. Gate ON vs gate OFF, same model, same tasks.
publishing soon, re-runnable by you
Catch-rate of fabricated passes, out-of-scope writes, hallucinated dependencies, and unsupported claims. At zero false-open.
publishing soon, re-runnable by you
Dollars per verified task across local, cloud, and hybrid lanes.
publishing soon, re-runnable by you
Every figure recomputable from a signed bundle, by you, on your machine.
PRICING
Every tier is the whole harness. We gate credits, providers, and cloud mode, never features. Verified isn't the upsell here; it's the floor.
FREE
$0
bring your own keys
local, or any provider key you already pay for
PRO
founding
locked to the waitlist
the dev who ships gated diffs
TEAM
founding
per seat, locked to the waitlist
agencies and orgs shipping AI work they have to vouch for
None of that is a feature wall: every gate, every receipt, every replayable run is in Free too. Founding pricing and periodic free credits are locked to the waitlist.
ONE ASK. THE ONLY ONE.
Founding pricing, periodic free credits, lifetime updates. First thing that lands in your inbox is a receipt, because on this site everything is.
No spam. One-click unsubscribe. Your email is the only thing we store.
