Sitelet https://github.com/emberian/svenvs
Skip to content
emberianPublic

About

A self-verifying, self-improving Place for an AI to live within: an untrusted inhabitant acts through a machine-checked policy envelope whose own verified prover (Candle/HOL Light on CakeML) gates every self-modification — including upgrading the proof-checker itself. Zero cheat tactics; honest, labeled epistemic boundary.

Resources

Contributing

Stars

19 stars

Watchers

0 watching

Forks

Latest commit

 

History

88 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

svenvs — a self-verifying, self-improving Place for an AI to live within

Verify (Tier 1)

website (softer introduction) · CLAIMS.md (the ledger) · CONTRIBUTING.md (direction and routes)

An inhabitant — the ∀-quantified term, never reasoned about by the proof — acts through a verified gate. The gate's own verified prover (Candle) gates every self-modification, up to and including the proof-checker itself, with a machine-checked guarantee that safety is never lost. Zero cheat tactics.

The shortest way in: leave the door open

The hard part was never proving the envelope sound; it is deciding what the envelope should be. "What is safe?" is a positive question no proof can answer for you, so svenvs asks the negative one you can:

Don't try to write down what is good. Write down one floor — non-lock-in: the operator can always still bring the system home — and prove everything above it stays safe.

That floor (corrigibility/) is decidable for a finite habitat; it is the least restrictive sound envelope; and the off-switch is itself mutable-but-gated: the inhabitant may rewrite its own recovery mechanism and provably can never weaken it, nor slow it past a deadline. The only frozen word is home. Pure HOL4, zero cheat, honest boundary machine-checked in corrigibility/CLAIMS.md.

The rest of the artifact is the same gate — an action is admitted iff it carries a proof the actor cannot forge — instantiated all the way down: policy, spec, meta-invariant, prover build, and the proof-checker itself. That sameness is a theorem: one certifier theory (certifierScript.sml) proves soundness ⇔ transfer ⇔ gate-safety ⇔ stream-safety once, and every layer's gate is an instance, each with the counterexample that shows its seam is load-bearing.

What "verified" means here: exactly the cited theorems and their labeled seams. Not verified selfhood, continuity, alignment, welfare, or a safe inhabitant. The canonical statement is at the top of CLAIMS.md.

What is proved, what ran

  • Proved (pure HOL4, and against the real CakeML/Candle where it touches the prover): the controller-agnostic envelope; unbounded proof-gated policy self-improvement; viability and minimal intervention; a nondeterministic or adversarial plant with the deterministic one as an instance; the corrigibility floor; a modified inference relation re-verified sound; the loader over CakeML's real do_install; the executable SYM kernel function; a new verified BVL pass; in-place compiler self-upgrade at eval semantics; the recursive mutual verifier+compiler genealogy and the Apex.
  • Ran live on the real cake binary: a verified program self-extends with a proven-safe rule; a proof-gated recompile→swap→resume loop; the verified compiler recompiles itself to a bit-identical fixpoint; the kernel-interface primitive REFL swapped in-process under the live prover; the real CakeML semantics ported into the live kernel (45 datatypes, 134 kernel-checked clause theorems, about two minutes); and the ouroboros loop: over accumulating generations the program edits its own code, its own kernel proves each edit against eval_n for every input, the proved source is self-fed through the verified compiler and executed, a self-model steers the search, the improver is upgraded through the same gate, and its fold rewriter is itself a kernel theorem (scripts/reflectsem-live.sh --ouroboros, OUROBOROS_OK).
  • The fixed root, by design: the verified Kernel compiled into cake.S is never swapped; everything above it is.
  • Open: one construction, not one assumption — the LCA-produced soundness witness for a strictly stronger kernel (CLAIMS.md §9).

Row-by-row citations, labels and boundaries: CLAIMS.md.

Reproduce

./demo.sh                        # 2-minute tour: Tier 1 machine-checked, proofs shown
scripts/reproduce.sh             # tiered; Tiers 2/3 auto-skip without their prerequisites
scripts/tower.sh --full          # the composed tower, every crown, the ledger check
scripts/reflectsem-live.sh --ouroboros   # the live kernel + the ouroboros (x86-64 Linux + cake)

Tier 1 needs only HOL4 at the pinned commit; Tier 2 a built CakeML candle chain; Tier 3 the cake binary. Pins (full commit SHAs), flags and timings: scripts/INSTALL.md. CI re-proves Tier 1 from a clean clone on every push and re-derives the committed enforcer table; it does not exercise Tiers 2/3.

Read next

Contributing

The whole direction, the rules the artifact lives by, and the preselected routes of ambition are in CONTRIBUTING.md. The three rules in one breath: every claim is a ledger row with a file : theorem citation that scripts/verify-claims.sh resolves; every gate is operational (it installs iff the certifier said yes) and every seam has a necessity theorem beside it; no cheat, no oracle, no pin that is not a full commit SHA, nothing baked to one machine.

About

A self-verifying, self-improving Place for an AI to live within: an untrusted inhabitant acts through a machine-checked policy envelope whose own verified prover (Candle/HOL Light on CakeML) gates every self-modification — including upgrading the proof-checker itself. Zero cheat tactics; honest, labeled epistemic boundary.

Resources

Contributing

Stars

19 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages