[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-92714":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":15,"stars7d":15,"stars30d":16,"stars90d":15,"forks30d":15,"starsTrendScore":15,"compositeScore":17,"rankGlobal":10,"rankLanguage":10,"license":18,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":21,"hasPages":19,"topics":22,"createdAt":10,"pushedAt":10,"updatedAt":32,"readmeContent":33,"aiSummary":34,"trendingCount":15,"starSnapshotCount":15,"syncStatus":35,"lastSyncTime":36,"discoverSource":37},92714,"auto","RightNow-AI\u002Fauto","RightNow-AI","the agi compiler: records llm agent behavior, proves what repeats, and compiles it into verified, sandboxed wasm binaries that run for microdollars. nothing figured out twice, paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.04542","https:\u002F\u002Fwww.rightnowai.co\u002F",null,"Rust",69,4,52,0,9,43,"Apache License 2.0",false,"main",true,[23,24,25,26,27,28,29,30,31],"agent","ai-infrastructure","compiler","conformal-prediction","distillation","inference-optimization","llm","program-synthesis","webassembly","2026-07-22 04:02:06","# Auto\n\n**Paper:** [arXiv:2607.04542](https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.04542) · **Built by** [RightNow AI](https:\u002F\u002Fwww.rightnowai.co) · Apache-2.0\n\nAuto compiles recorded LLM-agent behavior into verified, capability-confined\nwasm binaries, with a tiered runtime that falls back to a frontier model for\nnovelty and recompiles the result. It records what an agent does, proves which\nparts are secretly symbolic, extracts those, distills the rest into small\nspecialists, verifies the whole against a behavioral contract, and emits a\ncognition binary (`.cbin`): code plus small models plus a measured manifest.\n\n![cumulative spend: paying the model every task vs auto](docs\u002Ffigure1.png)\n\nThe idea in one line: treat the frontier model as an interpreter, and build\nthe compiler that has always followed interpreters. The plain-language story,\nwith every measured number and its provenance, is in\n[docs\u002Fwhy-agi-compiler.md](docs\u002Fwhy-agi-compiler.md).\n\n`CLAUDE.md` records the engineering norms we held ourselves to while building\nthis; `spec\u002F` is written for external readers.\n\n## what it is\n\n- The compiler: `auto record` attaches to a live agent and captures traces,\n  `auto compile --contract` lowers them to a typed IR, runs the passes, and\n  verifies against the contract before emitting a `.cbin`.\n- The runtime: tier-1 is the compiled fast path. tier-0 is a frontier model as\n  interpreter, for inputs the guard flags as novel. A guard trip deopts to\n  tier-0, captures the trace, and recompiles. This is the ratchet: nothing is\n  figured out twice.\n- The artifact: content-addressed `.cbin` with a manifest that reports eval\n  scores, cost and latency bounds, capability requirements, and provenance.\n  The manifest reports measured numbers or `null`, never aspirational ones.\n\nCapability confinement is physical: emitted artifacts declare zero wasm\nimports and the loader refuses anything else, so a binary cannot exceed its\ndeclared effects.\n\n## measured results\n\nEvery number below is traceable to `paper\u002Fbench-results.md` (which carries the\neval-run ids and ledger lines) or `paper\u002Fclaims.md`.\n\n- Determinism census: 488 of 560 recorded effectful spans were\n  witnessed-deterministic (87.1% pooled), with three families at 100.0% and\n  free-text generation as the residue. This is the measured form of the claim\n  that most agent cognition is secretly symbolic.\n- The ratchet, on a 300-item novelty stream with three scheduled distribution\n  shifts (positions 50, 120, 200): marginal cost per item fell from about\n  59 µ$ (pure frontier) to 2 µ$\u002Fitem steady-state, 6.4x cheaper end-to-end,\n  with three recompile generations each landing one cycle behind a shift, zero\n  errors, and every recompile passing the same contract gate. Tier-1 answers\n  matched the reference agent on witnessed inputs 124 of 128 times (96.9%).\n- Latency ladder on the compiled path: 736 ms (frontier) to about 21 ms (serve\n  over HTTP) to 290 µs (stdio) to 54.1 µs (in-process python, pyo3) to 18.2 µs\n  (in-process node, napi). The in-process figures measure the call boundary on\n  trivial fixtures, not inference.\n\nThe same stream run at a deliberately loose guard measured the failure mode the\narchitecture exists to prevent: 48.9% of tier-1 answers silently wrong.\nCalibration, not model capability, decides whether cheap stays correct.\n\n## pipeline\n\n```\n  frontend            IR                 passes                 backend\n  --------            --                 ------                 -------\n  trace SDK   -->  typed task graph  --> extract  --> verify --> task.cbin\n  (prompts,        (capability and       distill      (contract   (wasm code\n   tool calls,      memory effects,      optimize      gate)        + small models\n   args, results)   uncertainty,                                    + manifest)\n                    resource bounds)\n\n  run:  input --> guard --> in-distribution --> tier-1 (compiled, guarded)\n                        \\-> novel --> deopt --> tier-0 (frontier) --> capture\n                                                        \\-> recompile --> tier-1\n                            (the ratchet: nothing figured out twice)\n```\n\nThe passes, in order: symbolic extraction (enumerative search or LLM-guided\nCEGIS, candidates checked in a wasmtime sandbox with no network), distillation\n(residual fuzzy nodes into small tree or MLP specialists), verification (the\ncontract is the type checker; differential testing against the reference model;\na failing or unmeasurable contract blocks emit), and optimization. Guards use\ntrigram-Jaccard distance with split-conformal calibration for calibrated\nabstention.\n\n## build\n\nRequirements:\n\n- rust 1.96.1 (edition 2024), pinned in `rust-toolchain.toml`.\n- `flatc` 25.12.19 exactly. It must match the pinned `flatbuffers` crate; the\n  IR build fails on a mismatch.\n\n`crates\u002Fauto-ir\u002Fbuild.rs` resolves `flatc` in this order: the `FLATC` env var,\nthen `tools\u002Fflatc\u002Fflatc[.exe]` (gitignored), then `PATH`. Install it from the\nofficial release, for example on Windows extract `flatc.exe` from\n`Windows.flatc.binary.zip` at\n`github.com\u002Fgoogle\u002Fflatbuffers\u002Freleases\u002Fdownload\u002Fv25.12.19\u002F` into\n`tools\u002Fflatc\u002F`. Linux uses `Linux.flatc.binary.clang++-18.zip` from the same\nrelease.\n\nThe gates CI runs on every pull request:\n\n```\ncargo fmt --all --check\ncargo clippy --workspace --all-targets -- -D warnings\ncargo test --workspace\n```\n\n## quickstart\n\nBuild the CLI, then run the end-to-end script. It records the toy agent through\nthe real python SDK, checks the measured determinism report, compiles a `.cbin`\nthrough the verification gate, runs it, and proves the negative paths: a wrong\nimplementation is blocked, a far input abstains, a deopt is ingested and\nrecompiled to tier-1, and a tampered registry artifact is refused.\n\n```\ncargo build -p auto-cli\nbash evals\u002Ftoy-agent\u002Fe2e.sh\n```\n\nThe same steps by hand, one task, record to run:\n\n```\n# record the toy agent twice, then read the measured determinism report\ncargo run -p auto-cli -- record --store store.db -- python evals\u002Ftoy-agent\u002Fagent.py\ncargo run -p auto-cli -- record --store store.db -- python evals\u002Ftoy-agent\u002Fagent.py\ncargo run -p auto-cli -- report --task toy-agent --store store.db\n\n# verify a contract against the recorded spans (writes a content-addressed eval run)\ncargo run -p auto-cli -- verify --contract evals\u002Ftoy-agent\u002Ffake-frontier.contract.toml --store store.db\n\n# compile the span into a .cbin; without --module the implementation is\n# synthesized from the recorded observations, and emit is verification-gated\ncargo run -p auto-cli -- compile --contract evals\u002Ftoy-agent\u002Ffake-frontier.contract.toml --store store.db --out fake-frontier.cbin\n\n# run the compiled artifact\ncargo run -p auto-cli -- run --artifact fake-frontier.cbin \\\n  --input '{\"prompt\":\"The quick brown fox jumps over the lazy dog near the riverbank.\"}'\n\n# inspect the artifact and its manifest\ncargo run -p auto-cli -- inspect fake-frontier.cbin\n```\n\nThe paid paths (LLM-guided CEGIS at `--synth llm`, and a frontier model as\ntier-0 via `--tier0 \"frontier:\u003Cmodel-id>\"`) are fail-closed: they need\n`OPENAI_API_KEY` and an explicit nonzero `--spend-cap-usd`. The default cap of 0\nrefuses every paid call, and every call is appended to `~\u002F.auto\u002Fspend.jsonl`.\n\n## benchmark\n\nAUTO-BENCH v1 asks the question the thesis lives on: does the system ever pay\nfor the same thought twice? The protocol is frozen before execution in\n`evals\u002Fbench\u002FDESIGN.md`. The measured results, with every number carrying an\neval-run id or a ledger line and a failures-and-refusals section at equal\nweight, are in `paper\u002Fbench-results.md`; per-position stream data is under\n`paper\u002Fevidence\u002F`.\n\nReproduction: `evals\u002Fbench\u002FREADME.md` has one leg per task family. Paid loops\nare gated behind `RECORD=1` or `JUDGE=1` with the spend cap passed as an\nargument. A stranger with an OpenAI key reproduces every table for well under\n$1. Total ledgered benchmark spend was $0.0621 of a pre-registered $5.00 cap.\n\n## layout\n\n| path | what |\n|---|---|\n| `crates\u002Fauto-ir` | typed task graph: effects, uncertainty, resource bounds; flatbuffers serialization with byte-stable round-trip |\n| `crates\u002Fauto-trace` | trace model, strict JSONL ingestion, sqlite store, determinism report, replay comparison |\n| `crates\u002Fauto-contract` | contract format, verification harness with three-valued verdicts, content-addressed eval runs |\n| `crates\u002Fauto-passes` | symbolic extraction (enumerative search or LLM-guided CEGIS), region synthesis, tree and MLP distillation drivers |\n| `crates\u002Fauto-dsl` | the closed extraction DSL: one evaluator, compiled natively for search and to wasm for artifacts |\n| `crates\u002Fauto-model` | distilled-model wire format and tree inference over frozen char-trigram features |\n| `crates\u002Fauto-backend` | `.cbin` container (content-addressed), honest manifest, differential checks, verification-gated emit |\n| `crates\u002Fauto-runtime` | tier-1 wasm execution (zero-import capability refusal, fuel and memory limits), conformal guards, deopt, recompile ingestion |\n| `crates\u002Fauto-frontier` | the only paid-API path: spend-capped client, pinned price table, append-only ledger |\n| `crates\u002Fauto-registry` | local content-addressed artifact store and detached ed25519 signing |\n| `crates\u002Fauto-serve` | `auto serve`: registry artifacts over HTTP, guard-gated tier-1 per request, 409 abstention on a trip |\n| `crates\u002Fauto-proxy` | `auto proxy`: record any OpenAI-backed agent with zero code changes |\n| `crates\u002Fauto-daemon` | `auto daemon`: the ratchet as a service; watch a store, recompile on new evidence, publish |\n| `crates\u002Fauto-cli` | `record`, `report`, `verify`, `compile`, `distill`, `run`, `registry`, `inspect`, `serve`, `proxy`, `daemon` |\n| `crates\u002Fauto-py` | in-process python embedding (pyo3, abi3 wheel) |\n| `crates\u002Fauto-node` | in-process node embedding (napi addon) |\n| `sdk\u002Fpython`, `sdk\u002Ftypescript` | recording and replay tracers, same wire format |\n| `spec\u002F` | `ir.md` and the dialect specs; `adr\u002F` for irreversible decisions |\n| `evals\u002F` | reference tasks, e2e scripts, and the benchmark under `evals\u002Fbench` |\n\n## honest limitations\n\nThe benchmark corpora are designed: realistic but synthetic. The production\nclaim is an operator rerun on their own recorded traffic, not this corpus. The\ndeterminism and parity numbers are measured on those tasks at benchmark scale\n(560 spans), not a capability eval of the underlying model, which is the\nreference interpreter and not the subject. Free-text generative behavior is the\nhonest residue: at 40 tickets the summarize family does not compile to a tree\nat its declared judged threshold and stays tier-0, and field-extraction is a\nfully deterministic behavior the v0 output algebra cannot yet compile, an honest\nrefusal at all three rungs. Guards are lexical (Jaccard and cosine); they\ncalibrate real vocabularies well but admit lexical cousins, and a loose guard on\nthe novelty stream produced 48.9% silently wrong tier-1 answers. Semantic\nembedding guards and sigstore signing are recorded targets, not claims; current\nsigning uses a single local ed25519 keypair, and recorded cost and token\nattributes are the agent's own declaration.\n\n## license\n\nApache-2.0. Copyright 2026 RightNow AI. See `LICENSE`.\n\n## citation\n\n\"Auto: The AGI Compiler\", Jaber Jaber and Osama Jaber, arXiv:2607.04542.\nPaper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.04542\n\n```bibtex\n@misc{jaber2026auto,\n  title         = {Auto: The AGI Compiler},\n  author        = {Jaber, Jaber and Jaber, Osama},\n  year          = {2026},\n  eprint        = {2607.04542},\n  archivePrefix = {arXiv},\n  primaryClass  = {cs.LG},\n  url           = {https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.04542}\n}\n```\n","RightNow-AI\u002Fauto 是一个面向LLM智能体的编译器系统，将运行时记录的代理行为自动提炼为经形式验证、能力受限的WebAssembly二进制（.cbin），实现低成本、低延迟的确定性推理。其核心包含行为录制、符号性识别、模型蒸馏、契约化验证与分层运行时（编译快路径 + 前沿模型兜底解释器），支持自动回退、重编译与成本收敛。适用于需高频调用、严格成本控制与可验证行为的生产级AI代理场景，如自动化客服、规则敏感型决策服务及边缘侧轻量AI部署。",2,"2026-07-10 02:30:15","CREATED_QUERY"]