惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

罗磊的独立博客
L
LangChain Blog
aimingoo的专栏
aimingoo的专栏
IT之家
IT之家
B
Blog
博客园_首页
博客园 - 司徒正美
有赞技术团队
有赞技术团队
博客园 - 聂微东
I
InfoQ
美团技术团队
GbyAI
GbyAI
阮一峰的网络日志
阮一峰的网络日志
H
Help Net Security
大猫的无限游戏
大猫的无限游戏
MyScale Blog
MyScale Blog
WordPress大学
WordPress大学
The GitHub Blog
The GitHub Blog
A
About on SuperTechFans
人人都是产品经理
人人都是产品经理
Microsoft Azure Blog
Microsoft Azure Blog
Engineering at Meta
Engineering at Meta
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
The Cloudflare Blog

Hacker News - Newest: "AI"

AI can't read an investor deck AI as an attorney? Student uses ChatGPT, Gemini to sue UW over alleged racial discrimination Hacking MCP Servers in AI Systems – The Rug Pull: Tool Changes After Approval GitHub - MeepCastana/KubeezCut: Free Web based video editor Can AI judge journalism? A Thiel-backed startup says yes, even if it risks chilling whistleblowers Coming soon: 10 Things That Matter in AI Right Now DARPA built an AI to fact-check enemy weapons claims What explains heterogeneity in AI adoption? When AI Meets Muscle: Context-Aware Electrical Stimulation Promises a New Way to Guide Human Movements - Department of Computer Science AI Changed How We Build. It Did Not Change What Matters. Linux rules on using AI-generated code - Copilot is OK, but humans must take 'full responsibility for the… Meta spins up AI version of Mark Zuckerberg to engage with employees Code Mode: Let Your AI Write Programs, Not Just Call Tools | TanStack Blog GitHub - Delavalom/graft: Go framework for building AI agents. Type-safe tools, multi-provider (OpenAI, Anthropic, Gemini, Bedrock), zero vendor SDKs. India's TCS tops estimates, says new AI models did not dent services demand Gen Z's fading AI hype Strong feeling: we are in a folded AI reality GitHub - machinarii/total-recall-catalog: A reference catalog of latest knowledge retrieval, memory & RAG systems GitHub - mensfeld/code-on-incus: Give each AI agent its own isolated machine with root, Docker, and systemd. Active defense detects and stops threats automatically.. Quantization, LoRA, and the 8% Problem: Benchmarking Local LLMs for Production AI Iran war: We spoke to the man making Lego-style AI videos that experts say are powerful propaganda Powell, Bessent discussed Anthropic's Mythos AI cyber threat with major U.S. banks GitHub - immartian/bellamem: Persistent belief-graph memory for AI agents. Retrieves decisive context by importance — not recency, not RAG, not /compact. recursive-mode: The Repo-Native Operating System for AI Engineering After the attack on Sam Altman's home, will AI CEO's go on the offensive? The biggest advance in AI since the LLM Opus 4.6 vs GPT 5.4 One Prompt Unity World Generation Test “AI polls” are fake polls Client Challenge Can AI be a 'child of God'? Inside Anthropic's meeting with Christian leaders
GitHub - provedex/provedex: Cryptographic evidence layer ...
adi-suresh · 2026-05-27 · via Hacker News - Newest: "AI"

ci license

Cryptographic evidence layer for AI agents. Every utterance, tool call, and model output is signed at emission time and chained to the one before it. Anyone with the public key can verify the log, offline, with no involvement from the operator who produced it.

Built for regulated AI: healthcare scribes, financial voice agents, legal intake bots, claims handlers. Same primitive works for any agent whose decisions land in front of a regulator or a court.

Why this exists

The EU AI Act Article 12 obligation begins on August 2, 2026. High-risk AI deployments in the EU must produce tamper-evident logs of agent reasoning. Penalties go up to 15M EUR or 3% of global revenue.

Existing logging stacks (Datadog, Splunk, OpenTelemetry) record what the operator's app says happened. They do not produce evidence that survives an adversarial review. Compliance SaaS vendors collect screenshots; the post-Delve market has learned what that is worth.

Provedex is the primitive underneath. Sign locally, chain locally, verify offline. The operator never has to trust a vendor for the integrity of the log.

Components

Crate / package Role Status
provedex-core signing primitives, hash chain, NDJSON ledger, export bundle shipped
provedex-cli provedex command-line tool: verify, replay, export shipped
provedex-agent localhost HTTP signing daemon for non-Rust customers (default integration) shipped
provedex-pipecat (Python) Pipecat FrameProcessor that signs every frame through the sidecar. PyPI. See bindings/python/provedex-pipecat/. shipped
provedex-langchain (Python) LangChain CallbackHandler that signs every LLM and tool call via the sidecar. Covers LangGraph by inheritance. PyPI. See bindings/python/provedex-langchain/. shipped

Framework adapters that wrap the sidecar are the integration layer. Native FFI bindings (PyO3 / napi-rs) are a future optional fast-path; the sidecar covers every other language via localhost HTTP. See ADR 0004.

The reference voice-agent integration (whisper.cpp + Ollama + Piper) lives in its own repo at provedex/demo-voice. It dogfoods this SDK against the published v0.1.0 tag.

Install

Pre-built binary (recommended)

Download the latest release for your platform from GitHub Releases and extract:

tar -xzf provedex-agent-vN.N.N-aarch64-apple-darwin.tar.gz
sudo install -m 755 provedex-agent /usr/local/bin/

Container (Kubernetes / Docker)

docker pull ghcr.io/provedex/provedex-agent:latest
docker run --rm -p 8765:8765 ghcr.io/provedex/provedex-agent:latest

Multi-arch: linux/amd64, linux/arm64. Customer apps in the same pod or Docker network POST to the agent's /v1/sign endpoint.

From source

cargo install --locked --git https://github.com/provedex/provedex --bin provedex-agent

systemd / launchd / Kubernetes manifests

See deploy/ for ready-to-adapt manifests.

Python bindings (PyPI)

pip install provedex-pipecat      # Pipecat voice agents
pip install provedex-langchain    # LangChain + LangGraph

Both packages POST to the local provedex-agent sidecar. Run the agent first, then wire the binding into your pipeline. See the per-package READMEs for a five-line quickstart.

Quickstart - sidecar

The sidecar is the default integration path for any non-Rust app.

git clone https://github.com/provedex/provedex
cd provedex
cargo build --release -p provedex-agent
./target/release/provedex-agent

The agent binds 127.0.0.1:8765 and auto-creates a keypair at ~/.provedex/keys/ed25519.key. Sign an event from any language:

curl -X POST http://127.0.0.1:8765/v1/sign \
  -H 'content-type: application/json' \
  -d '{"event":{"type":"SessionStarted","payload":{"agent_id":"demo","model_id":"gpt-4o","session_id":"s1"}}}'

Verify the chain:

curl -X POST http://127.0.0.1:8765/v1/verify

Per-language clients (Python, Node, Java, Go, Ruby, PHP) live in docs/integration/sidecar.md.

Quickstart - Rust crate

For Rust apps, link the crate directly:

[dependencies]
provedex-core = "0.1"
use provedex_core::{AgentEvent, Ledger, LedgerSession, SigningKeypair};

let keypair = SigningKeypair::load_or_create("./key")?;
let ledger = Ledger::open("./ledger.ndjson")?;
let session = LedgerSession::open(keypair, ledger, "session-1".into())?;

let signed = session.seal_and_append(AgentEvent::SessionStarted {
    agent_id: "agent-1".into(),
    model_id: "gpt-4o".into(),
    session_id: "session-1".into(),
})?;

Run the minimal end-to-end example:

cargo run -p provedex-core --example basic_signing

Voice agent reference (separate repo)

A working voice-agent integration on top of this SDK lives at provedex/demo-voice. It records audio in the browser, transcribes via whisper.cpp, calls a local Ollama model, signs every step with provedex-core v0.1.0 (consumed via git tag), and optionally speaks the reply via Piper.

That repo is the reference dogfood: how a customer integrates provedex-core into a real voice pipeline.

Repository layout

crates/
  provedex-core/    signing primitives, hash chain, NDJSON ledger, export bundle
  provedex-cli/     `provedex` command-line tool
  provedex-agent/   localhost HTTP signing daemon (default integration)
bindings/
  python/           PyO3 wrapper (planned)
  node/             napi-rs wrapper (planned)
docs/
  spec/             byte-level normative specs (canonical-json, event-schema-v1, signature-scheme, ledger-format, openapi.yaml)
  adr/              architecture decision records
  integration/      framework-specific integration guides
  compliance/       regulator clause mappings (planned)
examples/           runnable integration examples
deploy/             systemd, launchd, Kubernetes sidecar manifests
benchmarks/         load-test scripts for the sidecar HTTP API

The voice-agent reference demo lives in a sibling repo: provedex/demo-voice.

CLI

provedex verify                              # verify the local ledger
provedex replay                              # human-readable transcript
provedex export --output ./bundle.json       # signed export bundle for an auditor

provedex verify exits non-zero if the chain is broken.

Specs

Normative documents that bindings, auditors, and third-party verifiers implement against:

A binding implementation that follows these specs produces signed events byte-identical to the Rust reference.

Build and test

cargo fmt --all -- --check
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo test --workspace --all-features

CI runs the same three checks plus cargo audit and cargo deny on every push and pull request. Mutation testing on provedex-core is documented in CONTRIBUTING.md.

Performance

Numbers from cargo bench -p provedex-core on an Apple M4 Pro running rustc 1.89.0, criterion default sample size (100), 3-second warmup. Reproduce: cargo bench -p provedex-core.

Operation Median time / event Throughput
canonical_json (one ModelInvoked event) 940 ns 1.06M events/sec
compute_self_hash (canonical-JSON + SHA-256) 2.7 us 366K events/sec
SignedEvent::seal (full sign, no I/O) 11.2 us 89K events/sec
LedgerSession::seal_and_append (sign + append + fsync_data) 3.8 ms 261 events/sec

seal_only isolates the crypto cost. The full append cycle is dominated by fsync_data on every event; customers that batch flushes will see the full-cycle cost approach the seal-only number. Voice agents in the typical 50 RPS regime stay well under the per-event budget either way.

Sidecar HTTP roundtrip

Numbers from bash benchmarks/agent-http/run.sh on the same hardware. Loopback HTTP via oha. Each scenario gets a fresh agent with an empty ledger.

Endpoint Concurrency p50 p95 p99 req/sec
GET /v1/healthz 50 0.91 ms 2.78 ms 7.65 ms 41,456
POST /v1/sign 1 3.98 ms 4.69 ms 5.43 ms 253
POST /v1/sign 10 35.2 ms 39.9 ms 42.8 ms 280
POST /v1/sign 100 353 ms 388 ms 420 ms 283
POST /v1/verify (100 events) 1 2.09 ms 2.18 ms 2.38 ms 476
POST /v1/verify (1k events) 1 3.97 ms 4.12 ms 4.28 ms 251
POST /v1/verify (10k events) 1 10.9 ms 11.4 ms 11.9 ms 90

Reproduce: bash benchmarks/agent-http/run.sh.

The /v1/sign cost is roughly 0.2 ms HTTP overhead on top of in-process seal_and_append (3.8 ms with fsync). Sign throughput plateaus near 280 req/sec regardless of concurrency because fsync_data on every event serializes the writer. The handler offloads the sync work to tokio's blocking pool via spawn_blocking, so other in-flight requests on the runtime are not starved during a sign. Customers that need higher per-key throughput either run multiple agents (one per signing identity) or batch flushes (planned). Verify scales linearly with chain size; verify on a 100k-event ledger takes roughly 100 ms.

/v1/healthz is fast enough for Kubernetes liveness probes at any reasonable cadence (43k req/sec single-agent throughput).

Versioning

Pre-1.0. The public API of provedex-core may change between minor versions until the schema is settled. Any breaking change to the canonical-JSON format, the hashed-field set, or the AgentEvent variants requires:

  • A new docs/spec/ document with a bumped version suffix.
  • A new ADR superseding the affected prior decision.
  • A bump of ExportBundle::schema_version.

Bindings, the sidecar, and the CLI track provedex-core semver.

License

Apache-2.0. See LICENSE.

Contributing

See CONTRIBUTING.md for development setup, commit-message format, and pull request expectations. Security reports go through SECURITY.md.