Evidence you can inspect before installing.

This viewer turns a checked-in KAI OS run into the product surface: process table, syscall ledger, replayable capsule, and baseline gate evidence.

If this local-first evidence layer should exist for JVM/Kotlin agents, star the repo and try the no-key tour when you are ready.

Star / Fork on GitHub Open proof pack Open sample artifacts Why star KAI OS Start here

One run, five evidence signals.

Source: examples/evidence-sample, generated from kaios review against a tiny Kotlin workspace change.

Run run-9c888ffe
Agent processes 3
Total tokens 582
Syscalls 3
Replay valid

Agent as process.

The run leaves process-style rows instead of an opaque transcript.

PID Agent State Tokens Context Syscalls Worker
1 planner SUCCEEDED 145 1131b 1 in-process
2 executor SUCCEEDED 192 1550b 1 in-process
3 validator SUCCEEDED 245 1919b 1 in-process

Tools as syscalls.

Tool use is recorded as a syscall ledger with permission, pid, duration, cost, redacted args, and denial status.

Call PID Agent Tool Permission Status Redacted args
sys-1 1 planner echo ECHO allowed message=planning:5f3f275e
sys-2 2 executor mock-http NETWORK allowed url=mock://kaios/tasks/15a7b387
sys-3 3 validator echo ECHO allowed message=validated:0edf3419

Capsule, replay, and contract status.

The same run is packaged so another machine can validate and replay it without calling a model provider.

Capsule contract

kaios.run-capsule/v1

Validation status is valid; the capsule embeds the snapshot, process trace, provenance hashes, and replay commands.

Offline replay

deterministic

Replay rebuilds the trace from the embedded snapshot and checks it against the saved trace.

Try it locally after installing the CLI:

kaios capsule --file examples/evidence-sample/change-review.capsule.json --check
kaios replay --file examples/evidence-sample/change-review.capsule.json

Baseline gate catches behavior drift.

Valid capsules can still differ. KAI OS compares stable runtime signatures and can fail CI on changed behavior.

Diff result

different

kaios.run-diff/v1 reports same=false and exits 1 under --check.

Stable deltas

  • token total: +216
  • input tokens: +216
  • context bytes: +3060
  • changed fields: task, output hash, metrics, processes
./scripts/evidence-samples-smoke.sh
kaios diff examples/baseline-gate/capsules/baseline.capsule.json \
  examples/baseline-gate/capsules/current-different.capsule.json \
  --check

Open the underlying artifacts.

Every panel above points back to checked-in files rather than hosted telemetry.

Review artifact Human-readable Markdown output from kaios review. Process trace kaios.process-trace/v1 with processes, scheduler, syscalls, cost, and events. Run capsule kaios.run-capsule/v1 with embedded snapshot and replay commands. Baseline diff Normalized stable diff used by the smoke test.

Want this Evidence OS direction to exist?

Star or watch KAI OS on GitHub, then use the no-key tour to see the same process table, syscall ledger, replay capsule, and baseline gate from your terminal.

Star / Fork Proof pack Evaluate in 5 minutes