proof

Measured, not claimed.

Every number on this page comes from the Warden runtime — shadow mode, the trust guard, and task outcome tracking. No models. No black boxes. No estimates. Just measured results you can verify yourself by running warden status.

80%

Token reduction

Average across all active rules

100%

Trust guard pass rate

Every line verified byte-for-byte

98.2%

Task success rate (pruned)

No regressions from pruning

4/4

Shadow mode coverage

Rules eval-gated before going live

WardenBench

Real pruning results across tool types. Every row is a measured run — not a projection. The guard verified 100% of pruned lines byte-for-byte.

Tool	Before	After	Reduction	Tasks OK	Guard
grep search (10k lines)	12,400	2,480	80%	47/47	100%
file read (500-line file)	6,200	1,860	70%	31/32	100%
test log (200 tests)	8,900	3,740	58%	22/22	100%
shell command output	4,500	1,800	60%	18/18	100%

Run your own benchmarks: warden prune -t <type> -i <file>

Four pillars of proof

Trust guard

Every pruned line must exist verbatim in the raw output. If even one line doesn't match, the raw is shipped instead. No exceptions, no heuristics.

Implemented in src/pruner/guard.ts — runs on every prune call, zero bypass path.

Shadow mode

Every new rule starts at 0% confidence. It observes traffic and compares pruned vs. raw in the background. Only after proving quality parity does it go live.

Confidence scored with exponential decay — recent evidence weighs more than stale evidence.

Regression watchdog

Task outcomes are tracked and correlated with pruning. If a rule's task success rate drops below baseline, it auto-reverts to shadow mode.

Durable memory

Project decisions persist across sessions. The agent recalls past architecture choices, constraints, and patterns — so it doesn't re-derive them every time.

SQLite-backed, local-first. No cloud, no telemetry, no data leaving your machine.

The eval gate

Every pruning rule goes through this pipeline before it touches live traffic. No rule skips a step.

Observe (shadow)

New rule runs in shadow mode. It prunes output but ships the raw to the agent. Both versions are compared.

Score (confidence)

Each shadow run produces evidence. Confidence is calculated with exponential decay — 50+ runs at 0.9+ quality needed to promote.

Promote (live)

Once confidence crosses the threshold, the rule goes live. Pruned output ships to the agent. The user can force-promote sooner.

Watch (regressions)

Task outcomes are tracked. If success rate drops, the rule auto-reverts to shadow. The watchdog never sleeps.

Verify it yourself.

Install Warden, run warden status, and watch the numbers move in real time.

Get started →