CAIRN · by fraQtl
Every night, your agents
go to work.
One session · watched live
▸ read src/routes.py4,102 tok
▸ pytest tests/auth1,876 tok
▸ read src/routes.py ← again
▸ pytest tests/auth ← again
▸ read src/routes.py ← again
0
TOKENS PAID FOR REPEATED WORK
The obvious fix is a trap
10:41 ▸ pytest tests/auth → 3 passed
10:52 ▸ … the code changes …
10:53 ▸ cache serves:
"3 passed""3 passed" — STALE
reality: 1 FAILED — and nobody knows.
4,152,027 real agent commands · measured
0%
of the re-reads a cache would have served had actually changed.
97% real content changes. Not timestamps.
So CAIRN never guesses
CERTIFIEDProvably identical
Provenance and output hash both match. Reuse is byte-for-byte correct.
DELTAChanged — serve the diff
Same work, output moved. The agent gets the change, not the whole thing.
BLOCKEDUnprovable — re-run
State drifted or output unobservable. Nothing stale ever reaches the agent.
Every call, sorted — live
CERTIFIED · 12%
DELTA · 20%
BLOCKED · 68%
The red stream is what every other cache would have served stale.
Priced the way due diligence prices it
$0
upper bound — no provider prompt caching
$0
the floor — carried context at provider cache-read rates. We quote this one.
Two prices, both defensible. The false-hit rate is printed on every receipt.
The pool provider caches can't touch
0
across runs
(nightly re-scans)
6.6× more certified-safe reuse — same engine, same proof standard.
And at fleet scale — who spends what
CAIRN CONSOLE · SHADOW MODE · ILLUSTRATIVE FLEET
| dev | tool calls | certified reuse | saved / night |
|---|
| nightly-fleet-04 | 88,204 | 1,412 | $0 |
| pr-review-bot | 64,118 | 948 | $0 |
| monorepo-sweeper | 52,377 | 762 | $0 |
| (8 more) | 208,109 | 1,912 | $0 |
Per user. Per team. Per agent. Nothing leaves your walls.
One command · your own number
MEASURED
$ pip install cairn-security-agent-audit
$ cairn-shadow report --model claude-sonnet-4.5
CAIRN shadow receipt — 1 day, 1,204 tool calls
re-reads: 214 (17.8%)
certified: 38 false hits blocked: 61
est. savings: $2.17 net of provider prompt cache
Certify the reuse.
Audit the spend.
a cairn marks the path that is safe to take.
github.com/fraqtl-ai/cairn-security-agent-audit