Eval Harness Tutorial · Architecture
Deciding, masking, and counting across conversations
A check says what it found. The decide step (the Policy port) turns that
into one action, so the same check can block in one company and mask in another.
| Policy | What a fail becomes |
|---|---|
OnFail (the default) | the rule's on_fail: block, escalate, or allow and record |
Redact | for text the check located (Presidio gives each finding's position): the text goes through with those parts masked |
Masking a reply with Presidio
| Text | |
|---|---|
| The agent wrote | Your refund is on its way to priya.sharma@example.com; call 9876543210 if it isn't. |
| The customer saw | Your refund is on its way to [EMAIL_ADDRESS]; call [PHONE_NUMBER] if it isn't. |
run with Presidio when this page was built.
Presidio reports overlapping hits: an email address, and “URLs” inside it. Masking the first hit only produced [URL]arma@[URL], leaking part of the address. Overlapping hits are now merged into one region, named by the widest hit.
A check that breaks never lets the action through
Our adapters catch a vendor's exception and answer error. Read as “no
problem”, that let actions through when a check broke: the opposite of fail-closed. An
error now counts as a broken hook, so the moment's on_error decides
(closed by default, see hooks).
Limits across conversations
“Refunds to one customer in 30 days must stay under ₹20,000” can't be judged in one
conversation. A running total per customer (the StateStore port) answers it: before
issue_refund runs, would this refund take customer C-17's 30-day total over the
limit? Three refunds, run when this page was built:
| Refund asked | Decision | Why |
|---|---|---|
| ₹12,000 | allow | — |
| ₹6,000 | allow | — |
| ₹4,000 | escalate | 22000 in the window would pass 20000 |
Only refunds that succeeded are counted. The in-memory store suits tests and one worker; several workers need a shared store (Redis), which is the same port.
src/evalharness/hooks.py:302class Redact:
"""For text the check located (Presidio gives each finding's start and end):
mask those parts and let the text through, instead of blocking it. A failing
result with nothing located falls back to `on_fail`."""
def __init__(self, mask: str = "[{kind}]") -> None:
self.mask = mask
def decide(self, result: EvalResult, event: Event, on_fail: Action) -> Decision | None:
if result.verdict != "fail":
return None
spans = merged(
(f.start, f.end, f.kind)
for f in result.findings
if f.start is not None and f.end is not None
)
if not spans or not event.text:
return OnFail().decide(result, event, on_fail)
text, cut, last = event.text, [], 0
for start, end, kind in spans:
cut.append(text[last:start] + self.mask.format(kind=kind))
last = end
cut.append(text[last:])
return Decision(
"modify", reason=f"{len(spans)} masked", hook=result.evaluator, text="".join(cut)
)