You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Proposal] §6.3: normative disposition for an unscoreable MODIFY composition (fail-closed redaction semantics) #296
A conformance vector cannot be scored because the wire-side disposition for an
invalid or unscoreable MODIFY composition is not normatively specified. The
quarantine record that surfaced it lives on #33; this issue is the normative
question its closure would resolve.
The vector (canonical input SHA-256 01410bfff073d3835775373ec63be2418378de41aae00d5e6a508689079abb5f, vector_id01410bfff073d383) is a steps/userMessage request carrying a
content string and modify: {redact: ["card"]} — a PII redaction request.
When the suite asks "what disposition does a conformant Guardian return for
this?", the answer is under-specified: §6.3 tells an Observed Agent how to
compose a valid modifications object and that an invalid one must fail
closed as DENY, but it does not state the normative expected disposition the
conformance suite should assert for a redaction composition it cannot itself
validate on the wire. Is the expected outcome MODIFY with the redacted
content, DENY (fail-closed), ASK, or is the vector itself not in scope for
a wire-level conformance claim? Without that decision the row cannot be scored
pass or fail.
What this issue is not
Not §6.3 needs the same no-DENY hook exception that §6.5 has for postCompact #270: that issue asks what an implementation must do at the five
hooks where DENY is not a legal disposition when it receives an invalid modifications object. This issue asks the normative decision the conformance
suite asserts about redaction compositions — a different row.
A normative disposition (or an explicit "not a wire-level claim") for a
MODIFY composition the requesting party cannot validate — e.g. a redaction
request whose redact field the Guardian cannot confirm is well-formed.
The expected reason code(s) on that disposition, if any, drawn from the
§17.2 registry.
Whether the conformance suite should score such vectors at all, or whether
they belong in a separate "composition-validity" bucket.
Opening this so the conformance ledger has a row whose closure resolves the
vector's uncertainty; happy to supply the wire transcript or a minimal fixture
if useful for triage.
Activity