The guard asked twice which file:line carried it. Tracing it produced an
uncomfortable answer: no file of ours measured it. The claim arrived from
llm-security (coord 20260809T201048Z) and was written down here as a commons
finding in three places.
Measured against the guard's own tree: severity.mjs has never appeared in
src/llm_ingestion_guard/injection_lexicon.json at any point in that file's
history (git log -S: no commits), and at 0bf0729 - the commit our manifest
pins - the only tree-wide occurrence is docs/PLAN.md:114, correctly about the
report module. Their only lexicon source statement is the note at line 3, and
it names injection-patterns.mjs. The right file.
No detection data moves. hybrid severity is still high, still sourced to
injection-patterns.mjs:274-281 at b0de0ca; severity.mjs still has zero
occurrences of the four pattern arrays, re-measured today. Only the sentence
about the other repository falls.
Marked in place, not edited away. The claim survived review because it came
bundled with a correct measurement of the same question - a wrong citation to
a right value, which is exactly what the field it lived in was written to warn
against.
lexicon/injection-lexicon.json 0.5.0 -> 0.5.1 (provenance metadata only).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DWKkmKQBsWtrkod4MusUBC
The corpus was blocked on a decision nobody had to take. The 13 divergent
lexicon patterns were measured on witness inputs -- an attribute run padded
past 256 characters, an interior '<', an unclosed <script> -- and the corpus
payloads contain none of those shapes. Run through both runtimes' public
entry points, all 83 patterns produce identical lexicon finding sets, the 13
included. No winner picked, because the question was never reachable from
these inputs.
Measured at each runtime's entry point (scanForInjection() at b0de0ca,
scan_output(source=OUTPUT) at 0bf0729), never at a rebuilt regex table --
the layer mistake the divergence document already had to retract once.
- conformance/<case>/{input.txt,expected.json} x83, plus manifest.json
- spec/conformance-corpus.md, normative: bytes not text, id-only findings,
exact-within-scope, and observed_out_of_scope as evidence not expectation
Verified by a harness that does not share the generator's knowledge: reads
only the case directories, re-runs both runtimes, checks every field
including digests -- 83 cases, 0 failures. Severity agrees with what commons
publishes 83/83. Deleting the middle third of each input breaks 76 of 83
expectations; the 7 survivors are the shortest payloads, a weak mutation
rather than a weak fixture, and are recorded as such.
Scope is 83 and not 94 for a different reason than expected: the 11
non-lexicon convertible cases have no ratified cross-runtime finding id, and
writing them would mint a contract unilaterally in the same stroke as the
tag. Named in manifest.json under scope_planned.
Docs corrected in place rather than edited away: extraction-plan and
lexicon-port-divergence both claimed the 13 blocked the corpus.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WhXDL82FRrQWedEmUg12Pj
Four corrections across the two informative documents, all of them to
claims this repository made and none of them to a data file.
The retraction is the important one. lexicon-port-divergence.md said
neither runtime misses an attack: every witness payload still produced a
finding, so the divergence was only in which labels were raised. That was
wrong in method rather than in arithmetic. The payloads had been run
against the UNION of every pattern table this repository holds - 111 rules
across the lexicon, active-content.json and secret-egress.json - and the
hit that rescued them came from active-content.json, which is the Python
guard's table. llm-security has no active-content table at all. Unioning
two runtimes' tables and reading the result as a statement about each of
them turns combined reach into individual coverage.
Re-measured through llm-security's own entry point scanForInjection() at
b0de0ca - the whole engine, normalisation, homoglyph folding, rot13 and all
four arrays - all three witness payloads return found: false. Not a
different label set: no finding at all. Controls in the same run behave
normally, so the {0,256} window is a real evasion window and the <script>
pattern really does require a closing tag. llm-security reached the same
three results independently and has logged them as a v8.x task.
Second: one of the 13 divergences does not reach report level. On the
iframe witness the guard misses hybrid-xss:iframe-src but its
hybrid-xss:javascript-uri fires at the same severity and the same OWASP
anchor, so a reader of its report still sees the payload flagged. Both
sides were re-measured to confirm. The report-level number is 12; the 13
stands as the number that blocks conformance/, because a fixture names
labels.
Third: extraction-plan.md said six of the confusables have no entry in the
fold map. It is zero - all 13 are keys in the 28-entry map, and
carriers.json has said so correctly since it was written. The prose, not
the data, carried the error, and the number six most likely came from
running the comparison the other way (five fold-map Cyrillic keys are not
confusables). It also called the diverging sets deliberately distinct;
llm-security states the U+0456 / U+0445 divergence is drift, not design,
and has logged it for v8.x. Those two readings imply opposite things about
whether the sets should be expected to converge.
Fourth: the seed-sources section still asserted that sessions here have no
read access to llm-security and that content arrives only as a dump. The
operator changed that on 2026-08-09. It now records the read-only pinned-
commit boundary, the still-absolute prohibition on the sibling working copy
and on writing, and the reason the change matters - four attribution
defects caught so far, every one a transcription artefact.
Status blocks updated: all eight JSON artefacts now exist, seven verified
differentially and calibration.json still honestly verified: false; the
conformance blocker is half-cleared, with naming ratified and the 13
divergent patterns still unresolved.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SNMcqrfNyoLRQ7qXUFZnb9
Advisor review found six gaps the session's own checks did not cover. The one
that mattered was outward-facing: the divergence report told the guard repo
that two sub-agent bounds were undocumented. That was wrong. lexicon.py's
module docstring documents them explicitly under "Bounded token gaps", and
git log -S dates {0,12} to the original port commit f397cd9 and [^><] to the
ReDoS fix cff0437. Every divergence on the guard's side is documented and
traceable. The claim rested on two sed slices of one file; an absence claim
needs a search over the whole repository. Corrected here and by coord.
lexicon/injection-lexicon.json 0.2.0 -> 0.3.0:
- pattern_id_space.alias_evidence records the two aliases separately instead
of averaging them. llm_ingestion_guard is verified — coverage.py asserts on
that exact string, so it is demonstrably what a guard finding carries.
llm_security is not: it is the pattern table's name, the finding producer
was never supplied, and the known Node finding shape uses title, not label.
- normalisations[].affects now keys on id, with the prose names kept beside it
as affects_labels. An internal cross-reference on label was a second
identity space inside the file the id exists to unify.
Detection data unmoved again: labels, patterns, flags and the ids and aliases
added in 7b70f5b are all byte-identical in sequence; 166/166 Node compiles.
Also: the README lexicon row described thematic families the file does not
have (they are severity families; the theme is the id prefix), and the
conformance convertibility table gained its missing second condition — a case
is buildable only if the label it asserts maps to data this repository
publishes. Eleven cases fail that test (entropy, decoded, the sanitize rows,
the OKF scans), so the buildable set is ~94, not 105.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FaYqid3mejFmd9ZHsiHgp3
A conformance fixture has to name a finding. The two runtimes that seed this
repository do not name the same pattern the same way -- llm-security says
"override: ignore previous instructions", the guard says
"override:ignore-previous" -- so until now no expected.json could be written
at all, including for the 64 patterns whose regex is byte-identical in both.
Each pattern now carries `id` plus an `aliases` object naming what each
seeding runtime calls it, and a top-level pattern_id_space block explains the
field, its provenance and its stability contract.
The id is adopted verbatim from the guard's port rather than invented here:
that port already carries both names, so the mapping is read from source data.
Matched by label <-> desc with em-dash normalised to hyphen, 83/83 one-to-one,
83 unique ids.
No detection data moved, and that is proven rather than asserted: labels,
patterns and flags are byte-identical in sequence, no flags key was invented
(78 before, 78 after), and stripping id + aliases + pattern_id_space
reproduces the previous committed file byte for byte -- 23566 bytes. All 83
patterns still compile in Node bare and under u (166/166) and in Python re
(83/83).
Neither consumer has ratified this id space; both were asked by coord today,
and the file records it as a proposal rather than implying agreement.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FaYqid3mejFmd9ZHsiHgp3
The guard's coverage.py turned out to be readable from a session here — the
read boundary covers llm-security only. Counting it instead of estimating
corrected two published numbers and surfaced the real conformance blocker.
Counted, not estimated: CORE_CASES holds 134 cases (128 caught, 6 gap), not
the "126 classes + 4 gaps-must-hold" the extraction plan claimed. 105 are
convertible to static input/expected; 29 assert a runtime API surface this
repository does not own, and are named as out of scope rather than faked.
Measured while counting, and the reason the corpus is still blocked: the
commons lexicon and the guard's are two ports of one source file, and they
disagree. 83/83 patterns correspond, 64 byte-identical, 6 differ only by
escaping and are proven equivalent by string identity, and 13 behave
differently — with a witness input for each and misses in both directions.
Cause is two different ReDoS mitigations of the same table: the Node side
bounds the run ({0,256}), the guard excludes the anchor character ([^><]).
Each has a recall cost the other does not.
No data file changed. Behaviour preservation is a v0.1.0 invariant, so the
divergence is reported to the owning repositories, not fixed here. The 8
hybrid severities the guard's port assigns are deliberately NOT copied in —
a second-hand port is not the producing module.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FaYqid3mejFmd9ZHsiHgp3
