Structural analysis: findings that live in the relationship between two places in a prompt, which no phrase dictionary can see because neither place is wrong on its own. Both are advisory — Trazum points, it does not cut.
- Fixes a corruption bug in
duplicate-lines. The rule was deleting the sharedOutput:line from a second few-shot example, leaving it with an input and no output. Two examples mapping different inputs to the same answer is often exactly why both are there. Labelled example fields (Input:,Output:,Q:,A:, and Spanish equivalents) are now exempt from line deduplication. This affected thesafelevel, so it could silently damage a prompt anyone ran through Trazum. - New
contradictory-instructionsadvisory across four axes: response language, output format, response length, and whether to show the reasoning. Reported as a warning with both conflicting sentences quoted. It carries no dollar figure — being wrong has no price tag. - New
redundant-examplesadvisory: few-shot examples that are near-copies of an earlier one, with the tokens they cost per month. It detects copy-paste accumulation (~0.89 similarity for a copied example with one field changed), and deliberately not paraphrases (~0.54), which sit too close to genuinely distinct examples (~0.20) to separate without a model. - Advisories now sort by severity before money. Sorting purely on the dollar figure buried an overflowing context window — and now a contradiction — underneath a saving of a few dollars.
- New public API:
findContradictions,analyzeExamples,findExamples, and thejaccard/normalizeForComparesimilarity helpers, which moved to a shared module so the duplicate rules and the structural analysis cannot disagree about what "near-duplicate" means. - Adding a contradiction axis now fails to compile until every catalogue names
it, the same guarantee
RuleIdgives rules. - Tests grow from 75 to 94.
Security#
Hardening for an open repository taking outside contributions. Full reasoning in SECURITY.md.
- Fixes four SSRF filter bypasses. The web app's private-host blocklist
allowed
https://[::ffff:169.254.169.254]— the IPv4-mapped IPv6 form of the cloud metadata address, which Node normalises to[::ffff:a9fe:a9fe]and the old patterns did not match. Also allowed: a trailing-dot hostname (localhost.), the carrier-grade NAT range (100.64.0.0/10), and credentials embedded in the URL, which would have been forwarded to whatever the host resolved to and written into any log recording the endpoint. - The filter moved from the Next.js route into
@trazum/coreasvalidateLlmEndpoint/isPrivateHost, so the most security-sensitive code in the project is unit-tested instead of living untested in an API handler. It returns a reason code rather than a message, so callers localise it and tests assert on the decision. - Fixes two ReDoS denial-of-service bugs, both reachable from the public
HTTP endpoint, both found by CodeQL after the first round of ReDoS tests had
passed:
whitespace— asafe-level rule present since 0.1.0. Its trailing-whitespace pattern restarted at every position inside a whitespace run and failed from each one when the run did not end the line: 17 seconds on a 100 KB line of spaces, well inside the 400 KB the API accepts. Anchored to the start of a run, it is now 3 ms at 400 KB.- The few-shot label patterns added in this release ended in three adjacent unbounded quantifiers, measured at O(n²) — 651 ms at 40 000 spaces, about a minute at the size cap. Their quantifiers are now bounded.
- The ReDoS suite gained the fixture shape it was missing. The original fixtures were all repeated tokens, which exercise the happy path over and over; neither bug needed that, they needed a plausible prefix followed by a long run that never completes the match.
- New
security.test.jsenforcing four invariants on every pull request: the SSRF filter fails closed, the core and CLI carry zero runtime dependencies,fetchappears only in the two modules that exist to make calls, and no regex exhibits catastrophic backtracking under pathological input. - Workflows run with
permissions: contents: readby default,npm ci --ignore-scripts, andpersist-credentials: false. - Added CodeQL (
security-extended), dependency review, a weeklynpm audit, Dependabot,CODEOWNERS, and an importable branch ruleset at.github/rulesets/main-branch.json. SECURITY.mddocuments the threat model, private reporting, the settings an admin still has to switch on, and the limits that are not covered — DNS rebinding, per-instance rate limiting, and actions pinned to tags.