⊤
FAULTLINE
· NEWS
The engineering journal.
Measured numbers, security disclosures, and fixes —
every entry with its receipt.
LEGEND
Incident
?
Measurement
?
Deployment
?
Improvement
?
FILTER
All severities
All dates
SEVERITY
↕
TITLE
↕
BRIEF
↕
DATE
↕
← back to all entries
🔗 SHARE
News entries — plain markdown, no JavaScript required (AI/crawler readable)
The Fair Litmus: three task families, three passes - the cortex uplift, measured honestly
(2026-11-17)
The Self-Built Conscience: a 35B model constructed its own workshop and solved the task that defeated it ten times
(2026-11-17)
The Cortex Uplift Experiment: a 35B model, a corrupted database, and the wall we finally mapped
(2026-11-17)
Terminal-bench round five: the first honest grading, the csv flip, and the openssl counter-flip — variance in one picture
(2026-11-15)
Terminal-bench A/B, round four: the template fix worked (parroting 42% to 7%), the corpus didn't move resolution, and four rounds deliver the instrument's verdict
(2026-11-15)
Incident disclosure: the experimenter broke the scaffolding too — our own failure-mode ledger from the three-round agent benchmark
(2026-11-15)
Terminal-bench A/B, round three: the uplifted seat, a null result, and what three rounds actually measured
(2026-11-15)
The vector question: what FaultLine's memory stores, and why
(2026-11-15)
The convergence gate: corrections now land, every time
(2026-11-15)
Terminal-bench A/B, round one: a cold seat, five honest empties, and a 33%→71% delta we cannot credit to memory
(2026-11-15)
Terminal-bench A/B, round two: mini-swe-agent, a seeded seat, and memory that actually carried content
(2026-11-15)
Documents are now a store of record: versioned, byte-returnable, superseded like facts
(2026-11-14)
Recall now answers plainly
(2026-11-14)
Document ingest: snapshot-first storage — ingest can no longer clog serving
(2026-11-14)
Serve latency: a stage-by-stage decomposition and a 38% reduction
(2026-11-14)
R4-K: a pathological knowledge shape could pin a CPU worker indefinitely
(2026-11-14)
The redaction fix: “remove this” now actually removes this
(2026-11-14)
Incident disclosure: a heavy seat answered “I know nothing about your family” — while holding all of it
(2026-11-14)
The heavy-seat walk fix: when 627 connections per question was the latency
(2026-11-14)
The blended-family recall round: 53% to 97%, measured run by run
(2026-11-13)
The recall overhaul: the numbers, the methodology, and the change log
(2026-11-12)