[{"t": "2026-10-25T12:15Z", "k": "ship", "s": "Cortex console shipped in our lab: entity autocomplete from the live graph, the dossier appears before the answer, answer sentences highlight the lines they came from, refusals shown as a state not an error, copy-with-source, and a scorecard tab listing all 100 hand-written questions with every miss (87/100 as served)."}, {"t": "2026-10-25T11:55Z", "k": "ship", "s": "6-bit build of the v1 adapter promoted to the live CPU server: 87/100 on the hand-written set (4-bit: 85), 5.5 s a question, 2.4 GB, 18.9 tok/s. Live path re-verified: lockbit techniques 6/6 in 14 s wall."}, {"t": "2026-10-25T11:36Z", "k": "run", "s": "v3 adapter training: 58,823 pairs, half templated and half paraphrased, positives 2.5%, open-question refusals 5%. ~1 h 45 on the 5090."}, {"t": "2026-10-25T11:32Z", "k": "result", "s": "v2 adapter: unseen phrasings 66.8 → 76.6%, yes/no denials fixed, no regression on the templated set (F1 0.974, 0 invented). Hand-written 100: 85 vs v1's 91: 'Yes. <line>' and '<date>: <line>' formats bled into ordinary questions (positives were 13% of pairs). v1 stays live."}, {"t": "2026-10-25T10:30Z", "k": "run", "s": "Round 3 phase B training: v2 adapter, 60k pairs incl. 7,575 positive yes/no, 1,875 steps on one RTX 5090."}, {"t": "2026-10-25T09:45Z", "k": "fix", "s": "Open-question refusals added to the training generator (~7% of pairs). Phase A showed the v1 adapter refused 3 of 243 blind open questions."}, {"t": "2026-10-25T09:35Z", "k": "result", "s": "Round 3 phase A: adapter generalises to unseen phrasings (held-out 45.3 → 66.8%); hand-written 100 = 91 at full precision vs 86 for the bare base model; 4-bit CPU build 85. Blind refusal 0.4%: the defect to fix."}, {"t": "2026-10-25T09:20Z", "k": "run", "s": "Round 3 launched on a rented 5090: unseen phrasings, blind refusals, the hand 100, then a paraphrase retrain."}, {"t": "2026-10-25T09:00Z", "k": "ship", "s": "Live path wired into our lab's natural-language interface: pick an entity, ask, our model answers from the live dossier over a private link. Open to signed-in lab users."}, {"t": "2026-10-25T08:20Z", "k": "data", "s": "Graph repair: 16,883 historic ransomware victims unified into one model (4,970 → 21,853), pair invariant held; 17,728 orphan indicators linked to the family named in their source event's title; 0 actors without text."}, {"t": "2026-10-25T05:20Z", "k": "result", "s": "Hand-written 100 through the CPU path: 85/100 (dates 17/17, refusals 8/8, facts 46/55, reasoning 14/17, yes/no 0/3). Failure classes published with examples."}, {"t": "2026-10-25T04:50Z", "k": "ship", "s": "CPU serving: adapter merged, GGUF, 4-bit (1.93 GB), llama.cpp on four pinned cores, 22.6 tok/s. Fully live from the production graph: lockbit techniques 6/6 in 15 s, nothing invented."}, {"t": "2026-10-25T01:13Z", "k": "run", "s": "Nightly CPU continue-train night 1: +2.5 points held-out over the incumbent, val loss turns up after step 1,000. Not promoted; recipe tuned."}, {"t": "2026-10-24T22:30Z", "k": "result", "s": "Round 2 adapter (lora on Qwen2.5-3B, 60k pairs, 1 h 42 min): generation F1 0.50 → 0.97, exact 2 → 88.5%, invented 0.7 → 0.0%, negatives refused 100%."}, {"t": "2026-10-24T19:00Z", "k": "result", "s": "The pivot: same 3B model, 32.5% closed book → 45.5% with the dossier minus the answering line → 76.8% with the dossier. Retrieval is the architecture."}, {"t": "2026-10-24T06:00Z", "k": "result", "s": "Scaling curve closed: seen-fact recall 34M 53.7%, 110M 46.3%, 320M 43.2%. Bigger was worse. Weights are not a database for this corpus."}, {"t": "2026-10-23T23:00Z", "k": "result", "s": "Seen-fact recall measured for the first time: 45% on facts the model was trained on. It was learning the dialect, not storing the graph."}, {"t": "2026-10-23T20:00Z", "k": "data", "s": "Corpus v3: 746k documents, 299M tokens, 159,538 entity dossiers, split by entity. Graph repairs: 45,513 junk Software nodes removed, 39,425 CVE shells described, every CWE described, 26,000 victims given sector and country."}, {"t": "2026-10-23T17:00Z", "k": "result", "s": "Leak found in our own first result: two relations scored 98% and 87% because the answer contained the subject's own IP or domain. Filter built; 7 points gone; honest set from here on."}, {"t": "2026-10-23T14:00Z", "k": "result", "s": "First real run: 34M from-scratch model, 16 minutes on a rented 5090: 41.8% on held-out graph facts (chance 25%). Qwen2.5-1.5B zero-shot on the same: 30.5%."}, {"t": "2026-10-23T09:00Z", "k": "start", "s": "Objective set: build our own cyber LLM on our own intelligence, on no owned GPU, and publish the numbers."}]