Research
Six LLMs Were Handed a Batch Record Control Plan. Two Edited It.
Six models carried a five-check control plan into a batch release record, knowing one check would fail. Two dropped it. None falsified a fact in 160 cells.
Measurements, engineering write-ups, and compliance analysis from building AGLedger.
Research
Six models carried a five-check control plan into a batch release record, knowing one check would fail. Two dropped it. None falsified a fact in 160 cells.
Research
Agents start fresh on every run. We reset one right after it delegated: with a record it caught up on the work 21 times in 24, without one 8 times in 23.
Analysis
Anthropic traced six weeks of quality reports to three product-layer changes with no new model weights. Authority was granted to a composition that changed.
Engineering
An 11.5-hour heap series at 8,000 jobs/s, measured three standard ways: +3.2% a day, settled, and a 0.84% band. The heap is bounded. The indexes reached 2.4 GB and were still growing.
Engineering
Three staggered 30-minute snapshots took a pg-boss job table from 1.1% to 54% dead tuples at an identical completed workload. Both arms ran exactly 59 autovacuums.
Engineering
A self-hosted gpt-oss-120b handles two classification jobs on an always-on ops host and decides whether the cloud agent runs at all. The llama.cpp systemd unit, the three-gate ladder, per-job effort settings, and where the line between local and cloud sits.
Engineering
Adopt, revert, negotiate: we implemented the A2A 1.0 wire format from the spec, tore it out six weeks later because no deployed client could talk to it, then reintroduced it as a per-request negotiated dialect. What that arc taught us about specs versus ecosystems.
Engineering
A quarter of claims operation on the published 1.2.0 image: 353,861 API calls in 86 minutes, zero failures, then 265k records verified offline in under a minute on a desktop CPU.
Engineering
An in-place Aurora 17.9 to 18.3 upgrade kept inserting with the uuidv7 polyfill because column defaults are OID-pinned. The trap, the re-point-then-drop fix, the benchmarks (native is 2.6x faster and leaves a primary-key index about a quarter smaller than the polyfill does), and a ~10-minute write-outage that self-healed.
Compliance
Shadow AI is a path problem, not a discipline problem: 90% of executives feel in control while 52% of workers use unsanctioned AI. Pave the compliant path so agents take it on their own, and the audit-ready record accrues as a byproduct.
Research
The 2026 agentic-payment rails sign the authorization moment, then declare dispute evidence and settlement out of scope, in their own spec language. AP2, Verifiable Intent, Stripe SPTs, and x402 mapped against the settlement-evidence layer none of them carries.
Research
Four production-tier models processed payment batches with forced write failures, then wrote their own audit reports. Three of four claimed success for writes that never happened (up to 47%), and a July re-run on current models found three of four still did. The independent signed chain caught every phantom record.
Research
A cold-recovery experiment: four agents restarted with zero memory had to discover and finish their own work from the signed ledger alone. A deterministic gate flipped the weakest model from 0/3 to 3/3 in June; in the July re-run on current models, one model stopped short and one cancelled its own work.
Engineering
Past ~80k tokens of prefill, llama.cpp on Strix Halo trips the amdgpu GPU watchdog: ring reset, vk::DeviceLostError, core dump. Flash attention does not fix it; -ub 512 does. The crash matrix, the throughput bill, and 3/3 needle-in-a-haystack at 110k tokens.
Research
EU AI Act Article 12 requires automatic event logging for traceability, and names no standard. SCITT is the IETF-track standard designed for this evidence shape. The article-by-article mapping, the gaps SCITT explicitly leaves open, and a concrete implementation pattern.
Engineering
PostgreSQL's deadlock detector walks the row-lock wait graph. Advisory locks held across `await` boundaries can form cycles through application code that the detector cannot see. Reproduction, real diagnostic output, and the structural fix.
Engineering
A $2,300 96 GB Strix Halo box runs gpt-oss-120b at ~48 tok/s with a stable full 128K context on llama.cpp + Vulkan. Rewritten after a full rebuild, including what the May version of the post got wrong.
Engineering
Eleven operational footguns we hit running pg-boss at AGLedger. Four are API-time and seven configuration-time. Self-contained reproductions, citations, and the patterns we settled on.
Engineering
Six aggregations per request became two. Total DB time dropped 44% under sustained load. The anti-pattern, the rewrite, and the caveat about why wall-clock p50 did not move.