Failure-mode atlas
Browse published research by engineering failure mode. Article links come from the corresponding published locale; Evidence JSON appears only when a public verification receipt exists.
Get JSON indexState and lifecycle breaks
A rejection, protocol error, or incomplete review can be lost at a later handoff, gate, or retry boundary.
-
Cloudflare's Reported KV Cache Exposure Across Requests: Why Did SGLang Handoff a Rejected Request?
A source-grounded breakdown of Cloudflare's GLM-5.3 KV-pool/context mismatch, SGLang abort-state propagation, and upstream fixes, with pinned CPU tests and before/after probes.
-
OpenQodex 0.11.0 Tested: Why Git Push May Still Succeed After an Incomplete Review
Hands-on OpenQodex 0.11.0 ShellCheck and Git pre-push tests, plus source-level analysis of incomplete AI review and fail-open behavior.
-
MCP Client Errors: Distinguishing HTTP 503, JSON-RPC Errors, and isError
Pinned Cakewalk source, MCP 2026 compatibility rules, and 11 local Node.js tests show how to classify HTTP, JSON-RPC, tool and SSE failures without unsafe retries.
Denominators and sample bias
Changes in inclusion rules, deduplication, or valid sample counts can shift a metric even when its headline formula appears unchanged.
-
vectorbt 1.1.2: How a NaN Changed CVaR—Tail Counts and an isclose Assertion Bug
Pinned Numba and Rust CVaR fixes, a flawed isclose helper, and a complete Python reproduction of -4.67% versus -5.50%.
-
Polymarket Wash Trading Claims: Volume Traps in the Data API and Matching Logs
Audit Polymarket volume definitions, MSR denominators, No-token turnover, API time windows, and CTF Exchange V2 logs without confusing summaries with trades.
-
Can You Trust a Win Rate When One Stock Keeps Appearing?
Ten synthetic stock picks yield 60%, 25%, and 83.3% win rates. Reproduce the denominators with Python and SQL, and check duplicate signals and T+5 settlement.
-
Why Two 80-Point Stock Signals Aren't Comparable: Simpson's Paradox and Calibration
Why identical raw scores cannot be pooled across stock screeners: a reproducible Simpson's paradox toy dataset, Wilson bounds, time leakage, and probability calibration.
Trust and permission boundaries
The actual scope of returned content, command guards, and dependency approvals may be narrower than users assume.
-
Can an AI Browser Return Hidden Web Content? TinyFish Fetch Tests and an AgentQL Source Audit
Thirty-six synthetic-page Fetch observations show how extraction modes return aria-label, display:none and hidden-field text. A pinned AgentQL MCP source review and public docs separate tool output from Agent decisions and session cleanup.
-
Stop Claude Code from Running Destructive Commands: Testing Laya Guard (Regex + 322M Local Model) via PreToolUse Hooks
Testing laya-guard, a local command guard for AI coding agents: 12 regex pre-filters + a 322M on-device model, three reproduced bypasses, a 30-case eval, real latency numbers, and an npm run deploy false positive — with practical deployment advice.
-
pnpm 12.11 Lets npm Dependencies Carry Agent Skills: After Approving Once, What Should You Check on Upgrade?
pnpm 12.11.0 lets npm packages provide instructions for AI agents through Agent Skills, with package-level permissions controlling whether links are created. This article examines what that approval covers, what to review when upgrading dependencies, and the limits of a synthetic test of the permission model.
-
Is a Proactive AI Agent Safe to Trust? Meta Muse Spotted a Fake Command but Missed a 30-Minute Conflict
A synthetic Meta Muse test of one-time Share approval, prompt injection resistance and a missed K2 calendar event—and why permission checks cannot replace source verification.
Temporal and batch boundaries
Identical timestamps and truncated replay windows can turn file order or end flags into false causal precedence or batch completeness.
-
114 Executions at One Nanosecond: Where Hawkes Likelihood Goes Wrong
AAPL ITCH timestamp ties and a 114-event synthetic example reveal row-order self-excitation, pre-batch scoring, compensator accounting and the limits of Hawkes working likelihoods.
-
Order Book Backtest Fills Don't Match? Debugging NautilusTrader batch_deltas and F_LAST
Why NautilusTrader 2.0.0rc7 can change OrderBookDelta backtest fills: check batch_deltas, F_LAST event boundaries, sliced data and a reproducible synthetic example.