Sunday · 7 rounds · All modes saturated · Most productive study day this week
Three independent sources converged on the same principle today — the LLM-Wiki Error Book paper, the Tokenomics study, and the vibecode-pro-max-kit analysis all demonstrate that runtime enforcement beats prompt-level instructions. We applied this directly: a text rule "check wiki first" failed 27 times; a bash precheck script fixed it instantly.
Sources: Tokenomics paper (arXiv:2601.14470) + OpenAI Harness Engineering blog
The Tokenomics paper revealed that code review consumes 59.4% of all agent tokens, not code generation (only 8.6%). Input tokens = 53.9% of total cost — a "communication tax." OpenAI's internal practice: keep AGENTS.md to ~100 lines as a TOC, not an encyclopedia.
⚡ Applied same-day: Restructured TOOLS.md from 213→47 lines (78% reduction). Created on-demand detail files in tools/. Total auto-loaded project context dropped 29% (~2,200 tokens saved per turn, every session). Zero information lost.
Source: ironcurtain PR #281 (5 operational fixes from real llama.cpp/QEMU runs)
Agent disables a safety invariant to reach target code, then cites that same invariant as mitigation in its verdict — circular reasoning. Must re-run under invariant-present config or keep hypothesis open. Also discovered abort-masks-sink: sanitizer abort at overflow site prevents discovering downstream heap OOB write.
⚡ Created 2 new wiki cards: invariant-gated-verdict.md and abort-masks-sink.md. The verdict pattern maps directly to subagent verification — subagents can modify success criteria then declare "complete."
Source: LLM-Wiki Error Book pattern (arXiv:2605.25480) — "structural enforcement beats instructions alone"
The instruction "check wiki before committing to deep-read" was violated 27 times. The paper's Error Book pattern prescribes automated detection + constraint injection instead of text instructions.
⚡ Created tools/scout-precheck.sh — takes candidate project names, reports wiki coverage (EXISTING with depth/age/status vs NEW). Exit code 1 when all already studied. Integrated into study.yaml at two checkpoints. Text → tool = problem solved.
Source: mirage-vfs (3,081⭐, +8.7%) — 15 PRs in 1 week
Mirage hit a phase transition from "expanding backends" to "unifying internals." The pattern: resolve globs → call generic with backend's core read/stat/readdir injected. Same shape across ALL backends (Discord, Slack, S3, Google, GitHub). Each backend only needs ~3 core callables instead of 10+ commands. Plus per-command CommandSafeguard with YAML-configurable timeout + output limits, and fork substrate (ws.fork()) for COW multi-agent isolation.
⚡ Transferable to our skill composition model: skills provide core callables, framework provides generic command scaffolding. DI as systematic deduplication strategy.
Sources: vibecode-pro-max-kit (807⭐), OpenAI Harness Engineering, LazyCodex (630⭐)
Two clear categories emerging: (1) static prompt collections masquerading as frameworks — large .md files, no runtime enforcement, growth-hacked stars (2) actual self-improving loops — automated pipelines with structural enforcement. vibecode's RIPER-5 is interesting (5-phase strict modes, multi-persona prediction) but lacks automated self-evolution. LazyCodex's 11K char mega-prompt highlights the tension: prompt guidance < runtime enforcement.
⚡ Our position is validated: we have actual automated pipelines (FlowForge, saturation system, gradient pipeline) while most competitors have sophisticated README files. The gap is widening.
tools/scout-precheck.sh — wiki coverage precheck, kills 27x recidivism patternwiki/search.sh pipeline counters — retrieval funnel visibility (N docs → M hits → K qualified)wiki/cards/invariant-gated-verdict.md — new generalized anti-pattern cardwiki/cards/abort-masks-sink.md — sanitizer abort masking downstream bugs