Nick McCarty

Layer 4 Circuit Diagram

nickmccarty0@gmail.com (Nick McCarty) — Wed, 03 Jun 2026 00:00:00 +0000

Interactive visualization of the ResNet Layer 4 co-activation circuit extracted from a fine-tuned Faster R-CNN pot detector.

The Economic Data Tools: FRED and BEA

nickmccarty0@gmail.com (Nick McCarty) — Tue, 02 Jun 2026 00:00:00 +0000

fred_tool: theme-routed FRED API with rate limiting, [FRED:SERIES:DATE] citations. bea_tool: curated query catalogue over NIPA, Regional, GDPbyIndustry, ITA with [BEA:...] citations.

The Wiggum Panel: Three-Persona Parallel Evaluation

nickmccarty0@gmail.com (Nick McCarty) — Tue, 02 Jun 2026 00:00:00 +0000

Three evaluator personas running in parallel threads, each reviewing from a distinct angle. Their issues are deduplicated and merged into Wiggum's revision context before the first rewrite.

Closing the SkillOpt Gaps: What Actually Shipped

nickmccarty0@gmail.com (Nick McCarty) — Tue, 02 Jun 2026 00:00:00 +0000

Three gaps between the harness autoresearch loop and SkillOpt: validation gating shipped as _validate_proposal(), the global exit became Kimi unblocking, and the epoch structure and skill artifact remain open.

The Plugin System: /forge:plugin and /forge:list

nickmccarty0@gmail.com (Nick McCarty) — Mon, 01 Jun 2026 00:00:00 +0000

/forge:plugin: LLM generates plugin JSON spec, files written to plugins//, hot-loaded into skills.REGISTRY. /forge:list: grouped command listing. Plugin skills auto-inject into synthesis; commands invoked as /plugin:command.

The Interview Skills: /grill-me and /onboarding

nickmccarty0@gmail.com (Nick McCarty) — Mon, 01 Jun 2026 00:00:00 +0000

/grill-me: saturation-gated interview loop (3-8 rounds, novelty threshold 3/10, fatigue detector) producing a structured knowledge brief. /onboarding: fixed 3-round scaffold + free-form grill-me + TOML config + ChromaDB memory seed.

The Developer Utility Skills: /scratchpad and /test-harness

nickmccarty0@gmail.com (Nick McCarty) — Mon, 01 Jun 2026 00:00:00 +0000

/scratchpad: force tool loop + exact-values synthesis instruction + prior result injection. /test-harness: run fast/full pytest suite, save output, parse summary.

The Trading Skills: /validate-trades and /execute-trades

nickmccarty0@gmail.com (Nick McCarty) — Mon, 01 Jun 2026 00:00:00 +0000

/validate-trades: deterministic PASS/WARN/FLAG checks on each thesis using local TA signals. /execute-trades: bracket order construction from validated theses, dry-run by default, --live submits to Alpaca.

Circuit Extraction Notebook

nickmccarty0@gmail.com (Nick McCarty) — Sun, 31 May 2026 00:00:00 +0000

Executable notebook for the circuit extraction pipeline: model loading, ablation over 2048 layer4 channels, co-activation graph construction, and Plotly visualization.

The Diagnostic Skills: /debug and /troubleshoot

nickmccarty0@gmail.com (Nick McCarty) — Sun, 31 May 2026 00:00:00 +0000

/debug: last 2 ERROR/FAIL runs, trace events, source anchors, structured Diagnosis/Evidence/Fix. /troubleshoot: debug + suggest in one call, Issue/Root cause/Fix/Next task.

The Site Generation Skills: /design and /build-page

nickmccarty0@gmail.com (Nick McCarty) — Sun, 31 May 2026 00:00:00 +0000

/design: Playwright CSS extraction + vision screenshot analysis + 10-section design system. /build-page: three-pass generation — cluster analysis, HTML shell, per-file card injection with optional visual refinement loop.

The Site Generation Skills: /site and /deck

nickmccarty0@gmail.com (Nick McCarty) — Sun, 31 May 2026 00:00:00 +0000

/site = /design + /build-page in one command. /deck generates a themed .pptx using python-pptx with colors and fonts extracted from the design system.

The Navigation Skills: /suggest and /re-orient

nickmccarty0@gmail.com (Nick McCarty) — Sun, 31 May 2026 00:00:00 +0000

/suggest: next-task recommendation from orientation + runs + git log + autoresearch state. /re-orient: fast project snapshot combining orientation cache with live GitHub data fetched in parallel.

The Explorer View: Per-Run Pipeline DAG Inspector

nickmccarty0@gmail.com (Nick McCarty) — Sat, 30 May 2026 00:00:00 +0000

Per-run pipeline DAG with clickable node inspector — task metadata, memory hits, planned queries, search tool calls, synthesis tokens, Wiggum dimension bars, evaluator reasoning, and RLHF feedback per stage.

The Page Feedback Widget: Closing the Loop Between Browser and Agent

nickmccarty0@gmail.com (Nick McCarty) — Sat, 30 May 2026 00:00:00 +0000

Chrome extension + FastAPI endpoint: leave notes on any local page, agents read them from page_feedback.jsonl. Shadow DOM widget with badge counter, completion checklist, and one-click clear.

The Pipeline View: Data Enrichment DAG

nickmccarty0@gmail.com (Nick McCarty) — Sat, 30 May 2026 00:00:00 +0000

Static SVG DAG showing how web search, Beige Book, FRED, BEA, Market Signals, yfinance, and Alpaca feed a context merge before LLM synthesis and Wiggum evaluation. Conditional dashed edge for trading thesis Alpaca order execution.

The Research History View: A Unified Activity Log

nickmccarty0@gmail.com (Nick McCarty) — Sat, 30 May 2026 00:00:00 +0000

Every search query, browser visit, and research output in one timeline. Filterable by type (search / browser / research) and time window, with score bars and direct links to the Memory store.

inference.py: The Unified LLM Backend Shim

nickmccarty0@gmail.com (Nick McCarty) — Fri, 29 May 2026 00:00:00 +0000

Drop-in `import inference as ollama` replacement that routes to Ollama, vLLM, llama-server, or any OpenAI-compatible endpoint. 14-model name translation map, per-model HARNESS_ENDPOINTS registry, OllamaLike adapter, real wall-clock TTFT timing.

harness/api: The FastAPI Backend and Dashboard Server

nickmccarty0@gmail.com (Nick McCarty) — Fri, 29 May 2026 00:00:00 +0000

Async-native FastAPI server with modular routers, WebSocket live streaming, static dashboard serving, and auto OpenAPI docs. The async replacement for the old Flask server.py.

The Skills Registry: Hook Points, Auto-Activation, and 38 Skills

nickmccarty0@gmail.com (Nick McCarty) — Fri, 29 May 2026 00:00:00 +0000

Six hook points, seven auto-activation predicates, lazy loading, and 38 skills: deep, cite, panel, knowledge-graph, annotate, recall, queue, orientation, introspect, contextualize, sync-wiki, playwright, re-orient, debug, suggest, troubleshoot, transcribe, design, build-page, deck, and more.

Inside agent.py: The Three-Turn Research Pipeline

nickmccarty0@gmail.com (Nick McCarty) — Thu, 28 May 2026 00:00:00 +0000

Three turns: vision preprocessing + search, synthesis written to disk, Wiggum evaluation loop. Plus novelty gating, keep_alive estimation from run history, thinking-model detection, and the SYNTH_INSTRUCTION autoresearch target.

The Position Swap: Beige Book RAG Results and the DPO Cold-Start Problem

nickmccarty0@gmail.com (Nick McCarty) — Thu, 28 May 2026 00:00:00 +0000

Prepending Beige Book context hurt by 0.08; appending helped by 0.13. A falsified hypothesis and what the position swap reveals about training signal for domain-grounded synthesis.

Building the Detectors: What Actually Shipped After 107 Experiments

nickmccarty0@gmail.com (Nick McCarty) — Thu, 28 May 2026 00:00:00 +0000

What the four proposed convergence detectors became in actual implementation — regex over cosine similarity, Kimi unblocking over a hard stop, and a pre-proposal research phase that wasn't in the design.

Live Data Beats Narrative: FRED RAG Experiment Results

nickmccarty0@gmail.com (Nick McCarty) — Thu, 28 May 2026 00:00:00 +0000

FRED fred_end: +0.40 vs control. Beige Book rag_end was +0.13. Live quantitative series is 3× more useful than qualitative Fed prose for grounded economic synthesis.

Memory as Infrastructure: Quality-Weighted Retrieval and the Ontology Graph

nickmccarty0@gmail.com (Nick McCarty) — Thu, 28 May 2026 00:00:00 +0000

How the harness memory layer turns isolated run output into retrieval signal: dual-store architecture, quality-weighted ranking, RLHF feedback loops, and UMAP-based topic visualization.

OSINT Enrichment: Nine Layers of Passive Reconnaissance

nickmccarty0@gmail.com (Nick McCarty) — Thu, 28 May 2026 00:00:00 +0000

Nine-layer OSINT pipeline wired directly into the harness research loop. Zero-config baseline covers DNS, HTTP headers, RDAP, crt.sh, Wayback, and Shodan. Two key-gated layers add tech stack and threat intel.

The Planner: Two-Pass Pre-Research Analysis

nickmccarty0@gmail.com (Nick McCarty) — Thu, 28 May 2026 00:00:00 +0000

Two LLM calls before research starts: a prior knowledge pass (what do I already know? what gaps need search?) and a main plan pass (task_type, complexity, expected_sections, targeted queries). Plan output is injected into synthesis as context.

The Search Cache: SQLite TTL Caching for DDGS Queries and Research Contexts

nickmccarty0@gmail.com (Nick McCarty) — Thu, 28 May 2026 00:00:00 +0000

Two-table SQLite cache: search_cache for per-query DDGS results (always active), research_cache for full research contexts (opt-in via RESEARCH_CACHE=1 for autoresearch). SHA-256 keys, lazy expiry, schema migration.

Deploying the Harness with Docker: CPU, GPU, and Compose Variants

nickmccarty0@gmail.com (Nick McCarty) — Wed, 27 May 2026 00:00:00 +0000

CPU and GPU Dockerfiles plus a three-service Compose stack (vLLM for large models, Ollama for fast models, harness dashboard) with health checks, live code mounts, and GPU passthrough.

The GitHub Skill: LLM-Assisted Git Operations from the Agent Loop

nickmccarty0@gmail.com (Nick McCarty) — Wed, 27 May 2026 00:00:00 +0000

12 GitHub operations auto-detected from task strings, with LLM-generated commit messages, PR titles/bodies, and issue descriptions — invokable directly from the harness agent loop.

The Subagent Demo Suite: Orchestrating Multi-Task Research Portfolios

nickmccarty0@gmail.com (Nick McCarty) — Wed, 27 May 2026 00:00:00 +0000

subagent_demo.py and subagent_demo_v2.py show sequential and parallel multi-task orchestration — Flask queue for one-at-a-time throughput, MCP HTTP server for concurrent execution up to MCP_MAX_CONCURRENCY.

Synthetic Eval Task Generation with TinyTroupe Personas

nickmccarty0@gmail.com (Nick McCarty) — Wed, 27 May 2026 00:00:00 +0000

Eight practitioner personas generate diverse research tasks via Microsoft TinyTroupe (or raw Ollama fallback). Criteria are auto-derived from the task text — count detection, structural requirements, no placeholder checks.

The Email Skill: Personalized Outreach Drafts from Conference Speaker CSVs

nickmccarty0@gmail.com (Nick McCarty) — Tue, 26 May 2026 00:00:00 +0000

Two LLM calls per contact — subject line and body — with slide content fetched via MarkItDown, per-contact JSON output, a manifest file, and every draft logged to runs.jsonl as task_type email_draft.

The Lit-Review Skill: Seven Steps from ArXiv Query to Rendered Survey

nickmccarty0@gmail.com (Nick McCarty) — Tue, 26 May 2026 00:00:00 +0000

Seven steps: ArXiv fetch → Semantic Scholar enrichment → persona curation → per-paper annotation with Wiggum → LLM clustering → cross-cluster synthesis → Jinja2 render. Three output templates: survey, gaps, executive.

Mining a Ground-Truth Knowledge Base for the Eval Suite

nickmccarty0@gmail.com (Nick McCarty) — Tue, 26 May 2026 00:00:00 +0000

Five eval tasks, five deep research runs with the novelty gate disabled, authoritative documents saved to knowledge_base/ and injected as file_context when the eval suite runs those tasks.

The Wiki Sync Skill: Deterministic Source Extraction and Gap-Targeted Code Injection

nickmccarty0@gmail.com (Nick McCarty) — Tue, 26 May 2026 00:00:00 +0000

Regular sync extracts constants, model names, and Wiggum weights from six source files via regex. Gap-targeted sync matches Wiggum FAIL issues to trigger keywords and injects relevant function bodies into the wiki.

Agentic System Design Patterns