![\"Sir](\"assets/svg/banner.svg\")
\n\n \n Run your model. Keep your memory. See every sensitive action before it happens.
\n A local-first AI assistant and knowledge workspace for Windows—without the account, telemetry, or invisible agent.\n
\n Download for Windows\n · \n Quick Start\n · \n How Control Works\n · \n Benchmarks\n · \n Research Method\n · \n Star the Project\n
\n\n\n ![\"CI](\"https://github.com/raydeStar/sir-thaddeus/actions/workflows/ci-pr.yml/badge.svg?branch=master\"){kind=icon}
\n ![\"Latest](\"https://img.shields.io/github/v/release/raydeStar/sir-thaddeus?color=1B3F6E&label=release\")
![\"Apache](\"https://img.shields.io/github/license/raydeStar/sir-thaddeus?color=C9973E\")
![\"Windows\"](\"https://img.shields.io/badge/platform-Windows-0078D6\"){kind=icon}
![\"No](\"https://img.shields.io/badge/telemetry-none-111111\"){kind=icon}
![\"MCP](\"https://img.shields.io/badge/tools-MCP-C79239\"){kind=icon}
Your model. Your memory. Your permission.
\n\n**Sir Thaddeus is an open-source, local-first AI assistant and local LLM agent\nworkspace for Windows.** Connect [LM Studio](https://lmstudio.ai/),\n[Ollama](https://ollama.com/), or another OpenAI-compatible endpoint, then\ngive your model a private workspace: threaded chat, permissioned MCP tools,\ndurable memory, a Markdown wiki, routines, sources, traces, and an emergency\nstop.\n\nUse a model on your machine and leave network tools off to keep the core experience offline. Turn on web access or a hosted endpoint only when *you* decide the job needs it.\n\n| **Your model** | **Your workspace** | **Your boundary** | **Your receipts** |\n| --- | --- | --- | --- |\n| LM Studio, Ollama, or any compatible `/v1` endpoint. | Threads, memory, wiki, routines, and settings on your disk. | Capability groups set to **Off**, **Ask**, or **Always**, plus per-tool overrides. | Sources, outcomes, permission decisions, diagnostics, and per-turn traces. |\n\n## What We Are Proving\n\nSir Thaddeus tests a specific claim: **a fixed small model can complete more useful everyday work when it is given well-designed deterministic capabilities, evidence, state, permissions, and verification.** Replacing it with a larger or newer model may be a sound deployment choice, but it is not evidence that the harness improved the original model.\n\nOptimization work therefore changes one generalized mechanism at a time, compares it with the same raw model and unchanged product, and keeps only repeatable gains. The active priorities, gates, and stop rules live in the [calibrated improvement plan](docs/CALIBRATED_IMPROVEMENT_PLAN.md); the [inference-method gap map](docs/research/INFERENCE_METHOD_GAP_MAP.md) ranks research ideas against measured Thaddeus failures; and the durable record of what worked, failed, or remains uncertain lives in the [research findings](docs/research/README.md).\n\n\n## The Improvement Loop\n\nThis project is no longer developed by asking an AI to \"make it better.\" Each\nbehavior change is treated as a falsifiable experiment against a frozen\nbaseline:\n\n```mermaid\nflowchart LR\n H[\"One hypothesis\"] --> T[\"Cheap reject-only checks\"]\n T --> D[\"Paired development run\"]\n D --> R[\"Exact repeat\"]\n R --> V[\"Disjoint validation\"]\n V --> P[\"Product, safety, and resource gates\"]\n P --> K[\"Merge the win or delete the code\"]\n```\n\nModel capacity, harness capability, and product quality are scored separately.\nA calculator can improve a user's outcome without making the model better at\nclosed-book math; a faster route is not a win if it causes unsafe actions; and\na benchmark fluctuation is not evidence until the exact candidate repeats on\nunseen inputs.\n\nThat distinction is the research contribution. Read\n[Registered Sequential Agent-Harness Experiments](docs/RESEARCH_METHOD.md) for the complete\npublic protocol, representative findings, anti-benchmaxxing boundary, and a\ntemplate for reproducing the work.\n\n\n## Control Is the Feature\n\nMost assistants ask for trust. Sir Thaddeus gives you a boundary you can inspect.\n\n\n\n\n ![\"Sir](\"assets/images/permission-flow-demo.gif\")
\n
Actual Sir Thaddeus UI · deterministic Playwright capture · no simulated product screens
\n\n### Nothing vague. Nothing buried.\n\nWhen a protected capability is set to **Ask**, the turn pauses. You see the tool, its arguments, and the scope of the decision.\n\n| **Deny** | **Allow once** | **For session** | **Always** |\n| --- | --- | --- | --- |\n| Stop this call. | Approve only this action. | Remember it until you leave. | Update the visible policy. |\n\nFile, web, system, screen, and memory capabilities stay independently configurable. Individual tools can override their group.\nWiki mutations are deliberately stricter: each state-changing Wiki call asks\nfor a fresh **Allow once** decision, even when its broader group or tool policy\nis **Always**. Read-only Wiki access continues to follow the configured policy.\n\n## Answers Should Show Their Work\n\nSearch results are not dumped into a footnote graveyard. Sir Thaddeus turns them into readable source cards with titles, domains, snippets, publication context, and one-click access—then keeps the tool outcome in the local trace.\n\n\n\n## More Than a Chat Window\n\n| Research | Private knowledge | Real work |\n| --- | --- | --- |\n| Web search, page fetch, places, weather, time zones, feeds, and holidays. | Durable memory plus a Markdown wiki with roots, folders, pages, revisions, search, import, export, and AI editing. | Scoped file reading for PDF, DOCX, XLSX, CSV, RTF, Markdown, and text; repeatable routines; system tools behind policy. |\n| **Inspectable** | **Model-flexible** | **Interruptible** |\n| Activity, diagnostics, audit events, permission decisions, and per-turn traces. | Local or hosted OpenAI-compatible models, with direct provider and context controls. | Stop active turns and kill sidecar work instead of hoping an agent eventually listens. |\n\nVoice, ASR/TTS, push-to-talk, tray integration, the compact panel, clipboard tools, and screen tools are available as beta features.\n\n\n## Three Minutes to Private AI\n\n### 1. Launch\n\nDownload the [latest Windows release](https://github.com/raydeStar/sir-thaddeus/releases/latest), verify the published checksum, extract it, and run `Launch Sir Thaddeus.cmd` or `Thaddeus.Runtime.exe`.\n\n### 2. Connect your model\n\nOpen **Settings → Models** and test one of these endpoints:\n\n| Provider | Default base URL |\n| --- | --- |\n| LM Studio | `http://127.0.0.1:1234/v1` |\n| Ollama | `http://127.0.0.1:11434/v1` |\n| Custom | Your compatible `/v1` endpoint |\n\n### 3. Draw the line\n\nOpen **Settings → Permissions**. Leave sensitive groups on **Ask**, turn unwanted capabilities **Off**, and allow silent access only where you trust it.\n\nWindows may show SmartScreen while release binaries remain unsigned. The [first-run guide](docs/FIRST_RUN.md) covers setup, local data paths, and backups.\n\n## Built Local-First, Not Painted Local Later\n\n```mermaid\nflowchart LR\n UI[\"Windows shell + React workspace\"] -->|\"rotating token · loopback only\"| RT[\"Local .NET runtime\"]\n RT --> MODEL[\"Your model endpoint\"]\n RT --> POLICY[\"Capability policy\"]\n POLICY -->|\"Off · Ask · Always\"| MCP[\"MCP process boundary\"]\n MCP --> TOOLS[\"Web · files · memory · system\"]\n RT --> DATA[\"Local threads · wiki · routines · traces\"]\n```\n\nThe runtime binds to `127.0.0.1` on an ephemeral port. A bearer token rotates each launch. Protected tool execution crosses a separate MCP process boundary and resolves the policy you configured before it runs. Read the [public architecture overview](docs/ARCHITECTURE_PUBLIC.md) for the full boundary and its explicit non-promises.\n\n\n## Measured, Not Guessed\n\n\"Agentic\" is cheap copy. Sir Thaddeus ships the harness that can prove—or disprove—the claim.\n\nRepresentative results show both the value and the boundary of the approach:\n\n| Question | Result | What it means |\n| --- | ---: | --- |\n| Can existing capabilities improve ordinary verified outcomes for the same 1.2B model? | **4 / 20 raw → 15 / 20 Thaddeus** | The harness captured 13 of 15 gold-supported positive outcomes, but this is not a closed-book intelligence claim. |\n| Can a narrow file-path seam remove mechanical model work safely? | **5 / 16 → 12 / 16**, repeated | Seven paired wins, zero losses, full validity, fewer calls and lower latency. |\n| Can deterministic action policy improve Wiki reliability? | **4 / 18 → 13 / 18** | Nine wins, zero losses, and all deferred requests preserved state. |\n| Can a machine-clock utility improve speed without a model call? | positive p50 **211.5 ms → 5.5 ms** | Two correctness gains, one harmful route repair, and no observed loss. |\n| Did the harness raise stabilized MMLU-Pro? | **10 / 20 raw = 10 / 20 harness** | No reproducible capacity uplift; MMLU optimization is paused. |\n| Did more inference help? | voting **37.9% → 27.9%** | No. The slower mechanism was rejected and removed. |\n\n### Local LLM and AI agent harness benchmark\n\nCan a tool-using AI agent harness improve small local models, a larger\nmixture-of-experts model, and a frontier-class hosted model? The current\nconfirmation used a private, human-reviewed bank of 100 fabricated tasks in 25\nsemantic families, four controlled arms, and two complete repeats. Primary\nscoring was deterministic or based on independently observed final state—not\nan LLM judge.\n\n| Fixed model | Raw | Production prompt, no tools | Direct tools | Full harness | Strict families: no tools → full |\n| --- | ---: | ---: | ---: | ---: | ---: |\n| LFM 2.5 1.2B Q4_K_M | 15/100 | 12/100 | 30/100 | **32/100** | 1/25 → **1/25** |\n| LFM 2.5 8B-A1B Q4_K_M | 24/100 | 22/100 | 42/100 | **48/100; 49/100** | 3/25 → **9/25** both |\n| Gemma 4 26B-A4B Q4_K_M | 23/100 | 5/100; 8/100 | 43/100; 46/100 | **56/100 both** | 0/25 → **10/25** both |\n| Luna, low reasoning | 22/100; 21/100 | 24/100; 23/100 | 55/100; 58/100 | **66/100; 58/100** | 5/25 → **12/25; 9/25** |\n\nThe strict primary result was supported in both repeats for **LFM 8B**\n(`p=0.03125`) and **Gemma 26B-A4B** (`p=0.001953`). Luna improved in both\nrepeats but remained variable; the first was significant and the second was\nnot. LFM 1.2B gained 20 individual outcomes without gaining a complete 4-of-4\nfamily, exposing the floor where tools help but model capacity still limits\nconsistency.\n\nThe direct-tools arm answers the obvious objection that any model with tools\nshould beat one without them. Tools explain much of the lift, while Gemma\nshows the clearest additional gain from production orchestration and\nverification: 43 and 46 cases with direct tools versus 56/100 in both\nfull-harness repeats.\n\nThis is **repeat-tested engineering evidence, not a model leaderboard or an\nintelligence claim**. The formal analyzer still marks the campaign\n`publish_ready: false` because evidence-schema attestations and causal failure\nautopsies remain incomplete. The deterministic totals are useful; “hole-proof\nresearch proof” would overstate them.\n\nRead the\n[complete local LLM agent harness benchmark report](docs/research/SEALED_2026S3_HARNESS_EVIDENCE.md)\nfor the methodology, statistics, all arms, repeat stability, latency, tokens,\nVRAM, host-restart handling, audit gaps, historical context, and reproduction\ncommands in one document.\n\nRun [`dev/model-intake.ps1`](dev/model-intake.ps1) to evaluate a model on your\nown machine against Sir Thaddeus's production baseline.\n\n