docs: add evaluation methodology for benchmark claims

Document the +122% improvement benchmark methodology referenced
on the landing page. Covers setup, per-task scores, scoring
criteria, what TaskWing provides, and limitations.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: josephgoksu

docs/development/EVALUATION.md

@@ -0,0 +1,106 @@
+# TaskWing Evaluation Methodology
+
+## Overview
+
+We evaluated whether injecting project-specific context via TaskWing
+improves the quality of LLM-generated architectural responses compared
+to a baseline (no context) scenario.
+
+**Result: +122% improvement** (3.6 → 8.0 average score).
+
+## Setup
+
+| Parameter        | Value                                      |
+|------------------|--------------------------------------------|
+| **Codebase**     | Production Go/React monorepo               |
+| **LLM judge**    | gpt-5-mini                                 |
+| **Tasks**        | 5 architectural questions                  |
+| **Scoring**      | 1–10 per task, averaged                    |
+| **Conditions**   | Baseline (no context) vs TaskWing-injected |
+
+## Tasks
+
+Each task required the LLM to answer an architectural question about
+the codebase. Correct answers required knowing:
+
+1. The primary language (Go, not TypeScript)
+2. Correct file paths and project structure
+3. Correct build/generate commands
+4. Architectural patterns and constraints
+5. Technology decisions and their rationale
+
+## Results
+
+### Per-Task Scores
+
+| Task | Without Context | With TaskWing | Delta |
+|------|---------------:|-------------:|------:|
+| T1   | 6              | 8            | +2    |
+| T2   | 3              | 8            | +5    |
+| T3   | 3              | 8            | +5    |
+| T4   | 3              | 8            | +5    |
+| T5   | 3              | 8            | +5    |
+| **Avg** | **3.6**     | **8.0**      | **+4.4** |
+
+**Improvement: +122%** (8.0 / 3.6 - 1)
+
+### Without Context (Baseline)
+
+The LLM without context consistently:
+- Assumed TypeScript instead of Go
+- Referenced nonexistent files like `src/types/openapi.ts`
+- Suggested `npm run generate` instead of `make generate-api`
+- Missed architectural constraints entirely
+
+Only T1 scored above 3, likely due to generic reasoning.
+
+### With TaskWing (Context Injected)
+
+TaskWing's MCP integration provided the LLM with:
+- **Decisions**: Technology choices and their rationale
+- **Patterns**: File structure conventions and API patterns
+- **Constraints**: Build requirements and deployment rules
+
+The LLM consistently identified Go, referenced correct file paths
+(`internal/api/types.gen.go`), and used correct commands.
+
+## Scoring Criteria
+
+- **8–10**: Correct language, correct paths, correct commands,
+  respects constraints
+- **5–7**: Partially correct; right language but wrong paths,
+  or right paths but wrong commands
+- **1–4**: Wrong language or fundamentally incorrect assumptions
+- **Rule**: Wrong tech stack identification = automatic score ≤ 3
+
+## What TaskWing Provides
+
+During the evaluation, TaskWing injected the following context
+via the MCP protocol:
+
+```
+Decisions:  22 (e.g., "PostgreSQL over MongoDB", "OpenAPI codegen")
+Patterns:   12 (e.g., "internal/api/handlers/ convention")
+Constraints: 9 (e.g., "No .env in production — use SSM")
+```
+
+This context was extracted automatically by `taskwing bootstrap`
+in under 3 seconds.
+
+## Reproducing
+
+1. Clone any Go or multi-language repository
+2. Run `taskwing bootstrap` to extract context
+3. Ask the same architectural questions with and without
+   TaskWing's MCP server connected
+4. Score responses on a 1–10 scale using the criteria above
+
+## Limitations
+
+- Single codebase evaluated (Go/React monorepo)
+- Single LLM judge model (gpt-5-mini)
+- 5 tasks may not capture all architectural reasoning scenarios
+- Scores are relative — absolute quality depends on the model used
+
+We plan to expand this evaluation to more codebases and models
+in future iterations.