infra(neon): decommission neon-64gb and redirect all database and LLM endpoints to nixos-laptop

4-Infrastructure/AGENTS.md

@@ -98,10 +98,10 @@ Dynamo-style S3-compatible store written in Rust. Replaced rclone serve s3.
 | Node | Tailscale IP | k3s | Garage | Zone | Disk | SSH |
 |------|-------------|-----|--------|------|------|-----|
 | **qfox-1** (this machine) | 100.88.57.96 | ✅ worker | ✅ 780 GiB | local | 1.8 TB NVMe | local |
-| **cupfox** | 100.72.130.76 | ✅ control-plane | ✅ 69 GiB | fra | 125 GB | key OK (361395) |
+| **cupfox** | 100.115.119.40 | ✅ control-plane | ✅ 69 GiB | fra | 125 GB | key OK (361395) |
 | **nixos-laptop** | 100.102.173.61 | ✅ worker | ✅ 347 GiB | ord | 459 GB NVMe | key OK |
 | **racknerd** | 100.80.39.40 | ✅ worker | ✅ 954 MiB | vps | 9.1 GB VPS | key OK |
-| **neon-64gb** | 100.92.88.64 | ❌ rebuilt (standalone k3s) | ❌ | netcup-arm | 2 TB | key OK (allaun) |
+| **neon-64gb** | 100.92.88.64 | ❌ decommissioned (rebuilt) | ❌ | netcup-arm | 2 TB | key OK (allaun) |
 | **steamdeck** | 100.85.244.73 | ✅ worker | ✅ 373 GiB | gpu | 476 GB NVMe | key OK |
 | rs-vps (netcup) | — | ❌ | ❌ | — | 2 TB | SSH via password |
 | dracocomp | 100.100.140.27 | ❌ | ❌ | — | — | unreachable
@@ -474,15 +474,15 @@ Never claim:
 - `4-Infrastructure/shim/rrc_bosonic_tensor_gpu.py` — GPU-accelerated Bosonic Tensor Network Centrality: computes exp(-i*A*theta) using adaptive RK4 on GPU with `wgpu` (Vulkan) to scale N to 10000+ without CPU eigendecomposition bottleneck. Receipt: `rrc_bosonic_tensor_gpu_receipt.json`.
 - `4-Infrastructure/shim/rrc_bosonic_db_buffer.py` — Asynchronous database ingestion buffer: queues, batches, and flushes PostgreSQL inserts for bosonic tensor network receipts and metrics in thread-safe worker pools.
 - `4-Infrastructure/shim/pist_trace_classify_offline.py` — Offline, token-free trace classifier implementing Lean's `Semantics.PIST.Classify` color-space model and executing the local `rrc-watchdog` binary in the container.
-- `4-Infrastructure/shim/lake_build_ingest.py` — Pure-I/O shim that runs `lake build <target>` and ingests the result into `ene.sessions`, `ene.packages`, `ene.receipts`, and `ene.ingest_events` on the canonical neon-64gb Postgres (`arxiv-pg` container). No admissibility logic, no Float arithmetic.
+- `4-Infrastructure/shim/lake_build_ingest.py` — Pure-I/O shim that runs `lake build <target>` and ingests the result into `ene.sessions`, `ene.packages`, `ene.receipts`, and `ene.ingest_events` on the canonical nixos-laptop Postgres (`100.102.173.61`). No admissibility logic, no Float arithmetic.
 - `4-Infrastructure/k3s-flake/tests/chat-verify.spec.ts` — E2E Playwright verification spec for Chat (Hermes) SSO login.
 - `4-Infrastructure/k3s-flake/tests/audiobookshelf-verify.spec.ts` — E2E Playwright verification spec for Audiobookshelf SSO login.
 
 ## Canonical database locations
 
-- **Postgres (canonical):** `arxiv-pg` podman container on **neon-64gb** (`100.92.88.64`). Databases: `arxiv` (arXiv papers + pgvector HNSW), `ene` (ENE memory substrate with full schema from `ene_substrate_schema.sql`).
+- **Postgres (canonical):** PostgreSQL service on **nixos-laptop** (`100.102.173.61`). Databases: `arxiv` (arXiv papers + pgvector HNSW), `ene` (ENE memory substrate with full schema from `ene_substrate_schema.sql`).
 - **Gremlin (canonical):** Azure Cosmos DB endpoint in `.env.gremlin` (`mathblob.gremlin.cosmos.azure.com`). Current graph: 44,804 vertices, 29,466 edges.
-- **Local qfox Postgres:** none. Do not spin up local Postgres containers on qfox for production data; they are not replicated to neon and create confusion.
+- **Local qfox Postgres:** none. Do not spin up local Postgres containers on qfox for production data; they are not replicated to nixos-laptop and create confusion.
 
 ## Compute Dispatch (WGSL → any substrate)
 
@@ -549,7 +549,7 @@ blessed Compiler surface:
 - Command: `python3 "4-Infrastructure/shim/lake_build_ingest.py" Compiler --actor opencode`
 - Runs in background so `git commit` is not blocked
 - Log: `~/.cache/lake-build-ingest.log`
-- Destination: `ene.sessions`, `ene.packages`, `ene.receipts`, `ene.ingest_events` on neon-64gb (`arxiv-pg` container)
+- Destination: `ene.sessions`, `ene.packages`, `ene.receipts`, `ene.ingest_events` on nixos-laptop (`100.102.173.61`)
 
 To reinstall the hook on a fresh clone, append the background invocation from
 `.git/hooks/post-commit` to the Git LFS post-commit hook.

4-Infrastructure/infra/ene-rds/systemd/ene-api-wrapper.sh

@@ -14,7 +14,7 @@ if [[ ! -x "$BIN" ]]; then
   exit 127
 fi
 
-export RDS_HOST="${RDS_HOST:-100.92.88.64}"
+export RDS_HOST="${RDS_HOST:-100.102.173.61}"
 export RDS_PORT="${RDS_PORT:-5432}"
 export RDS_USER="${RDS_USER:-postgres}"
 export RDS_DB="${RDS_DB:-postgres}"

4-Infrastructure/infra/vikunja_mcp.py

@@ -19,7 +19,7 @@ from urllib.error import HTTPError
 
 SERVER_NAME = "vikunja-mcp"
 SERVER_VERSION = "0.1.0"
-VIKUNJA_URL = os.environ.get("VIKUNJA_URL", "http://100.92.88.64:3456")
+VIKUNJA_URL = os.environ.get("VIKUNJA_URL", "http://100.102.173.61:3456")
 _token_file = os.environ.get("VIKUNJA_TOKEN_FILE", "")
 VIKUNJA_TOKEN = os.environ.get("VIKUNJA_TOKEN", "")
 if not VIKUNJA_TOKEN and _token_file:

4-Infrastructure/rds_probe/src/db.rs

@@ -12,7 +12,7 @@ pub struct DbConfig {
 impl DbConfig {
     pub fn from_env() -> Self {
         let host = env::var("RDS_HOST")
-            .unwrap_or_else(|_| "100.92.88.64".to_string());
+            .unwrap_or_else(|_| "100.102.173.61".to_string());
         let user = env::var("RDS_USER").unwrap_or_else(|_| "postgres".to_string());
         let password = env::var("RDS_PASSWORD").unwrap_or_else(|_| "".to_string());
         

4-Infrastructure/shim/batch_embed_artifacts.py

@@ -19,7 +19,7 @@ from shim.utils import sha256_text, utc_now
 BATCH_SIZE = 50
 TOKEN_REFRESH_SEC = 600
 
-HOST = os.environ.get("RDS_HOST", "100.92.88.64")
+HOST = os.environ.get("RDS_HOST", "100.102.173.61")
 PORT = int(os.environ.get("RDS_PORT", "5432"))
 USER = os.environ.get("RDS_USER", "postgres")
 DB = os.environ.get("RDS_DB", os.environ.get("RDS_DBNAME", "postgres"))

4-Infrastructure/shim/chentsov_fusion.py

@@ -50,7 +50,7 @@ ENDPOINTS = {
         "temperature": 0.1,
     },
     "neon-qwen": {
-        "url": "http://100.92.88.64:11434/v1/chat/completions",
+        "url": "http://100.102.173.61:11434/v1/chat/completions",
         "model": "hf.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF:Q4_K_M",
         "name": "Qwen3.6-35B-A3B (neon, CPU)",
         "timeout": 3600,
@@ -58,7 +58,7 @@ ENDPOINTS = {
         "temperature": 0.1,
     },
     "neon-step": {
-        "url": "http://100.92.88.64:11434/v1/chat/completions",
+        "url": "http://100.102.173.61:11434/v1/chat/completions",
         "model": "step3.7-flash",
         "name": "Step-3.7-Flash (neon, CPU)",
         "timeout": 3600,
@@ -68,7 +68,7 @@ ENDPOINTS = {
     # DeepSeek-Prover-V2-7B: purpose-built Lean 4 / math proof model (~4-5 GB Q4_K_M).
     # Best for hard construction sorries; trained on Lean proof search data.
     "neon-deepseek": {
-        "url": "http://100.92.88.64:11434/v1/chat/completions",
+        "url": "http://100.102.173.61:11434/v1/chat/completions",
         "model": "hf.co/irmma/DeepSeek-Prover-V2-7B-Q4_K_M-GGUF:latest",
         "name": "DeepSeek-Prover-V2-7B (neon, CPU)",
         "timeout": 3600,

4-Infrastructure/shim/dataset_ingest_rds.py

@@ -34,7 +34,7 @@ logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(mess
 log = logging.getLogger("dataset_ingest_rds")
 
 # Config
-RDS_HOST = os.environ.get("RDS_HOST", "100.92.88.64")
+RDS_HOST = os.environ.get("RDS_HOST", "100.102.173.61")
 
 STACK_ROOT = Path(os.environ.get("STACK_ROOT", "/home/researcher/stack"))
 DATA_DIR = STACK_ROOT / "shared-data" / "data" / "ingested_datasets" / "2026-05-18"

4-Infrastructure/shim/gccl_transfer_pipeline.py

@@ -30,7 +30,7 @@ except ImportError:
     sys.path.insert(0, os.getcwd())
     import gccl_waveprobe as gw
 
-DEFAULT_TARGET_HOST = "100.92.88.64"  # neon-64gb IP
+DEFAULT_TARGET_HOST = "100.102.173.61"  # neon-64gb IP
 
 # ── Manifest Generation ─────────────────────────────────────────────────────
 

4-Infrastructure/shim/gremlin_lean_report.py

@@ -215,12 +215,12 @@ def emit_markdown(report: dict, out_path: Path) -> None:
 
 def push_to_appflowy(report: dict) -> None:
     """Push top modules to AppFloyo Cloud workspace."""
-    url = os.environ.get("APPFLOWY_URL", "http://100.92.88.64:8000")
+    url = os.environ.get("APPFLOWY_URL", "http://100.102.173.61:8000")
     token = os.environ.get("APPFLOWY_TOKEN", "")
 
     # If no token, try to get one from GoTrue
     if not token:
-        gotrue_url = os.environ.get("GOTRUE_URL", "http://100.92.88.64:9999")
+        gotrue_url = os.environ.get("GOTRUE_URL", "http://100.102.173.61:9999")
         email = os.environ.get("GOTRUE_ADMIN_EMAIL", "admin@researchstack.info")
         password = os.environ.get("GOTRUE_ADMIN_PASSWORD", "admin123")
         try:

4-Infrastructure/shim/lake_build_ingest.py

@@ -9,7 +9,7 @@ lake_build_ingest.py — Run `lake build` and ingest the result into ENE on neon
 This is a pure-I/O shim: it runs the build, parses the output, and writes rows to
 Postgres. It contains no admissibility logic and no Float arithmetic.
 
-Target database: arxiv-pg container on neon-64gb (Tailscale 100.92.88.64).
+Target database: arxiv-pg container on neon-64gb (Tailscale 100.102.173.61).
 Tables used:
   ene.sessions     — one row per build invocation
   ene.packages     — one row per build target/receipt
@@ -46,7 +46,7 @@ from typing import Any
 
 ROOT = Path("/home/allaun/Research Stack")
 LEAN_DIR = ROOT / "0-Core-Formalism/lean/Semantics"
-NEON_HOST = "100.92.88.64"
+NEON_HOST = "100.102.173.61"
 CONTAINER = "arxiv-pg"
 DB = "ene"
 PIST_CLASSIFY = LEAN_DIR / ".lake" / "build" / "bin" / "pist-classify-trace"

4-Infrastructure/shim/rds_connect.py

@@ -28,7 +28,7 @@ def _resolve_params() -> dict:
                         params["sslmode"] = v
         return params
 
-    host = os.environ.get("RDS_HOST", os.environ.get("PGHOST", "100.92.88.64"))
+    host = os.environ.get("RDS_HOST", os.environ.get("PGHOST", "100.102.173.61"))
     port = int(os.environ.get("RDS_PORT", os.environ.get("PGPORT", "5432")))
     user = os.environ.get("RDS_USER", os.environ.get("PGUSER", "postgres"))
     dbname = os.environ.get("RDS_DB", os.environ.get("RDS_DBNAME", os.environ.get("PGDATABASE", "postgres")))

4-Infrastructure/shim/sync_wiki_to_rds.py

@@ -17,7 +17,7 @@ from shim.utils import sha256_text, utc_now
 STACK_ROOT = Path(os.environ.get("STACK_ROOT", "/home/allaun/Research Stack"))
 WIKI_ROOT = Path(os.environ.get("WIKI_ROOT", str(STACK_ROOT / "6-Documentation" / "wiki")))
 
-HOST = os.environ.get("RDS_HOST", "100.92.88.64")
+HOST = os.environ.get("RDS_HOST", "100.102.173.61")
 DB = os.environ.get("RDS_DB", os.environ.get("RDS_DBNAME", "postgres"))
 
 

4-Infrastructure/storage/cache-offload.sh

@@ -32,7 +32,7 @@ TIER0_PGDUMP="/mnt/stackcache/pgdump"
 TIER2_REMOTE="gdrive:research-stack-offload"
 TIER1_THRESHOLD_GB=2
 
-RDS_HOST="${RDS_HOST:-100.92.88.64}"
+RDS_HOST="${RDS_HOST:-100.102.173.61}"
 RDS_PORT="${RDS_PORT:-5432}"
 RDS_USER="${RDS_USER:-postgres}"
 RDS_DB="${RDS_DB:-postgres}"

4-Infrastructure/storage/garage/db-consolidate.sh

@@ -39,7 +39,7 @@ export AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY
 export AWS_DEFAULT_REGION="${AWS_DEFAULT_REGION:-garage}"
 export AWS_ENDPOINT_URL="${AWS_ENDPOINT_URL:-http://localhost:3900}"
 
-RDS_HOST="${RDS_HOST:-100.92.88.64}"
+RDS_HOST="${RDS_HOST:-100.102.173.61}"
 RDS_PORT="${RDS_PORT:-5432}"
 RDS_USER="${RDS_USER:-postgres}"
 RDS_DB="${RDS_DB:-postgres}"

4-Infrastructure/storage/restic/backup.sh

@@ -38,7 +38,7 @@ fi
 export AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY AWS_DEFAULT_REGION
 export RESTIC_REPOSITORY RESTIC_PASSWORD_FILE
 
-RDS_HOST="${RDS_HOST:-100.92.88.64}"
+RDS_HOST="${RDS_HOST:-100.102.173.61}"
 RDS_PORT="${RDS_PORT:-5432}"
 RDS_USER="${RDS_USER:-postgres}"
 RDS_DB="${RDS_DB:-postgres}"

5-Applications/tools-scripts/llm/deepseek_v4_flash_lean_harness.py

@@ -84,13 +84,13 @@ PROVIDERS: dict[str, dict] = {
         "notes": "OpenRouter, deepseek-v4-flash",
     },
     "neon-deepseek-prover": {
-        "api_base": "http://100.92.88.64:11434/v1",
+        "api_base": "http://100.102.173.61:11434/v1",
         "api_key": "",
         "model": "hf.co/irmma/DeepSeek-Prover-V2-7B-Q4_K_M-GGUF",
         "notes": "neon-64gb, DeepSeek-Prover-V2 7B Q4_K_M, 62GB RAM CPU-only",
     },
     "neon-goedel-prover": {
-        "api_base": "http://100.92.88.64:11434/v1",
+        "api_base": "http://100.102.173.61:11434/v1",
         "api_key": "",
         "model": "hf.co/mradermacher/Goedel-Prover-V2-8B-GGUF",
         "notes": "neon-64gb, Goedel-Prover-V2 8B GGUF, 62GB RAM CPU-only",

5-Applications/tools-scripts/mcp/cloud_bootstrap.sh

@@ -5,7 +5,7 @@
 #
 # Environment variables (set these in cloud env config):
 #   BUILD_SERVER_URL=http://100.88.57.96:8765
-#   NEON_OLLAMA_URL=http://100.92.88.64:11434
+#   NEON_OLLAMA_URL=http://100.102.173.61:11434
 #   LAKE_WORKDIR=/workspace/0-Core-Formalism/lean/Semantics
 
 set -e
@@ -54,7 +54,7 @@ try:
     print('  build_server: UP')
 except: print('  build_server: DOWN (tailnet required)')
 try:
-    r = urllib.request.urlopen('${NEON_OLLAMA_URL:-http://100.92.88.64:11434}/api/tags', timeout=5)
+    r = urllib.request.urlopen('${NEON_OLLAMA_URL:-http://100.102.173.61:11434}/api/tags', timeout=5)
     print('  neon_ollama:   UP')
 except: print('  neon_ollama:   DOWN (tailnet required)')
 "

5-Applications/tools-scripts/mcp/token_saver_mcp.py

@@ -50,7 +50,7 @@ def _ensure_singleton() -> None:
 # ── Resource URLs (configurable via env) ────────────────────────────────
 
 BUILD_SERVER = os.environ.get("BUILD_SERVER_URL", "http://100.88.57.96:8765")
-NEON_OLLAMA   = os.environ.get("NEON_OLLAMA_URL", "http://100.92.88.64:11434")
+NEON_OLLAMA   = os.environ.get("NEON_OLLAMA_URL", "http://100.102.173.61:11434")
 NEON_MODEL    = os.environ.get("NEON_MODEL",
     "hf.co/irmma/DeepSeek-Prover-V2-7B-Q4_K_M-GGUF")
 LAKE_WORKDIR  = os.environ.get("LAKE_WORKDIR",

AGENTS.md

@@ -58,6 +58,38 @@ fully-local Hermes-3 model on qfox-1's RTX 4070. OpenClaw is decommissioned.
     api_key: "sk-local"
   ```
 
+## Provider-nixos LLM Gateway (2026-07-03)
+
+`neon-rs1000` (Tailscale `100.79.14.103`, NixOS, 4 cores, 7.8Gi RAM) runs the
+FreeLLMAPI proxy + OpenCode server as the model inference gateway for the
+onklaud-5 pipeline.
+
+### FreeLLMAPI (port 3001)
+
+- **Repo**: `tashfeenahmed/freellmapi`, cloned to `~/freellmapi`, production build
+- **Dashboard**: `http://100.79.14.103:3001` — admin: `admin@researchstack.info` / `admin123`
+- **Unified API key**: `freellmapi-ef5610fa456735f2bcb6205faf03f8725fe41265f129cd56`
+- **Base URL**: `http://100.79.14.103:3001/v1`
+- **Systemd user unit**: `freellmapi.service` — enabled, auto-restart
+- **20 provider keys installed** (Google, Groq, Cerebras, NVIDIA, Mistral, OpenRouter,
+  GitHub, Cloudflare, Zhipu, Ollama Cloud, HuggingFace, OpenCode Zen, Kilo,
+  Pollinations, OVH, LLM7, AI Horde, Agnes, Reka, SiliconFlow, Routeway, BazaarLink)
+- **~1.7B tokens/month aggregate**, 6.1K used (~99.99% remaining)
+
+### OpenCode Server (port 4096)
+
+- **Version**: 1.17.9, installed via `nix profile install nixpkgs#opencode`
+- **Systemd user unit**: `opencode-serve.service` — enabled, auto-restart
+- **Provider**: FreeLLMAPI via `@ai-sdk/openai-compatible` (model: `freeapi/auto`)
+- **Config**: `~/.config/opencode/config.json`
+
+### Onklaud-5 Pipeline
+
+- **Path**: `~/onklaud-5`
+- **council.py**: Patched to route through FreeLLMAPI (`localhost:3001/v1`) instead of
+  OpenRouter directly. Both `KIMI_MODEL` and `GLM_MODEL` set to `auto`.
+- **Venv**: `~/onklaud-5/venv`, 31/32 tests pass
+
 ### Common failure modes
 
 - `500 Internal Error` from `chat.researchstack.info` after a model swap:
@@ -91,18 +123,11 @@ workstation.
   failures and writes corrections to AGENTS.md. Requires `claude` CLI available
   for the analysis LLM.
 
-### Neon proxy (2026-06-19)
+### DECOMMISSIONED — Neon proxy (2026-06-19)
 
-Headroom proxy also runs on neon-64gb (netcup ARM64, NixOS) for remote agent
+[INACTIVE / DECOMMISSIONED] Headroom proxy previously ran on neon-64gb:
-sessions:
+- **Tailscale endpoint**: `http://100.92.88.64:8787` (inactive)
-
+- **Status**: Decommissioned. All remote/local LLM sessions route directly through qfox-1 or external providers. Do not route ANTHROPIC_BASE_URL to neon.
-- **Tailscale endpoint**: `http://100.92.88.64:8787` (also reachable as
-  `http://neon-64gb:8787` via MagicDNS)
-- **Startup**: `/home/allaun/.headroom/headroom-proxy-start.sh` (wraps
-  `LD_LIBRARY_PATH` for NixOS libstdc++ compatibility)
-- **Installed via**: Python venv at `~/headroom-venv/`
-- **Config**: `--memory --learn --port 8787 --host 0.0.0.0`
-- **Route through it**: `ANTHROPIC_BASE_URL=http://100.92.88.64:8787 claude`
 
 ### Common failure modes
 
@@ -432,6 +457,114 @@ For files using PhysicsScalar.Q16_16:
 ### Migration Guide
 See `LEAD_Q16_16_MIGRATION_GUIDE.md` for detailed microsteps.
 
+## Gauge Theory Research Program
+
+Gauge theory is the most rigorous framework for mapping out math because it forces explicit
+distinctions between theorems, conjectures, and empirical observations. Use it as a probe,
+not as a claim.
+
+### What Gauge Theory Demands
+
+1. **Proven vs Conjectural**
+   - If you claim a correspondence, you must prove it or label it an ansatz
+   - Empirical correlations (r=0.41, r=-0.33, n=13, p>0.05) are data, not theorems
+   - Gauge theory requires: "Is this a theorem or an ansatz?"
+
+2. **Dimensional Honesty**
+   - AT model is 2D; gauge confinement requires 3+1D (or 2+1D for compact U(1))
+   - Confinement in 2D gauge theory is trivial — this is a theorem
+   - Dimensional mismatches must be explicitly addressed, not papered over
+
+3. **Category/Type Correctness**
+   - Frustration is boolean/Z₂
+   - Wilson loop is real-valued (trace of holonomy)
+   - Curvature is Lie-algebra-valued
+   - Confusing these categories is a type error
+
+### What Gauge Theory Does NOT Give You
+
+1. **Derivation for Free** — The Baker-Hopf coupling is an ansatz, not derived from gauge principles
+2. **Uniqueness** — Other connections (arctan, Li₂, etc.) could satisfy similar constraints
+3. **Predictive Power** — Correlations computed because the hypothesis suggested them
+
+### Success Criteria
+
+Success is NOT "SilverSight IS gauge theory."
+
+Success IS: "We attempted to derive SilverSight structures from gauge theory.
+We succeeded for X, failed for Y, and learned Z."
+
+Specifically:
+- **Attempt derivation for Baker-Hopf coupling** — proves it's gauge-theoretic or reveals it's something else
+- **Attempt derivation for frustration = Wilson loop** — proves the correspondence or reveals it's just a correlation
+- **Attempt derivation for Baker Λ = field strength** — proves the formal resemblance or reveals it's just a resemblance
+
+### Failure Mode
+
+Failure is NOT "correspondences don't hold."
+
+Failure IS: "We didn't attempt the derivation, so we don't know what the structures actually are."
+
+### Recommended Framing
+
+Use gauge theory as a research program:
+
+```
+"We investigate whether SilverSight can be derived from lattice gauge theory.
+We have proven X, observed Y empirically, and conjecture Z.
+Here are the falsification criteria."
+```
+
+Never claim:
+
+```
+"SilverSight IS gauge theory"
+```
+
+## Fraction Rules
+
+The Research Stack operates on fixed-point arithmetic for mathematical rigor.
+Float (`ofFloat`) is forbidden in compute paths.
+
+### Q16_16 Fraction Types
+
+| Type | Description | Use Case |
+|------|-------------|----------|
+| `Q16_16` | 16.16 fixed-point | General computation |
+| `Q0_16` | 0.16 fixed-point | Values in [0, 1) |
+| `Q12_20` | 12.20 fixed-point | High precision fractions |
+| `Q20_12` | 20.12 fixed-point | Large integer + fraction |
+
+### Fraction Construction Rules
+
+1. **Use canonical constructors only**
+   ```lean
+   Q16_16.ofNat n      -- for integers
+   Q16_16.ofRatio a b  -- for rational numbers
+   Q16_16.ofRawInt x   -- for already-scaled integers
+   ```
+
+2. **Never use `ofFloat` in compute paths**
+   - Allowed only at external boundary (JSON parsing, sensor input)
+   - Must be immediately bracketed: `ofFloat x |> toFixedPoint |> compute`
+
+3. **Bridge pattern for legacy code**
+   ```lean
+   import Semantics.PhysicsScalarBridge
+   -- Use PhysicsScalarBridge.add, .gt, etc.
+   -- Constants: PhysicsScalarBridge.one, .two, .three, .half, .quarter
+   ```
+
+### Fraction Comparison Pattern
+
+```lean
+-- WRONG: direct comparison with Float
+if x > 0.5 then ...
+
+-- RIGHT: fixed-point comparison
+if x > Q16_16.half then ...
+```
+
 ## Glossary (before reading further)
 
 These terms appear throughout all AGENTS.md files and the codebase: