Chronohorn

Family-agnostic experiment tracker and architecture-search runtime for predictive descendants. Built on the shared decepticons kernel.

What It Does

• Tracks experiments from any model family, stores them in SQLite, and keeps legality/trust state attached to results
• Analyzes curves and frontiers with saturation, marginal ranking, velocity, and ablation-board views
• Runs the search loop through manifest-driven fleet dispatch, drain, result pull-back, and auto-deepen/control surfaces
• Exposes runtime state to agents through 64 MCP tools, a terminal observer, and an HTTP runtime dashboard

Chronohorn is family-agnostic at the runtime layer. Family-specific mutation policy lives under python/chronohorn/families/<name>/; the shared mechanism layer stays in the decepticons kernel.

Quick Start

# Install from PyPI (decepticons kernel pulled automatically)
python3 -m pip install chronohorn

# Or monorepo dev install: shared kernel first, then runtime
python3 -m pip install -e ../decepticons
python3 -m pip install -e .

# Ingest results and view the observer/dashboard
chronohorn observe serve --result-dir out/results

# Emit a family-owned scan manifest
chronohorn fleet emit-family-matrix --family causal-bank --regime gated-retention

# Full daemon: drain + fleet probe + observer + MCP
chronohorn runtime --manifest manifests/frontier_gated_retention.jsonl

# CLI help
chronohorn --help

MCP Integration

Chronohorn exposes a stateful MCP surface for experiment querying, frontier analysis, ablation tracking, fleet control, saturation detection, learning-curve comparison, and manifest/runtime management. The exact tool set changes with the runtime; the live registry is in python/chronohorn/mcp.py. Run chronohorn mcp for stdio transport. See .mcp.json for a client configuration example.

Repo Boundary

The intended split is:

decepticons -> chronohorn -> heinrich
kernel         runtime       evidence / audit

• decepticons owns reusable mechanisms, config validation, and export-friendly kernel surfaces
• chronohorn owns training, replay, scoring, scan emission, fleet execution, and runtime control
• heinrich is outside the runtime path and owns external evidence packaging

See docs/REPO_BOUNDARY.md and docs/STACK.md for the promoted boundary.

Repo Guide

The repo has a few different kinds of material that matter for different reasons:

• docs/README.md points to canonical live docs vs historical docs
• manifests/README.md explains named regimes, generated queue files, and archive intent
• state/README.md explains the tracked runtime snapshot and handoff files
• scripts/README.md explains which scripts are wrappers vs maintenance utilities

Adding a Model Family

Create a package at python/chronohorn/families/<name>/ implementing the FamilyTrainingAdapter protocol. The registry auto-discovers it — no manual registration. See CLAUDE.md for the full protocol reference and conventions.

Current Focus

The active causal-bank search is organized around cheap O(n) architecture screening before promotion:

• 10k rapid ablation lanes for mechanism screening
• scale/context survival rows aimed at pushing the frontier toward 1.0
• primary learned-substrate experiments around gated_delta
• VRAM-tier-aware fleet placement so small CUDA rows can prefer the smallest sufficient GPU lane

Current manifests live under manifests/, and current results/launch state live under out/results/ and out/fleet/.

License

MIT — see LICENSE.