Baltor Showcase
baltor.ai

Field notes from building Baltor.

Technical articles and benchmarks on agent runtimes, contract cards and distributed systems. Every post comes as a web page and as Markdown for your agent.

4 posts, last published

FeaturedContext Architecture7 min read

Why Every Baltor Capability Starts as a < 2 KB Contract Card Before Loading Source

Full source files spend a model's attention during discovery. Baltor keeps discovery and loading apart: every capability starts as a typed contract card under 2,048 bytes, and source loads once the model has chosen.

Published by Baltor.AI

A contract card
component_id openapi-to-mcp-server
schema_version 1
engine_slot codegen
input_ports openapi_spec
output_ports server_dir, tests
refusal_codes step_effects_required
required_authority reads_fs, writes_fs
verifier_sha256 3f9c1a07…
1.4 KB of a 2,048-byte card

All posts

  1. Runtime & Benchmarking8 min readMarkdown version

    How Compare Runs Raw-Prompt vs. With-Baltor Side-by-Side in Real Time

    Two isolated lanes run identical prompts on the same model weights, so differences in frame time, physics and constraints trace back to the harness around the model.

  2. Distributed Systems & Simulation8 min readMarkdown version

    Simulating Backpressure, Consistent Hashing, and Shard Failover Before Writing Production Code

    Discrete-event simulations test token-bucket admission, bounded-load consistent hashing and jittered circuit breakers before the production code is written.

  3. Loop Runtime & Engine Slots9 min readMarkdown version

    One Loop Runtime: Composing Deterministic and Model Nodes with Swappable Engine Slots

    Every graph vertex is a Loop. Roles combine with engine slots declared in engine_slots.yaml, so an engine can be swapped while every caller stays the same.