Minimal, clean rebuilds of ideas from unify-agent, the agent harness unify built from April 2025, each tested against the strongest simple alternatives on data that reflects the conditions the idea is for, with today's models. Building a brain around a stateless model tells the story behind them, and what these results say about it.

Why this repo exists (Dan, 9 October 2026): the first experiments on these ideas ran the original unify-agent code as it stood when each idea went in. Most of what they measured was that code's bugs and toy settings in which a simpler design was bound to do as well. Here each idea is rebuilt from scratch, small and bug-free, so the only question is whether the idea itself holds.

1. One idea per folder, concepts/<name>/: the idea's minimal implementation, the baselines, the data generator, the runner, the analysis, tests and a README. A concept's code imports nothing from another concept, only from common/.
2. Minimal and correct. The smallest code that implements the idea fully. Unit tests cover the mechanism and every baseline. A sanity condition, in which every arm should succeed, runs first. Read the transcripts of a sample of runs, in every arm, before trusting a number. A bug found in any arm is fixed and every affected run re-run; discarded runs are kept in a labelled file with the reason.
3. Strong baselines. Each idea is compared with what a competent engineer would build instead, today, with today's models. That includes "put everything in one long context", which million-token contexts have made strong. Arms use the same model, prompts and data except for the mechanism.
4. Conditions written down first. Before any experiment run, the README gets a timestamped "Plan": the conditions the idea is built for (from its own reasoning), the arms, the measures, the items and seeds, and what result counts as the idea winning. It changes only for a documented fault. Every plan includes a small setting and the realistic one, ideally a sweep of scale, so the results show where the idea pulls ahead, or that it does not.
5. Report everything as it comes out. Effects with intervals (common/src/demos_common/stats.py), and accuracy, cost and latency side by side. If the idea loses, the README says so plainly. Never tune a setting until the idea wins.
6. Models: today's, with strength as an axis. Only models someone would reasonably choose today, on the cost and quality frontier; no older models, and no backdating (Dan, 9 October 2026). Stronger models need thinner harnesses, so each idea is tested at three tiers (below): its full design on the cheap tier, on two families, and its headline comparison on the mid and top tiers, showing whether its advantage holds or fades as models get stronger. Every arm in a condition uses the same model. A judge comes from a different family from the model under test where possible, and is checked by hand on a sample.
7. Every model call goes through demos_common.LLM, which writes each call to the concept's ledger with its exact cost. Never print, log or store the key.
8. Budget per concept: aim under $50 of model spend; stop and report before $80. Runs on the older models of the short-lived amendment of 9 October, kept apart and out of every analysis, do not count against the $80: that detour was the lead's, not the concept's (decided when the round was picked up after a usage limit cut it off).

common/                       shared, tested: demos_common.LLM (OpenRouter
                              client + cost ledger + retries + time to first
                              token), demos_common.stats, demos_common.ledger
concepts/<name>/              one idea: pyproject.toml (own dependencies, with
                              demos-common as a path dependency), src/, tests/,
                              data/, runs/ (ledger and raw records), results.json,
                              README.md (idea, plan, results, caveats, cost),
                              runs.tar.zst (runs/ archived for git)
bin/with-openrouter-key       runs a command with the key in its environment

Run records are too large for git loose (one concept wrote 180 MB in 20,000 files), so runs/ is ignored and each concept commits runs.tar.zst instead, about 70 times smaller: make it with tar -cf - runs | zstd -19 -T0 -o runs.tar.zst and restore it with zstd -dc runs.tar.zst | tar -xf -. An archive over 45 MB is committed in parts instead, since GitHub refuses files over 100 MB and warns over 50 MB: split -b 45m runs.tar.zst runs.tar.zst.part-, with runs.tar.zst itself in the concept's .gitignore; restore with cat runs.tar.zst.part-* | zstd -dc | tar -xf - (memory-consolidation does this).

A new concept's pyproject.toml depends on the shared package like this:

[project]
dependencies = ["demos-common"]
[tool.uv.sources]
demos-common = { path = "../../common", editable = true }

On the laptop: cd concepts/<name> && ../../bin/with-openrouter-key uv run python run.py ... (the key comes from Secret Manager as the team service account).

On a cloud machine: the round of 9 October 2026 ran on a Google Cloud VM, blog-exp-1 in saas-368716 (32 vCPU, 125 GB), deleted that night once every file was back here. For another round, rent one as brain's docs/operations/cloud-compute.md says, with a lease, give it the key in a file only its user can read (~/.experiments.env), copy a concept over with rsync -az --exclude .venv --exclude __pycache__ concepts/<name>/ <host>:demos/concepts/<name>/, run the same command there under nohup or tmux, and copy runs/ and results back after every batch. The machine is temporary: nothing may live only there.

The round of 9 October 2026 is over. Seven concepts finished, each with its verdict at the top of its README's "Results"; separate-loops was stopped before its analysis when Dan stopped the experiments that night, and its README says what ran.

OpenRouter list prices per million tokens on 9 October 2026 (all take images and about a million tokens of context):

Older generations (GPT-5.x, Claude 4.x, Gemini 2.5 and earlier) are left out: nobody would choose them today at their price. The fast models in a spoken agent come from the cheap tier.