From 12c4244b5339028d591eec1150a6cfed53192f6d Mon Sep 17 00:00:00 2001 From: chris harris <283105756+Chris-ObtuseAI@users.noreply.github.com> Date: Sun, 26 Jul 2026 16:45:37 -0500 Subject: [PATCH] Standardize portfolio presentation --- README.md | 66 ++++++++++++++++++++++++++++++++++++------------------- 1 file changed, 44 insertions(+), 22 deletions(-) diff --git a/README.md b/README.md index 123226f..e728cce 100644 --- a/README.md +++ b/README.md @@ -5,10 +5,10 @@ **An evidence-driven laboratory for systems that learn how to improve.** Generate diverse candidates. Measure the visible objective.
-Freeze the contender. Promote only when held-out evidence survives. +Freeze the contender. Promote only when versioned evaluation evidence survives. [![Tests](https://github.com/ObtuseAI/doofus/actions/workflows/tests.yml/badge.svg)](https://github.com/ObtuseAI/doofus/actions/workflows/tests.yml) -[![Held-out benchmarks](https://github.com/ObtuseAI/doofus/actions/workflows/heldout-benchmarks.yml/badge.svg)](https://github.com/ObtuseAI/doofus/actions/workflows/heldout-benchmarks.yml) +[![Reference-suite workflow](https://github.com/ObtuseAI/doofus/actions/workflows/heldout-benchmarks.yml/badge.svg)](https://github.com/ObtuseAI/doofus/actions/workflows/heldout-benchmarks.yml) [![Python 3.11+](https://img.shields.io/badge/Python-3.11%2B-3776AB?logo=python&logoColor=white)](https://www.python.org/) [![Quality diversity](https://img.shields.io/badge/search-MAP--Elites-F97316)](doofus/evolution/map_elites_archive.py) [![MIT License](https://img.shields.io/badge/license-MIT-2563EB)](LICENSE) @@ -19,13 +19,31 @@ Freeze the contender. Promote only when held-out evidence survives. Doofus is a local-first research platform for evolving optimizers, prompts, policies, mutation strategies, and explicitly registered project parameters. -It combines quality-diversity search, nested meta-evolution, calibrated metacognition, sealed held-out benchmarks, proof construction, runtime health, lineage, rollback, and an information-dense evidence dashboard. +It combines quality-diversity search, nested meta-evolution, calibrated metacognition, versioned reference-suite evaluation, proof construction, runtime health, lineage, rollback, and an information-dense evidence dashboard. The name is playful. The research question is not: > How do you let a system change itself aggressively without mistaking activity, a higher training score, or a lucky run for improvement? -Doofus answers with diversity before convergence, predictions before outcomes, held-out evidence before promotion, and recovery as part of the experiment rather than an afterthought. +Doofus answers with diversity before convergence, predictions before outcomes, reference-suite evidence before promotion, and recovery as part of the experiment rather than an afterthought. + +## Product truth + +| Contract | Current public state | +| --- | --- | +| **Portfolio role** | **Evolution and Metacognition Lab** — quality-diversity search, calibrated promotion, replay, and bounded optimization. | +| **Maturity** | **Research preview** | +| **Engineering evidence baseline** | [`c356ff97`](https://github.com/ObtuseAI/doofus/commit/c356ff973daa9ff3985cbb03fd84df064c79f0ed) · configured [tests](https://github.com/ObtuseAI/doofus/actions/workflows/tests.yml) and opt-in reference-suite workflow | +| **Proved now** | The public source implements MAP-Elites archives, challenger tournaments, lineage, candidate freezing, replay, calibrated metacognition, promotion audits, rollback, and bounded external-project experiments. | +| **Authority ceiling** | Research candidates remain shadow or bounded until explicit gates pass. Operators own external-project effects, provider use, retained source mutation, and changes to promotion policy. | +| **Clean demonstration** | Follow the [quick start](#quick-start) to run provider-free tests, the public reference suite, and the local evidence dashboard. | +| **Known limit** | The committed reference tasks are public and therefore are not a secret or independently administered holdout. External evaluation and clean public benchmark provenance remain required for stronger generalization claims. | + +**Designed for:** research teams that want optimization systems to expose diversity, uncertainty, calibration, lineage, and failed promotion—not merely a winning score. + +[Explore the ObtuseAI portfolio](https://github.com/ObtuseAI) · +[Inspect the learning loops](#the-five-nested-learning-loops) · +[Start a technical conversation](https://github.com/ObtuseAI/doofus/issues) ## Doofus in one minute @@ -34,7 +52,7 @@ Doofus answers with diversity before convergence, predictions before outcomes, h | **What is it?** | A metacognitive evolution runtime and research workbench for measurable, reversible improvement. | | **What can it evolve?** | Candidate configurations, optimization strategies, prompts, bounded genes, and declared numeric parameters in external projects. | | **How does it search?** | Mutation, novelty pressure, challenger tournaments, MAP-Elites archives, shadow elites, and higher-order strategy loops. | -| **How does it avoid fooling itself?** | Candidate freezing, held-out tasks, calibration scoring, lineage, proof density, repeatability, uncertainty, and rollback. | +| **How does it avoid fooling itself?** | Candidate freezing, versioned reference tasks, calibration scoring, lineage, proof density, repeatability, uncertainty, and rollback. | | **What does it show?** | Candidates, archives, scores, costs, safety events, metacognitive predictions, promotion audits, runtime health, and recovery state. | | **Does it need an API key?** | No for local tests, benchmarks, dashboards, and provider-free runs. Provider lanes are optional and separately configured. | @@ -67,7 +85,7 @@ flowchart LR VISIBLE --> ARCHIVE["Quality-diversity archive"] ARCHIVE --> CHALLENGE["Challenger tournament"] CHALLENGE --> FREEZE["Freeze candidate identity"] - FREEZE --> HOLDOUT["Sealed held-out suite"] + FREEZE --> HOLDOUT["Versioned reference suite"] HOLDOUT --> AUDIT["Promotion + safety audit"] AUDIT -->|"evidence clears floor"| ELITE["Retained elite"] AUDIT -->|"regression / uncertainty"| ROLLBACK["Rollback + lineage record"] @@ -91,7 +109,7 @@ flowchart TB ARCHIVE["Loop 1
population + MAP-Elites"] CANDIDATE["Loop 0
candidate mutation"] SCORE["Arenas + evidence"] - HOLDOUT["Frozen held-out evaluation"] + HOLDOUT["Frozen reference-suite evaluation"] GATE{"Promotion gate"} SUP --> META2 @@ -168,17 +186,19 @@ Promotion is assembled from multiple evidence surfaces: | **Near-miss analysis** | Why did an otherwise strong candidate fail the final gate? | | **Quarantine** | Is uncertainty material enough to remove the candidate from promotion? | -A high raw score can still be rejected for weak proof, instability, safety findings, or held-out regression. +A high raw score can still be rejected for weak proof, instability, safety findings, or reference-suite regression. -### 4. Sealed held-out benchmarks +### 4. Versioned reference-suite benchmarks -Doofus includes a versioned held-out suite across three task families: +Doofus includes a public, versioned reference suite across three task families. +The committed tasks support deterministic comparison and regression detection; +they do not claim secret evaluator custody or independence from the repository: - **optimization** — candidate behavior on objective landscapes not exposed to the visible loop; - **code repair** — bounded source-repair tasks evaluated through executable behavior; - **calibration** — confidence and prediction quality against known outcomes. -The candidate is serialized and hashed before the held-out tasks are evaluated. The report records: +The candidate is serialized and hashed before the reference tasks are evaluated. The report records: - suite version and engine reference; - frozen candidate identity; @@ -191,7 +211,7 @@ The candidate is serialized and hashed before the held-out tasks are evaluated. sequenceDiagram participant E as Evolution loop participant F as Freeze boundary - participant H as Held-out runner + participant H as Reference-suite runner participant G as Promotion gate E->>E: Optimize on visible objectives @@ -203,9 +223,11 @@ sequenceDiagram G-->>E: Promote, reject, quarantine, or abstain ``` -The slower held-out workflow is opt-in on pull requests through the `run-heldout` label and can also be started manually. This keeps routine CI fast while making benchmark promotion an explicit act. +The slower reference-suite workflow is opt-in on pull requests through the +historical `run-heldout` label and can also be started manually. This keeps +routine CI fast while making benchmark promotion an explicit act. -See the [held-out benchmark protocol](docs/benchmarks/HELDOUT_BENCHMARKS.md). +See the [reference-suite benchmark protocol](docs/benchmarks/HELDOUT_BENCHMARKS.md). ### 5. External project evolution @@ -282,7 +304,7 @@ It brings together: - current candidate and best-known state; - MAP-Elites and shadow-elite archive growth; - lineage and mutation variants; -- visible, held-out, proof-density, and promotion scores; +- visible, reference-suite, proof-density, and promotion scores; - metacognitive predictions and calibration; - model/provider usage and cost; - safety, budget, and quarantine events; @@ -342,7 +364,7 @@ python -m venv .venv .\.venv\Scripts\python.exe -m pytest tests -q ``` -Run the held-out baseline: +Run the versioned reference-suite baseline: ```powershell .\.venv\Scripts\python.exe -m doofus.benchmarks ` @@ -374,10 +396,10 @@ Review provider, cost, secret, and runtime policies before enabling any live mod | **Filesystem** | Registered roots, protected-root rejection, exact target registry, snapshot and rollback. | | **Subprocesses** | Operator-supplied fitness commands run with provider secrets and `DOOFUS_*` controls removed. | | **Providers** | Optional adapters, explicit secret sources, response validation, call ledger, cost and call ceilings. | -| **Promotion** | Score, proof density, replay, regression, safety, bloat, and held-out evidence remain separate checks. | +| **Promotion** | Score, proof density, replay, regression, safety, bloat, and reference-suite evidence remain separate checks. | | **Metacognition** | Prediction and recommendation in shadow mode until calibration and safety gates justify more. | | **Runtime** | Heartbeats, duplicate-process checks, watchdogs, checkpoints, emergency stop, and failover. | -| **Claims** | Internal, visible, held-out, simulated, backtested, and live evidence remain differently labeled. | +| **Claims** | Internal, visible, reference-suite, simulated, backtested, and live evidence remain differently labeled. | ## Release proof @@ -387,7 +409,7 @@ The public workflows separate routine correctness from expensive benchmark evide | --- | --- | --- | | **Tests / pytest** | Push and pull request | Installs the package on Windows and runs the clean-checkout suite without API keys. | | **Tests / Ruff** | Push and pull request | Blocks on the repository’s selected Python error and correctness rules. | -| **Held-out benchmarks** | Manual or `run-heldout` pull-request label | Reproduces the committed baseline, runs the sealed suite, and uploads the exact report for the source SHA. | +| **Reference-suite benchmarks** | Manual or historical `run-heldout` pull-request label | Reproduces the committed baseline, runs the public versioned suite, and uploads the exact report for the source SHA. | | **Public release readiness** | Inside the test job | Checks presentation, links, license, hygiene, and public-source constraints. | Routine CI intentionally does not claim a long-running local evolution history exists on a fresh runner. Machine-local live-runtime tests are marked and isolated from clean-clone proof. @@ -399,7 +421,7 @@ doofus/ |-- evolution/ candidates, archives, novelty, tournaments, lineage |-- metacognition/ predictions, calibration, memory, abstention, control |-- meta/ strategy, prompt, gene, and external-project evolution -|-- benchmarks/ frozen held-out runner and task families +|-- benchmarks/ frozen reference-suite runner and task families |-- scoring/ scorecards, proof density, promotion and quarantine |-- proof/ replay, stability, lineage, and rescue evidence |-- runtime/ health, failover, process identity, recovery @@ -417,7 +439,7 @@ runtime/ machine-local state, metrics, controls, and recovery Recommended reading: 1. [Metacognitive evolution](docs/DOOFUS_METACOGNITIVE_EVOLUTION.md) — runtime modes, evidence loop, and authority. -2. [Held-out benchmark protocol](docs/benchmarks/HELDOUT_BENCHMARKS.md) — task inventory and isolation. +2. [Reference-suite benchmark protocol](docs/benchmarks/HELDOUT_BENCHMARKS.md) — task inventory and isolation. 3. [External project guide](docs/EXTERNAL_PROJECTS.md) — fitness contracts and bounded source experiments. 4. [Safety boundary](docs/DOOFUS_SAFETY_BOUNDARY.md) — mutation, providers, budgets, and recovery. 5. [24/7 autonomy policy](docs/DOOFUS_24_7_AUTONOMY_POLICY.md) — unattended-runtime constraints. @@ -445,7 +467,7 @@ Live provider and external-project modes can incur cost or modify source when se Doofus does not: - guarantee that a fitness function measures what you intended; -- make a held-out suite immune to all forms of leakage; +- make a public reference suite immune to leakage or equivalent to an externally administered holdout; - turn internal score growth into real-world improvement; - remove the need for operator judgment; - authorize arbitrary self-modification;