SEED is a measurement and epistemic-control protocol for distinguishing implementation activity from verified progress in agent-driven software work. Formerly Bonkers; the repository was renamed from thefleet to seed-protocol on 2026-09-02 and the old name redirects. Not an agent runtime, a fleet orchestrator, or a claim of autonomy. Repository: thisisntjon/seed-protocol.
Status: HYPOTHESIS (C-004) · INSTRUMENT VALIDATED (sabotage 23/23, receipt) · case study AUTHOR-RUN
Research question (C-004): Does an incident-earned, machine-checked control plane reduce invalid completion claims, recovery cost, or unnecessary human intervention relative to a bare repository, without a material drop in correct task completion? The matched ablation is designed and paused (workflow/experiments/2026-08-08-p3-bare-vs-seed-ablation.md). A null result is a valid outcome.
Source: workflow/research/2026-08-24-pr-case-study/artifacts/summary.json field weekly. Object: poketcg (private repository) @ pin 9522a8a37078d00f46b99a586b825b789b01387d. Labels: docs_only = every changed path is .md/.txt; has_code = at least one .py/.go/.js/.ts/.rs/.java/.c/.cpp/.h; n = PR-linked first-parent commits that week (CODEBOOK.md). Campaign docs_only_rate = 0.432 (855/1979). Author-run, not independently reproduced. The clone that produced the census is not this repository.
Four numbers from that pin (PAPER.md, artifacts/summary.json): 1,979 PR-linked first-parent commits analyzed (author-run, n = 1 repository). 43.2% of those were docs-only (855/1,979; composition, not merge rate). 0/40 sampled PRs had an independent human GitHub review (40-PR sample; the only reviewer login was the author; artifacts/review_sample.json). 2/80 sampled PRs changed the playing agent (80-PR stratified inspection, single inspector; artifacts/sample80_summary.json).
python scripts/onboard_check.py
python scripts/schema_check.py
python -m unittest discover -s schema/tests
python scripts/sabotage_test.py
python scripts/poc_check.py
python workflow/blueprint/bin/selftest.py
python workflow/harvest/pack/assemble.py --check
python workflow/blueprint/bin/orient.py
CI runs the same commands (.github/workflows/verify.yml). A green onboard check means the docs match this tree. It does not mean the research question is answered. To rerun the case-study instrument against the pinned poketcg SHA, pass an explicit clone path of the private source repository to workflow/research/2026-08-24-pr-case-study/measure_pr_census.py; a judge who only clones this repository can read artifacts/ but cannot regenerate the census. Independent reproductions: file one with the reproduction issue template.
- C-004 is untested. Supported rows are about the instrument, not causal benefit.
- The PR case study is n = 1 repository, n = 1 principal, author-run. It does not support a treatment-effect claim. The public dataset is
workflow/research/2026-08-24-pr-case-study/artifacts/. The source mill (poketcg) is a private repository and is not this repository; the public dataset isartifacts/here. - The playbook line “encoded rules never failed again” is unverified (
workflow/canon/DECISIONS.md). Recurrence was not tracked in the source ledger. - Factory bootstrap (
workflow/blueprint/bin/bootstrap.py) is not yet a CI fixture. Continuity (switch.py) is not built. - This repository became public on 2026-09-02 (gate G-002 in
GATES.md). Public visibility is not evidence; the claim table above is.
Agents cold-start at START-HERE.md and trust nothing until python scripts/onboard_check.py is green. Humans scoring this tree start at JUDGES.md. Claims and falsifiers: CLAIMS.md. Incident-earned rules: LAWS.md. Machine-form artifacts: schema/README.md. Case study: PAPER.md.
Implementation throughput is a poor proxy for verified progress in agent-driven software work. SEED records that finding as a small, incident-bounded protocol: claims have falsifiers, guards are sabotage-tested before they are trusted, state lives in git artifacts, humans gate only irreversibles, and spend is scored against closed loops.
The protocol is an epistemic control plane. It does not schedule agents. A later factory extract packages portable ideas so a second project does not re-derive them. Whether using SEED improves recovery time, false-progress rate, or human intervention versus a bare project is hypothesis C-004 and is not claimed.
| ID | Claim | Status |
|---|---|---|
| C-001 | Verified decision/measurement loops are distinguishable from infrastructure activity | SUPPORTED |
| C-002 | The checker detects every defect class in its sabotage suite | SUPPORTED |
| C-003 | The portable core transplants into a second project without domain import or overwrite | SUPPORTED |
| C-004 | Using SEED improves false progress, recovery, or human load versus bare | HYPOTHESIS |
| C-005 | A negative result can be published without being rewritten as success | SUPPORTED |
Falsifiers and bound evidence: CLAIMS.md. Do not upgrade a row by repetition.
Case study (author-run, not independently reproduced).
Throughput is not progress — a census of 1,979 PR-linked commits on poketcg (a private repository) at pin 9522a8a. Pre-registration: workflow/experiments/2026-08-24-pr-case-study.md. Public dataset: artifacts/.
Instrument calibration. Cold-start comprehension and deceptive-green scoring are registered under pocs/. Transplant of the invariant core is receipt-bound (C-003).
What the instrument contains (not proof it helps). Factory bins orient.py, bootstrap.py, and selftest.py are runnable in this clone. Doctrine and harvest ledgers stay in-tree and are not the landing page.
| Path | Role |
|---|---|
JUDGES.md |
Score the instrument, not C-004 |
CLAIMS.md |
Public claims, status, falsifier, bound evidence |
LAWS.md |
Incident-earned rules and enforcement grade |
START-HERE.md |
Machine door |
workflow/experiments/ |
Pre-registrations (PASS / KILL / NULL / INVALID-INSTRUMENT) |
workflow/research/2026-08-24-pr-case-study/ |
Methods, paper, public artifacts/ |
workflow/receipts/ |
Banked work; “done” without evidence is invalid |
workflow/canon/RETRACTIONS.md |
Numbers that were published and are now false |
Watanabe et al. (arXiv:2509.14745) study merge rates of Claude Code pull requests. MSR 2026 mining-challenge papers taxonomize agent-PR communication and human intervention. Those datasets answer acceptance. The case study here asks a different question: what a single-operator factory’s merges are made of when GitHub is the coordination substrate.
Use CITATION.cff. Preferred citation is the case-study report, with the author-run caveat attached.
A DOI is minted via Zenodo from GitHub releases (.zenodo.json carries the deposit metadata); cite the DOI of the release you used once one exists.
MIT. See LICENSE.
See CONTRIBUTING.md. Research norms; author XOR verifier. Default mode is one strong agent. Fleet work is opt-in for independent, worktree-isolatable tickets only.
SEED measures whether agent-driven work constitutes verified progress. BigBoss controls which autonomous actions can occur and preserves human decision authority. The Council tests independent verification through heterogeneous model families. The Bus shows adversarial review terminating a bad architecture before further implementation. Godot Methodology tests whether the same verification principles generalize into software architecture. Founder-led, independent: simoneresearch.com.