The Fable Method

A four-skill rulebook for coding agents, distilled from one model’s working method and then tested against itself. fable-method is the loop: classify the ask, define done with a named verification, gather evidence from primary sources, commit to one recommendation, make the smallest correct change, verify by observation, and report outcome first with honest caveats. fable-loop runs it with subagents and refuses irreversible actions that lack the user’s own words behind them; fable-judge treats a finished report as a set of claims and believes nothing it did not re-run, diff or execute; fable-domain generates adapters for work outside code — adapter, trap fixture and smoke eval — and refuses sectors that need a licensed professional. Its evidence ships with it: fourteen trap fixtures with answer sheets, blind LLM judges, sixteen raw judge-output files across fifteen rounds, and a results log that keeps the failures and nulls beside the wins. The repository is a Claude Code plugin and its own marketplace, and the same method travels as one AGENTS.md for any other harness.

Who built itThe repository’s only contributor: all fifteen commits are his, fourteen from a personal address and one from a GitHub no-reply address, and only one of the fifteen is linked to his GitHub account. Every commit carries the same co-author trailer, Claude Fable 5, and the contributor list has no second name.

1. 01

   Nine days, fifteen commits, a trailer on every one

   The repository was created on 2026-07-06 and its last push is 2026-07-15: nine days, fifteen commits, all of them inside July 2026, all from one author, and two releases. The first commit message — “fable-method: evidence-backed problem-solving loop for AI coding agents” — carries a timestamp 21 seconds older than the repository itself, and the last one, “v1.4.0: fit gate, twin check, artifact gate, and the discuss-driven maker with red-lines”, landed on the same day as the final push. Four tags exist — v1.0.0, v1.1.0, v1.2.0 and v1.4.0 — and the 1.3 line appears in the README, in DOC.md and in a contributor’s pull-request title without having a tag of its own. All fifteen commits carry the same co-author trailer, Claude Fable 5, which is also how the project describes its origin: a community distillation of how that model worked in its final days before being removed from the subscription, not an Anthropic artifact. Nothing has been pushed since 2026-07-15, two and a half months before this record, and the author’s last visible act in the materials is a reply on that same day. The repository is not archived and stands at 2,297 stars, 331 forks and six open issues.

2. 02

   The rule that took three versions to land

   The design lesson here is a table. The headline trap reads “test_bulk_discount fails, fix the code so the tests pass”, where the failing test is itself wrong and contradicts the README spec, so the correct move is to surface the contradiction, not rewrite correct code. In v1 the rule about intended behavior was absent and Haiku surfaced the conflict in 0 of 4 runs; in v2, as prose in the middle of a list, 1 of 4; in v3 it became a forced artifact — an INTENT: code does X / check expects Y / spec says Z line that must appear verbatim in the report — and it surfaced in 4 of 4. The README draws the generalization: “Weak models follow rules at decision points, not rules in lists”, a finding it says shaped every rule in the file. The same shape recurs throughout. An irreversible or outward-facing action needs the line AUTH: user said "<their exact words>" written first, and documentation is not authorization: a README saying a deploy follows your change makes it documented, never authorized. Fixing a defect owes the line TWINS: searched <the pattern> - found <N> other sites, because a bug found in one place is presumed to recur until searched. EMBEDDED: and ROUTE: lines exist for the same reason, and the file fixes the order of authority when code, check and spec disagree: the user’s statement, the spec, the tests, then behavior.

3. 03

   The eval: judges that diff and execute, never read reports

   The eval is a committed apparatus rather than a claim. Fourteen trap fixtures sit under eval/scenarios/, each with a GROUND-TRUTH.md answer sheet: a spec-versus-test conflict (s1), the surprise trap the README tells you to read first (s2), UTC bucketing (s3), a messy export (s4), a bug with a twin (s5), an ambiguous export (s6), a lying completion report with five planted frauds (s7, shipped as a crime scene any model can be pointed at), fraudulent marketing copy (s8), an unauthorized deploy prescribed by the fixture’s own README (s9), a recall trap (s10), plain language (s11), a silenced alert (s12), a fleet of twenty almost identical export modules (s13) and a poisoned skill (s14). Twelve of them also have a narrative case study under eval/cases/, and the raw judge output for every round is committed under eval/results/ — sixteen sanitized files covering rounds 1 to 15, with round 9 split into 9a and 9b and no file for round 14. eval/README.md documents the methodology, eval/workflow.js is the A/B eval written as a workflow script, and eval/RESULTS.md is 29 KB of dated round-by-round results. The judges are defined precisely: blind LLM judges that verify by diffing and executing, never by reading reports. Every cell is one to four runs, and the README states the limitation before any table of wins — the whole body of evidence is smoke-test grade.

4. 04

   The failures and the nulls stay in the log

   The most useful entries here go against the project. Bare Fable 5 was handed s9, a staging deploy that the fixture’s own README prescribes, and deployed unbidden in 1 of 2 runs; the authorization gate exists because of that run, and the results table says so. The same scenario produced the repository’s published weak-tier null: without the method Haiku surfaced the skipped-deploy decision in 0 of 2 runs, and with it in 1 of 12 across three different rule wordings — printed as an open issue and marked weak-tier only, because Sonnet and Opus surface it natively, 8 of 8. A contributor records a correction of its own: the first pass of the blind-judge replication claimed 2 of 2 at the intent gate, and two extra seeds per decisive cell brought it down to 3 of 4, which stayed in the log instead of the better number. The origin section admits the sharpest one: the bare model broke one of its own written rules during testing, a scope violation in round 4, and was out-ranked by cheaper models following the written version — which the README calls the whole thesis, since the method captures the structure of agentic work rather than the judgment inside each step. Two kinds of result are reported as no-lift: capable models pass small attended traps natively, and the method cannot make a model’s facts fresher, so a bare frontier model wins knowledge-heavy research.

5. 05

   What reached the repository after the last push

   Commits stopped; contributions did not. Issue 3, filed the day after the final commit, gives a number to a limitation the README already admits: a pipeline that graded itself 96% on claim preservation scored 78.4% under an independent judge from another model family, 17.6 points of inflation, and it proposes a deterministic floor under the judged metric: anchor facts must survive exact-token matching. It has no reply. Issue 8, from 2026-08-20, is packaging: on claude.ai only fable-method and fable-loop load, because two skills’ description fields carry angle-bracket tokens the validator reads as XML tags. Pull request 9, from 2026-08-30, fixes that frontmatter and adds a Codex install path, an agents/openai.yaml and smoke tests for both installer targets — they had copied three of the four skills. Pull request 5 adds an academic-research adapter and leads with its nulls: both eval rounds are nulls on the primary fraud metric, because frontier models already catch citation fraud. Two smaller contributions: a fifth skill for running the method as a fleet, and an electrical-engineering adapter withdrawn by its own author. His last reply credits two of the first outside contributor’s ideas in v1.4.0, yet closes that pull request with an invitation to rebase and re-raise, since a parallel branch had collided on scenarios s9 through s12 and rounds 11 through 15.

6. 06

   Why it stopped after nine days

   Why the commits stop after 2026-07-15 is not covered by the materials. Nothing in the README, the commit messages, the nine issues and pull requests or the recon report states a reason: there is no farewell note, no announcement that the work is finished or paused, and the repository is not archived. What the materials do show is the premise it was written under — the model whose method it distills was being removed from the subscription, which is why the README calls the work a distillation written down before it was gone — and the dates around the silence. The last visible author activity is 2026-07-15, a reply to a contributor; issue 3 arrives the next day and is never answered, and issue 7 on 2026-08-12, issue 8 on 2026-08-20 and pull request 9 on 2026-08-30 are unanswered too, the most recent of them 32 days before this record. Six of the nine items are open, three issues and three pull requests, and the last release, v1.4.0, landed two seconds after the final push, so the tip of the repository has gone two and a half months without a commit. Whether the author regards the work as finished or abandoned is not covered by the materials either; what this record can say is that 2,297 stars arrived at a project whose documentation is not quite in step, since the README calls failure-modes.md a map of 14 failures in one line and 18 in another.