night-market-research-methodology

The discipline that turns a hunch into an accepted result in this repo. An "accepted result" is a change that survived the evidence bar and landed through change control as a rule, a skill module, a config gate, or an ADR. Everything else is either a local working note or a documented retirement. This skill covers the full path: score the idea, experiment behind a default-off flag, meet the evidence bar, land the durable artifact, or retire the idea on the record.

The evidence bar

A claim graduates from hunch to result only when it passes all four tests.

1. One mechanism explains all observations, including negatives. If the hypothesis explains the three failing cases but not why the fourth case passed, it is incomplete. Keep digging until a single mechanism accounts for everything you saw.

2. Predict numbers before running. Write down the expected measurement first, then measure. In-repo anchor: the forced-eval harness labels expected activations in prototypes/forced-eval/activation_cases.json before any run, then compares baseline against treatment with a McNemar paired test (a significance test for paired binary outcomes).

3. Survive assigned adversarial refutation. Assign a reviewer or agent whose explicit job is to break the claim. Use Skill(attune:war-room) for hard-to-reverse decisions and Skill(imbue:rigorous-reasoning) to counter agreement bias. A claim nobody tried to break is unproven.

4. Never let the generator judge itself. The agent that produced the work must not be its sole verifier. See plugins/imbue/skills/proof-of-work/modules/independent-verification.md. Prefer executable checks over an LLM judge, and prove the check can fail before trusting it (Guards 2 and 3 in plugins/imbue/skills/proof-of-work/modules/verifier-integrity.md).

Corollary from verifier-integrity: a green check proves the code satisfies the spec as written. It cannot prove the spec says what you meant, and it proves nothing if the check cannot fail. Validate the spec separately from the code, and mutation-test the check itself.

Idea lifecycle

An idea moves through four gates in order. Skipping a gate is how speculative infrastructure gets built and reverted.

Gate 1: score worthiness before building

Formula and thresholds from docs/backlog/queue.md (a local, gitignored working file):

Worthiness = (Business Value + Time Criticality + Risk Reduction)
           / (Complexity + Token Cost + Scope Drift)

Score	Action
> 2.0	Implement now
1.0 to 2.0	Discuss before proceeding
< 1.0	Keep in backlog

Queue rules: at most 10 active items. Items untouched for 30 days are archived to a GitHub issue (labels backlog,deferred) and removed from the queue. Because docs/backlog/ and docs/research/ are gitignored, the durable record of a deferred idea is the issue, not the queue file.

Gate 2: experiment behind a default-off flag

Exemplar: the egregore completion-integrity gate.

• Commit 83281337 added the gate with completion_integrity: bool = False in plugins/egregore/scripts/config.py (still False as of 2026-07-02).
• Commit cd903cbf added a test covering the raw-JSON opt-in path.

Pattern: land the mechanism off by default, cover the opt-in path with a test, and collect usage before proposing a default change.

Gate 3: data-collection window before structural change

ADR-0015 (docs/adr/0015-orchestrator-skill-simplification.md) requires 30 days of usage data before simplifying the over-built orchestrator skills. Apply the same bar to any promotion or simplification: name the data window in the PR, not an intuition.

Gate 4: adopt through change control or retire on the record

Adoption goes through the process in night-market-change-control. Retirement is written down, never silent. ADR-0012 (confidence-tagged claims) and ADR-0013 (Naur theory-building) carry Status: Superseded by ADR-0017, which is Accepted and rules "Do not build an enforcement mechanism. Permit voluntary use." A documented no is a valid result.

The research-to-rules pipeline

1. Run multi-channel research (Skill(tome:research) or manual) into a dated synthesis at docs/research/YYYY-MM-DD-<topic>.md. Match the shape of the existing docs: Thesis, What the evidence says, solution pattern, Mapping to the night-market ecosystem, Evidence gaps and caveats.

2. Map every gap against existing ecosystem assets before proposing new code. Most gaps turn out to be covered already (see case study 3).

3. Land each real gap as the smallest durable artifact: a .claude/rules/ file, a module inside an existing skill, or a config gate. A new skill is the last resort (.claude/rules/shared-utility-consumer-rule.md requires 2+ consumers within 30 days).

Caution: docs/research/ is gitignored, and at least one committed rule (prefer-invariants-over-fallbacks) cites a research-doc path that exists only on the authoring machine. Put the load-bearing evidence and citations into the rule or ADR itself. The research doc is background, not the record.

Case study 1: coming loop (one doc, two artifact types)

The 2026-07-01 synthesis of Armin Ronacher's "The Coming Loop" pulled in the METR randomized trial (arXiv 2507.09089: 16 developers, 246 tasks), GitClear 2025 (211M changed lines), and Karpathy's "mortal terror of exceptions". It produced two artifacts:

• a review-time rule, .claude/rules/prefer-invariants-over-fallbacks.md (commit 9f771794), and
• a runtime gate, egregore completion_integrity, default off (commit 83281337).

Lesson: one research doc can fan out into different artifact types. Match the artifact to where the failure occurs (review time versus runtime).

Case study 2: prover-verifier (module, not skill)

The 2026-07-01 prover-verifier synthesis landed as commit 29081fda: a 146-line module with six guards, plugins/imbue/skills/proof-of-work/modules/verifier-integrity.md, inside the existing proof-of-work skill. No new skill was created. Lesson: extend the consumer that already exists.

Case study 3: karpathy-derivation (build only the delta)

docs/karpathy-derivation/project-brief.md (tracked in git) maps four Karpathy principles against existing skills in a coverage matrix and concludes "~90% coverage exists." Only the delta was built: imbue:karpathy-principles, a compact synthesis with an anti-pattern catalog. The same matrix was later reused as the lens for the April 2026 skill audit. Lesson: run the coverage analysis first. The most common honest research outcome is "we already have this."

The audit protocol

As practiced in Discussion #449, the April 2026 skill-audit synthesis (category [Knowledge]).

1. Tier 1 first: git history and rg scans (Skill(pensive:tiered-audit)). Escalate only what Tier 1 flags.

2. Targeted parallel agents require output contracts, per .claude/rules/plan-before-large-dispatch.md (plan mode at 4+ agents). Contract schema in plugins/imbue/skills/proof-of-work/modules/output-contracts.md: required_sections, min_evidence_count (minimum [EN] evidence tags in findings), strictness (strict/normal/lenient). Findings carry file:line evidence. An empty findings list is a valid result. Report "no findings" as such rather than padding.

3. Findings land as waves of inline fixes.

4. Policy-shaped findings become issues, and landed issues become rules. Issues #454 (Exit Criteria required in every SKILL.md) and #457 (utility skills need 2+ consumers) both followed this path, are CLOSED, and live on as .claude/rules/skill-exit-criteria.md and .claude/rules/shared-utility-consumer-rule.md.

5. The deferred remainder gets a tracking issue: #574 (Wave-3 skill-audit backlog, OPEN as of 2026-07-02).

Reading Discussions requires GraphQL. The gh discussion subcommand does not exist:

gh api graphql -f query='query {
  repository(owner: "athola", name: "claude-night-market") {
    discussion(number: 449) { title body }
  }
}'

Proof-and-analysis recipes

Activation lift: does a skill actually fire?

prototypes/forced-eval/ (commit 5683e89b) measures whether a forced-evaluation hook lifts skill activation:

• activation_cases.json holds labeled prompts with expected Skill() activations, recorded before measurement.
• measure_activation.py runs each prompt via claude -p --output-format stream-json --max-turns 1 --allowedTools Skill, baseline (hook off) against treatment (hook on), and applies the McNemar paired test. True-negative cases count false activations, so a high positive rate alone is not treated as success.

Status: PROTOTYPE, not wired into any plugin.json. The harness is unit-tested but the live lift is unmeasured as of 2026-07-02 (the README says so). Run the harness tests:

uv run python -m pytest prototypes/forced-eval/ -q

Verified 2026-07-02: 20 passed.

Mutation testing: are the tests real?

Mutation testing mutates source code and checks whether the tests notice. A surviving mutant is a test that cannot fail on that behavior, which is the "hollow check" failure mode from verifier-integrity Guard 2. CI runs it weekly plus on dispatch (.github/workflows/mutation-testing.yml). Exit codes: 0 means no survivors, 2 means survivors found, anything else is a crash. Local, per plugin:

cd plugins/<plugin>
uv pip install mutmut --quiet
uv run mutmut run --paths-to-mutate=scripts/,src/ --tests-dir=tests/

Adjust --paths-to-mutate to the directories that exist. CI builds the list from the plugin's top-level scripts/ and src/ dirs.

Ratchet baselines: debt burndown you can prove

A ratchet baseline freezes today's debt count in a JSON file. The check fails only when new debt appears, and prints when the count drops so you can tighten the baseline and lock the win. Two live ratchets, both pre-commit hooks and standalone scripts:

python3 scripts/check_skill_graph_drift.py
python3 scripts/check_skill_exit_criteria_drift.py

Verified output on 2026-07-02: dangling Skill() refs at 5 against a baseline of 31, and SKILL.md files missing Exit Criteria at 1 against a baseline of 127. Each script names the baseline key to lower. The shrinking baseline diff is the burndown proof: cite it in the PR.

Where good ideas came from

Source	Path taken
External research	Ronacher, METR, Karpathy syntheses became rules and gates (case studies above)
PR-review pain	Recurring finding classes became pre-commit guards and .claude/rules/ entries
Audits	Discussion #449 became issues #454/#457, which became rules
Incident lessons	See night-market-failure-archaeology for the chronicle

When NOT to use

• Running tests, coverage, or the evidence gates for a concrete change: use night-market-validation-and-qa instead.
• Classifying, gating, and landing a change: use night-market-change-control instead.
• Mechanics of Discussions, the decision journal, or ADR practice: use night-market-collective-memory instead.
• Understanding settled incidents and reverts: use night-market-failure-archaeology instead.
• Choosing an open problem worth attacking: use night-market-research-frontier instead.
• Executing the completion-integrity work: use night-market-completion-integrity-campaign instead.

Exit Criteria

• The idea has a written worthiness score with all six factors, and any score at or below 2.0 was discussed or queued, not built.
• Predicted numbers were recorded before the measurement ran.
• One mechanism explains every observation, including the cases that did not fail.
• Verification names an independent check the generator cannot influence, and that check has been shown able to fail.
• Any experiment shipped behind a default-off flag with a test covering the opt-in path.
• The outcome is on the record: a landed rule, module, gate, or ADR, or a superseding ADR documenting retirement.

Provenance and maintenance

Compiled 2026-07-02 against repo v1.9.15, branch discussions-fix-1.9.14. Commit anchors (9f771794, 83281337, cd903cbf, 29081fda, 5683e89b) are stable. Volatile facts and one-line re-verification:

• completion_integrity still default False: rg -n "completion_integrity" plugins/egregore/scripts/config.py
• Ratchet counts (5/31 dangling refs, 1/127 missing Exit Criteria on 2026-07-02): rerun python3 scripts/check_skill_graph_drift.py and python3 scripts/check_skill_exit_criteria_drift.py
• Issue states (#454 CLOSED, #457 CLOSED, #574 OPEN on 2026-07-02): gh issue view 574 --json state -q .state
• Forced-eval lift still unmeasured: rg -n "not measured" prototypes/forced-eval/README.md
• Research and backlog dirs still gitignored: git check-ignore docs/research/x.md docs/backlog/x.md
• Mutation exit-code semantics: rg -n "Exit codes" .github/workflows/mutation-testing.yml
• Worthiness thresholds: reread docs/backlog/queue.md. It is a local file, absent on fresh clones. The thresholds are restated above.