Threshold Port

Threshold Port: keep the value, drop the ritual, move the number

Bob's sharpest line in the conversation draws the whole map: "it's probably a mistake to impose a human discipline on an agent. It is not a mistake to impose human values on the agent, but there may be thresholds that we need to change" (C17, ledger). This island turns that line into a repeatable port. TDD, a coverage floor, a complexity cap: each one crosses to the agent lane in three moves, and the number at the end moves only on evidence.

Scope: numbers, not markdown

This island targets numeric gate parameters. Complexity caps, coverage floors, CRAP cutoffs, mutation budgets are what gets ported here, never markdown files. Tuning a skill, a CLAUDE.md, or any prose context file is skill-tune's lane entirely.

The experimental engine this island leans on is skill-tune's too. The measure-edit-keep-iff-score-moved loop belongs there and is referenced here, never re-derived. Run that loop with the gate level as the one edited variable and lane outcomes as the score. This island supplies the rest: what to keep, what to drop, and what counts as a win for a number.

Move 1: keep the value

Name what the practice actually protects, as a property a tool can measure: tested, covered, small, clean. Values transfer intact (C17). If the value resists measurement, the port is not ready for move 3. Carry it as advisory prose and say so, until a measurable proxy exists.

Move 2: drop the ritual

The ritual is the human-wired procedure around the value, the part shaped by human working memory, boredom, and habit. Classify a rule as ritual when it constrains the order or rhythm of the work rather than a measurable property of the result.

The canonical case (C17) is strict TDD line-interleave, which Bob explicitly does not impose: "I don't think it makes any sense to make an agent write a single line of a test and then write a single line of the production code." Agents naturally "write a function and then write the test for that function", and "They always fall back on doing that… So I figure that's probably okay." The value crosses whole: every path covered by a test. The red-green rhythm stays home.

Move 3: retune the threshold by controlled experiment

Human-calibrated numbers carry human assumptions, so the ported level starts as a hypothesis. Bob's own path is the template (C17): "for a human I would keep crap numbers below four… but for the agents I've set this at six and… maybe I'll push it to eight." Here is skill-tune's loop, run on a number:

1. Baseline at the current level. Record real lane outcomes: defect escapes, thrash incidents (change-one-break-another loops), gate-failure rate, productivity margin.
2. Change one level: one variable, one step. Go 4→6, then 6→8; a 4→8 jump makes the delta unattributable.
3. Run real work at the new level, long enough to collect the same outcomes.
4. Keep iff no worse: outcomes held or improved. Otherwise revert and log why, so the next porter does not retry it.

The hard rule: evidence decides, agents propose

Agent opinions enter the experiment only as hypotheses in step 2. Bob polls his agents on threshold levels and treats the answers as color, never as a decision: "you can't trust any debate you have with an agent" (C18). A level that traces to an agent vote instead of a measured step is authority without evidence, wearing a port as a disguise.

The rationale is a tunable too

Bob justifies the looser agent level by "a huge short-term memory and a perfectly accurate short-term memory" (C17). That is his hypothesis, not established research, and research/crap-metric.md flags it unverified. Treat the rationale exactly like the number. If agents thrash at 8, the hypothesis lost, however good the story. Outcomes outrank rationales in both directions.

The port record

One row per ported practice, kept beside the gate config it calibrates:

practice    value kept (measurable)      ritual dropped           level path        evidence per step
TDD         all paths test-covered       line-interleave rhythm   n/a (no number)   C17 precedent + lane baseline id
CRAP gate   small + fully tested fns     manual per-fn review     4→6 (kept)        outcome rows per step, run ids

The evidence per step cell names the recorded outcome comparison for each level change. Where no comparison exists, it reads unverified — inherited human level, not yet tested. Formula, regimes, and per-language tooling for the CRAP example live in research/crap-metric.md; the values-not-disciplines source line is traced in research/martin-canon.md.

Done when

• the value is stated as a measurable property and the destination gate measures it;
• every dropped rule is classified ritual via the order-or-rhythm test;
• every level in force is the latest step of a one-variable experiment with recorded outcomes, or is explicitly marked unverified — inherited human level, not yet tested;
• no level traces to an agent poll;
• the port record shows all of the above.

Verify → fix → re-verify: walk the record row by row against these five conditions, repair the first miss, then walk again from the top until a full pass.

Enforced vs advisory

• Enforced (today) — this island's own shape: python3 scripts/validate-island.py skills/threshold-port, run from the pack root, exits 0. In a destination repo, a ported threshold is enforced only where its tool blocks on breach with a non-zero exit (crap4py already does; see research/crap-metric.md). Until that wiring exists, the level is advisory and the port record must say so.
• Advisory (v0) — the three-move classification, the one-step-at-a-time discipline, the no-agent-vote rule, and the port record format. No script here checks them yet. A later wave can add a port-record linter; until then this is honest guidance, and calling it enforced would be laundering.

No authority without evidence. Values transfer; rituals stay home; the number moves one measured step at a time.