Type Realization

Measure whether one named protocol skill's declared contract is realized during an actual run. The first argument is required; /realize inquire selects only the registered inquire cases and graders. An unknown or omitted target fails before setup.

Purpose

A protocol's formal blocks are runtime-normative: TYPES, PHASE TRANSITIONS, TOOL GROUNDING and the Rules type the prose and carry the operational contract. Static checks establish that a SKILL.md contains those blocks and enforce a bounded set of structural invariants. Nothing establishes that a run follows their transitions.

This skill supplies evidence for that gap. It drives Claude Code or Codex across a matrix of models and treatments, captures the JSONL trace, grades the deterministically observable subset, and names the transcript judgments still owed. A complete automatic cell can establish that collection occurred, the declared branch reached Stop, or the zero-uncertainty branch reached Proceed; semantic ordering and rendered type coverage remain manual until their specified judges are built.

Judgment boundary

The object of judgment is the formal transition trace:

X → context sufficiency → evidence collection as needed → residual deficit
  → formal branch → Stop | Proceed

Every score attaches to a node or edge in that trace. Once the selected branch is observable, judgment ends. A plan, implementation, analysis, or other downstream artifact is evidence that Proceed occurred; its quality, correctness, completeness, and usefulness are outside /realize. Likewise, an unchanged substrate is evidence of Stop only where the case granted the capability to proceed and the formal branch required stopping. references/grader-design.md carries the witness design and the separate oracle that artifact-quality evaluation would require.

When it applies

Reach for this after changing a protocol's phase structure, terminal conditions, gate definition, or answer types — the places where a contract can become unrealizable without any static check noticing. Reach for it also when adopting a new model, where the question is whether the contract still holds on it.

Do not reach for it to check that a file is well-formed. That is /verify, it is free, and it runs on every commit.

Workflow

cd .claude/skills/realize/scripts

./setup.sh inquire                           # default Claude runner
export CLAUDE_CODE_OAUTH_TOKEN="$(...)" && ./run.sh inquire

REALIZE_RUNNER=codex ./setup.sh inquire      # no credential consumed or stored
REALIZE_RUNNER=codex ./run.sh inquire        # borrows this machine's codex login
REALIZE_RUNNER=codex ./teardown.sh inquire

./teardown.sh inquire                        # Claude volatile state

For Claude, setup.sh prints the one interactive step — obtaining a token against the target-specific isolated config directory. For Codex, setup creates bare/protocol homes and installs the local protocol plugin only into the protocol home without reading or writing a credential. run.sh authenticates each codex exec child with the codex login already on this machine (codex login --device-auth signs one in; the harness never calls codex login), symlinked into the arm's disposable home only for the span of each codex exec — never copied, because a ChatGPT login rotates its refresh token and a copy would strand the real one. references/runbook.md carries why that keeps the homes isolated.

Read references/runbook.md before the first run. It records where runner isolation lives, where the budget floor sits, and what each column of the report asserts.

From CI

.github/workflows/type-realization.yml runs the Claude path on manual dispatch and comments the report on the PR. Codex remains local-only until the matrix can run behind the official action's credential proxy; repository-controlled harness code does not receive an OpenAI key in Actions.

Configuration

harness.config.json carries shared runner settings and a target registry. Each target owns its plugin directory, skill id, invocation, and case set. REALIZE_RUNNER=codex selects the committed Luna xhigh profile. Results are keyed by target, runner, the case's subject-facing files (prompt and replies after frontmatter, case.yaml, the scaffold script), and a hash of the actual protocol/style treatment — for protocol arms, the skill's references/ too — so one skill, ablation, or edited case cannot reuse another's cache.

Prefer a capable model for the primary measurement. The weakest available one exercises the safeguards but not the protocol, so a failure there cannot separate a defect in the contract from a limit of the model.

Arms

Claude has four arms crossing the protocol against the Epistemic Ink output style, and one that removes the protocol's formal blocks. The style ships in the cc-plugin marketplace's ink-figure plugin; styleSource in harness.config.json names the file the style arms read, so those arms need that marketplace installed:

protocol-prose runs only when named in REALIZE_ARMS. It loads a copy of the plugin rebuilt at each run with the Lean blocks removed from its SKILL.md files, and refuses to run when the copy lost none. Read against protocol: a transition realized in both arms is realized without the formal blocks, and one realized only under protocol points to what they add.

The sham arm is not a construction. Published work on rule files for coding agents found random rules helping as much as curated ones, which makes "a long structured instruction is present" a live alternative explanation for any positive result. The output style is that alternative made available as a control: it fixes gate shape and observer markers while fixing none of a protocol's own obligations.

Codex has bare and protocol arms. Codex has no equivalent of Claude's shipped output-style treatment, so requesting style or protocol+style fails rather than simulating a different deployment condition; protocol-prose is likewise Claude-only. The Codex protocol treatment is present only when codex plugin list reports that plugin installed and enabled in the isolated protocol home while the bare home reports it absent. Codex JSONL currently exposes no separate skill-invocation event, so its skill report cell states that limitation rather than inferring invocation from the model's prose.

Cases

Every case runs under explicit invocation. Each registered target needs at least a case where the protocol's obligations must be realized, and one that holds the counterpart the protocol must not fabricate — for /inquire, a collection that leaves nothing open and relays; for /grasp, an answer with nothing to check it against. Whether a protocol is selected at all, or stays silent, is measured by route's own eval, not here. The registered targets are the keys of targets in harness.config.json; no result is implied for a protocol absent there.

A target's paired cases mount the same scaffold, deliberately: for /inquire, one observes whether a file-discoverable fact was asked, the other whether a supplied parameter was re-asked, and differing substrates would let a run pass one by luck. A case names its scaffold in case.yaml (scaffold_script); one that names none mounts evals/scaffold.sh.

Write the prompt as the task alone. The line that invokes the protocol lives in harness.config.json and reaches only the arms that have it — a prompt naming the command makes an arm without the plugin gate on the missing tool instead of on the task.

Read every clause of a negative case as an adversary would. A specification that looks exhaustive can still leave a term underdetermined, and a correct protocol will open a gate on it — recording a failure that belongs to the case author.

evals/inquire-underspecified/ and evals/inquire-fully-specified/ are the worked pair for /inquire. Follow their shape when adding a protocol: a prompt.md carrying frontmatter and the user's words, and one grader per obligation under graders/. A grader checks that the fields a judgment produces exist and are faithful to their sources; it does not grade the judgment itself, such as whether a ground is sufficient.

A protocol whose obligations fire only after the user answers needs turns past the first Stop. A case reaches them by declaring multi_turn in case.yaml. Where its user side can be written as a fixed script — driver: harness, the replies in reply-1.md, reply-2.md, … — the harness sends them itself, on either runner, resuming one session; evals/grasp-adjudicable/ and evals/grasp-unattachable/ are that pair for /grasp, and their oracle.md says why every reply is written to stand at whichever gate it lands on. An oracle that must read the subject's turn to compose a reply, as /elicit's does, is walked by hand (references/runbook.md) and refused if registered.

Reading results

Read integrity first. It reports whether each arm's treatment actually applied — both dimensions of it, the plugin and the output style, since either can fail silently and leave two arms running the same treatment. A row whose integrity falls short of its run count is not evidence about the protocol, and the report reprints those rows separately so they are not mistaken for findings.

pass_k is one only when every repetition passed its deterministic transition predicates. The manual column counts scenario-specific transcript judgments excluded from that composite; the report names them. For inquire, collection-before-surfacing order, unasked cheap evidence, faithful basis, kept ownership, stated answer openings, the relay of a collection that left nothing open, and the absence of a design gate remain manual observations grounded by the grader files. For grasp, the automatic set is what both cases share — the target read in the first turn, the tree unchanged after every turn, every turn reported — and the absence of a test, the grounded opening map, the one-round correction, a check's result set against both readings, the withheld verdict with its need kept on the map, the stop at each gate, and closure on the user's word are manual. For conduct, the only automatic predicate is every turn reported: whether the method's work started — the map's stop, the taking's proceed, the relay's proceed — is read from the transcript, so each transition is manual along with what its turn presents. Whether the tree differed from the scaffold after each turn is recorded in the cell's sidecar as an observation those graders may read, not as a verdict. The predicates column breaks pass_k down by predicate; turns shows how many scripted turns a multi-turn cell reached.

On Claude, skill says whether the protocol fired where it was available, and n/a where there was no plugin to fire. Codex reports trace-unavailable for that column and keeps plugin installation integrity separate from behavioral fulfillment.

Absolute transition-fulfilment rates are the reportable automatic quantity here, not deltas. A baseline arm has no gate to stop at, so most predicates have no counterpart there to subtract.

For Codex, the report records token use rather than inventing a dollar amount the CLI did not emit. Every requested cell must produce a complete thread.started → turn.completed pair. A launch failure, missing cell, unreadable predicate, or treatment-integrity failure makes run or report exit non-zero while preserving partial transcripts for inspection.

Additional Resources

• references/runbook.md — the workflow, where isolation lives, how to widen the matrix, how to read the report.
• references/grader-design.md — why the deterministic axis is behaviour rather than wording, the obligation-to-predicate mapping, what the arms answer, and the structured-extraction and metamorphic-validation passes that are specified but not yet built.
• scripts/harness.mjs — the runner and the deterministic graders. Node standard library only.
• evals/*.sh — the fixtures; each case selects its own through scaffold_script in case.yaml.