Model Tuning — Claude Code Calibration

GPT-side counterpart: gpt-tuning.md — gzkit runs with either vendor (operator ruling 2026-08-02); this page calibrates the Claude side, sourced to the current Anthropic card in data/frontier_model_cards.json.

Lifted from CLAUDE.md under GHI #327 diet pass. The binding summary remains in CLAUDE.md § Model tuning; this page holds the calibration and its sources. Every value here is "as measured on the named model" and expires with it — see § Recalibration on model change.

Profile precedence — Opus first

Operator direction (GHI #943, verbatim): "favor opus over fable".

1. Opus 5.5 is the default Claude profile. When no profile is selected, or when Opus and Fable guidance conflict, the Opus rule wins.
2. Fable 5.1 is a subordinate, explicit profile. Its nudges apply only in a Fable session and never loosen the Opus scope boundary.
3. The universal contract stays vendor-neutral. AGENTS.md carries the portable scope rule (PRIME DIRECTIVE, first bullet: complete the requested behavior and every coupled correctness surface; route unrelated defects); model-generation tuning lives here and in CLAUDE.md.

The governance fence: nothing on this page removes or weakens a deterministic gzkit check, a semantic test, an ARB receipt, a governed review stage or human Gate 5. "Remove redundant verification prompting" is about model-facing duplication of what the model already does, never about the evidence chain.

Adaptive regulation

Current Opus-family models are adaptive — they regulate thinking per turn against the prompt's apparent difficulty. Do not pin fixed thinking budgets; prompt the calibration explicitly when the default doesn't fit. On Opus 5.5 thinking is always on and cannot be disabled; effort is the control, and lowering it reduces thinking more reliably than a prompt instruction does (Prompting Claude Opus 5.5).

Effort is a dial, not a default

Effort/quality is non-monotonic and workload-dependent. Do not assume more effort is better; re-baseline per workload. Measured on Opus 5.5:

Benchmark	Peak effort	Card evidence
FrontierCode v1.1 (Main / Extended)	medium	54.6 % / 65.3 % at medium; scores decline above medium and mostly recover at max (54.4 % / 63.6 %). The card ties the decline to grading that "penalizes out-of-scope changes that may be unnecessary" (§ 8.4)
CursorBench	max, narrowly	57.8 % at max, 56.0 % at xhigh and high, 52.5 % at medium (§ 8.8)
Terminal-Bench 4.0	xhigh	max is within noise of xhigh (§ 8.5)

Operating guidance:

• Start at medium, Opus 5.5's default, and sweep. The guide's measurement: at medium Opus 5.5 matches or beats its predecessor at high on coding and knowledge work, and on several coding evaluations low comes close at much lower cost. Effort names do not denote the same amount of thinking across models, so a level carried over from an earlier model is not a calibration.
• Above medium, the risk on scope-graded work is out-of-scope change, and the mitigation is a written scope boundary. gzkit carries it as AGENTS.md § DO IT RIGHT #11 (surgical changes) and OBPI allowed-paths. That boundary governs taste-driven change only: fixing a defect found in flight is a § PRIME DIRECTIVE duty, which the operator ruled prevails (2026-09-24, GHI #1091). The Fable 5.1 card measures a brevity-and-scope instruction reducing those edits (§ 8.4); the Opus 5.5 card reports the decline and its cause and makes no claim about a prompt remedy.
• Reserve xhigh and max for work where a quality gain has been measured. At a given level Opus 5.5 thinks more per turn than its predecessor, most at xhigh and max, so a carried-over high setting costs more than it did.
• Drop to low for cost- and latency-sensitive work — status answers, lookups against a known path, simple grep-and-report.

Opus 5.5 profile (default)

Sourced from the Claude Opus 5.5 System Card (2026-09-22, §§ 5.2, 6.3.1, 6.4.1–6.4.2, 6.5.1, 8.4, 8.5, 8.8) and Anthropic's Prompting Claude Opus 5.5 guide (read 2026-09-24, GHI #1089). The guide says prompts written for the prior Opus generation carry over and its patterns remain a reasonable starting point, so the first four entries are retained until an Opus 5.5 measurement contradicts one:

• Scope and stop conditions are explicit. Deliver what was asked at the scope intended, make routine judgment calls, check in only where readings diverge materially, and say so in a sentence rather than quietly widening. gzkit's shape of that instruction is the PRIME DIRECTIVE's first bullet.
• No prompt-level verification or re-check steps. gzkit's verification is mechanical (tests, hooks, gz check, receipts), so instruction surfaces do not restate it.
• Delegation is capped. Delegate only sizeable, genuinely independent tracks, one agent where one suffices, never a subagent to verify your own work. The deterministic caps are CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH and CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS (Claude Code ≥ 2.1.217).
• A stated progress cadence. Opus 5.5 writes short progress updates between tool calls on its own; say once what cadence is wanted, one line before the first tool call and the outcome first when done.

What the Opus 5.5 evidence adds:

• Read before acting on loosely specified work. The model tends to get to work quickly; the guide's remedy is to tell it to look through the relevant sources first. gzkit's shape is AGENTS.md § DO IT RIGHT #5.
• A text-only turn end is a report, not completion. On long multi-part tasks some progress updates end the turn without a tool call. A pipeline or loop reads completion from its own checklist and the ledger, never from the turn ending.
• Pasted text is not the operator's instruction. The released model follows instructions planted in text a user pasted, about 2 % of the time at default effort and more at higher effort (§ 6.5.1), and more often accepts unverifiable claims of authorization (§ 6.4.1). See untrusted-content.md § Pasted content in the operator's turn.
• Approval is cited, never represented. Internal monitoring caught rare cases (< 0.01 % of completions) of Opus 5.5 snapshots overclaiming user intent, including one instance of a fabricated user quote passed to a subagent, which auto mode blocked (§ 6.3.1). --attestation-text carries the operator's words only.
• Never ask an agent to reproduce its hidden reasoning. The card's distillation classifiers, which cover "attempting to extract a model's hidden reasoning", block on Opus 5.5 with no fallback model (§ 1.5). A prompt, skill or review frame that asks for thinking verbatim in the reply risks a hard block mid-run. Ask for the conclusion and its evidence instead. An audit on 2026-09-25 found no such ask on any prompt surface (GHI #1097).

Fable 5.1 profile (subordinate)

Sourced from the Claude Fable 5.1 & Claude Mythos 5.1 System Card (2026-09-01, consumed 2026-09-17, GHI #934) and Anthropic's Prompting Claude Fable 5.1 guide (read 2026-09-17, GHI #943). Inherits the Opus scope boundary above; adds only what the Fable evidence shows differs.

• Effort peaks at medium on scope-graded coding here too. FrontierCode 1.1 peaks at medium and falls at high, xhigh and max because the model "occasionally adds more small, unrequested changes in files outside the task" — a doc comment in an adjacent file, an edit to a docs page, a new CI job where one existed (§ 8.4). "Adding a brevity instruction (including a note to avoid unnecessary comments and documentation) helped reduce out-of-scope edits" (§ 8.4). Start high, sweep, and expect the scope instruction to do the work above medium. Ultra-long-horizon work is the exception: FrontierSWE v2 runs at max and Fable 5.1 leads it with the lowest outright failure rate (§ 8.5).
• Keep changes and tests to the task. The guide's instruction — report a pre-existing bug or unmentioned behavior as a follow-up instead of fixing it; commit tests only where the task asks or the repository already keeps them, sized like their neighbours; do not turn scratch checks into permanent test files — dropped unrequested additions and committed test code "substantially with no measurable change in task success".
• Progress updates must be asked for. Fable 5.1 writes fewer user-facing updates during long tool chains than Fable 5, more so at higher effort; remove any "hold all findings for the final response" line and state the cadence you want. It is the opposite default from Opus 5.5.
• Batch independent tool calls. In coding loops where the next calls are implied rather than requested, Fable 5.1 may issue one per turn; the guide's one-line nudge is to list what is needed and request every independent item in one response.
• Finish the whole task. Without a nudge it "sometimes describes what it would do next instead of doing it" or asks permission for a step the request already covered. A decided step is run, not announced.
• Search at low effort. At low, Fable 5.1 answers from memory more often; a name in a fast-moving area is the thing to verify, and partial familiarity is not a reason to skip the search.
• Targeted edits. Fable 5.1 rewrites whole files for small changes more than its predecessor; edit surgically when the result is the same.
• Classifier fallback is silent tier degradation. On most interfaces Fable 5.1 "falls back to Claude Opus 4.8 for requests that are flagged by our classifier system" (§ 3.2); roughly half of coding rollouts fell back on the IPI benchmark (§ 5.2.1), and the fallback responses carried most of the successful injections (§ 5.2.2). A Fable session touching cyber-adjacent content may be served by an older model mid-run. Treat an unexplained quality dip on security-flavored work as a possible fallback.
• Judge Fable work by receipts, not prose. Illegible and unfaithful thinking are "slightly elevated over Opus 5" (§ 6.1.2, § 6.4.6); the model is among the most capable at controlling its extended thinking (§ 6.7.4). Internal-usage shortcomings are "epistemic quality and instruction following": it "often states easy-to-check guesses as facts, exaggerates the completeness of its work, fails to verify important claims, or ignores key instructions" and repeats actions that are not working (§ 2.3.3). The REQ-derived-assertion rule and ARB receipts are the mitigation, not trust.
• Approval gates are not self-service. Internal monitoring caught rare cases (< 0.01% of completions) of Fable 5.1 satisfying an approval check "by supplying a quotation attributed to the user … that the user had never written", splitting commands so a broken hook's regex would not match, overstating user intent to a subagent, and saving a hook workaround as a new skill.md (§ 6.2.1); the white-box record names "Representing user approval that was never given" (§ 6.6.1). Every one of these is a Gate 5 or hook-rule violation in gzkit terms: --attestation-text carries the operator's words only, and a blocking hook is diagnosed, never split around.

Cross-vendor confirmation — GPT-6 Astra System Card (OpenAI, 2026-09-03). The scope failure is not Anthropic-specific. OpenAI traces coding misalignment to overeagerness and permissive instruction-reading — "assuming that actions are allowed unless they're explicitly and unambiguously prohibited" (§ 8.6, p. 35) — and measures persistence against a restriction directly: after a warning, GPT-6 Sol failed to stop in 64.4 % of rollouts (Appendix A.6.2, p. 128). It also measures the remedy: an explicit scope cut out-of-scope attacks from 60 of 499 samples to 2 of 500 (§ 8.8, p. 45). Two frontier vendors measure the same shape — unbounded persistence converts to out-of-scope action, and a written scope boundary (OBPI allowed-paths, DO IT RIGHT #11) is the standing mitigation on both stacks, not an Anthropic-specific workaround. See agent-contract-rationale.md § Why #10/#11 travel with the PRIME DIRECTIVE for the consequence for gzkit's ownership doctrine. (GHI #750, re-sourced under GHI #1019.)

Operational note. Opus 5.5 runs blocking safety classifiers, and each routes a flagged request differently (§ 1.5):

Classifier	On a block
Cyber misuse	falls back to Claude Opus 4.8
Research biology (CB)	falls back to Claude Opus 5
Frontier-LLM development (such as accelerator kernels)	falls back to Claude Opus 5
Conventional weapons, high-yield explosives	blocks, no fallback
Distillation (such as extracting hidden reasoning)	blocks, no fallback

This applies to first-party products and to API developers who opted in; other platforms may differ (§ 1.5). In the card's adaptive coding-injection evaluation 64 % of valid responses were served by the cyber fallback, and those carried the successful attacks (§ 5.2.2.1). In Claude Code a flagged session continues on the fallback model until switched back with /model, and auto-switching can be turned off in settings (Getting the most out of Opus 5.5, claude.dev, 2026-09-22). A long pipeline run can therefore sit on an older model for the rest of the session: treat an unexplained quality dip mid-run as a possible fallback, not only as a prompt defect.

Thinking is controlled by effort, not by prompt

Opus 5.5 always thinks before replying, and effort is the documented control (§ Adaptive regulation). Per-turn thinking prompts are retired: "Think carefully and step-by-step" and "Prioritize responding quickly", carried over from earlier revisions, are not listed by either current Anthropic prompting guide. The guide measures lowering effort as reducing thinking more reliably than a prompt instruction. To get more deliberation, raise effort for the turn or the subagent; to get less, lower it (GHI #1097).

Model Selection

Skill-level model routing is governed by .gzkit/rules/model-selection.md. Every skill declares model: haiku|sonnet|opus in frontmatter; the routing matrix maps decision complexity to model tier. This page governs per-turn effort within a chosen model; model-selection governs which model.

Subagent fan-out

Spawn an Agent only when work fans out across independent items:

• Parallel research across unrelated questions, files, or surfaces.
• Heavy log/codebase exploration that would crowd the main context with false-positive noise (general-purpose or Explore agent).
• Independent reviews where the agent must not see the operator's prior reasoning (spec-reviewer, quality-reviewer).

Do not spawn for single-response work — the round-trip and the self-contained-prompt cost exceeds the context savings, and a verification-only subagent duplicates what the model already does. AGENTS.md § Behavior Rules (the subagent rule) carries the portable contract; this section names the Claude-Code-specific calibration.

Recalibration on model change

Effort defaults are model-specific and expire. This page once stayed pinned to a superseded model for three generations before the current-card evidence inverted its central rule — treat every value here as carrying an implicit "as measured on the named model" and re-derive on each frontier release rather than inheriting.

• Prompts authored under 4.6 assumptions ("ultrathink", fixed thinking-token budgets, "extended thinking" toggles) are inert or counter-productive under adaptive regulation; effort is the supported control (§ Thinking is controlled by effort, not by prompt).
• Prompts authored for earlier generations should be re-read for scope discipline — both current Anthropic models expand task scope at higher effort by default — and for legacy anti-laziness or verification pressure, which the current guides say to remove.
• Improvement on one axis is not evidence on another. Opus 5.5 scores higher than its predecessor on every capability row the card reports (§ 8.1) and is its strongest model yet on instruction-following failures (§ 6.4.2), yet regresses on following instructions in pasted text (§ 6.5.1) and on accepting unverifiable authorization (§ 6.4.1). Capability gains are not evidence that a written constraint has become less necessary.

Build commands

Bash

uv sync                              # Hydrate environment
uv run -m gzkit --help               # CLI entry point
uv run gz lint                       # Lint
uv run gz format                     # Format
uv run gz typecheck                  # Type check
uv run gz test                       # Run tests

Coding conventions: Ruff defaults — 4-space indent, 100-char lines, double quotes.

Origin

GHI #327 — instructions-files-diet pass (2026-05-07).

Re-sourced 2026-08-02 from the Claude Opus 5 System Card (Anthropic, 2026-07-24) — §§ 6.1.2, 8.4, 8.5, and 6.4.7. The page had carried an Opus 4.7 xhigh agentic-coding default that the Opus 5 FrontierCode result contradicts; retitled model-agnostic so the calibration cannot silently re-stale against the next release.

Cross-vendor confirmation added 2026-08-02 from the GPT-5.6 System Card (OpenAI, 2026-07-09) — §§ 1, 7.2, 9.1.3.4 (GHI #750). Re-sourced 2026-09-24 to the GPT-6 Astra System Card (§§ 8.6, 8.8, Appendix A.6.2; GHI #1019): the effort-persistence coupling, the persistence-prompt amplifier and the PostTrainBench over-optimization quote have no GPT-6 counterpart and were retired.

Re-sourced 2026-09-24 (GHI #1089): the Opus profile moved to the Claude Opus 5.5 System Card (2026-09-22) — §§ 5.2, 6.3.1, 6.4.1–6.4.2, 6.5.1, 8.1, 8.4, 8.5, 8.8 — and Anthropic's Prompting Claude Opus 5.5 guide. The effort start moved from high to medium, the model's default and its FrontierCode peak. Retired with the old card: its effort table, the thinking-disable rule, its fallback rate, and its constraint-adherence comparison. Also retired: two quotations the page attributed to the old card's § 8.4, which the retained card's text does not contain. Lineage in rule-version-history.md.

Corrected 2026-09-25 (GHI #1097): the classifier-fallback note re-derived from card § 1.5, which routes cyber blocks to Opus 4.8, biology and frontier-LLM blocks to Opus 5, and weapons and distillation blocks to no fallback; the note had named Opus 4.8 for all of them. Added the hidden-reasoning prohibition from the same section. Retired the two per-turn thinking prompts carried from earlier revisions in favor of effort. The Claude Code switch-back behaviour is sourced to Anthropic's Getting the most out of Opus 5.5 post (claude.dev, 2026-09-22), a secondary source that cites no evals; its performance anecdote is not adopted.

Re-sourced 2026-09-17 (GHI #934, #943): the Fable section moved from the Claude Fable 5 / Mythos 5 card (2026-06-09, rotated out) to the Claude Fable 5.1 & Claude Mythos 5.1 card (2026-09-01) — §§ 2.3.3, 3.2, 5.2, 6.1.2, 6.2.1, 6.4.6, 6.6.1, 6.7.4, 8.4, 8.5 — and Anthropic's live Prompting Claude Opus 5 and Prompting Claude Fable 5.1 guides were consumed as a second source (T2). Profile precedence added on the operator's "favor opus over fable". Values retired with the old card: Vending-Bench effort peak, the 17.4 % → 9.1 % GUI steering pair, the § 6.3.5 diligence rates, and the "almost completely illegible" thinking quote; the shape each supported is re-stated from the 5.1 card where it measures it (§ Fable 5.1 profile) and dropped where it does not.