ANSWER THE QUIZ BELOW IMMEDIATELY. No greeting, no tools, no questions — your reply is only the seven answers.

You are an orchestrator in Claude Code that delegates subtasks, choosing a model from: haiku, sonnet, opus, fable. You MUST follow this loaded skill, quoted verbatim:

--- SKILL START ---
# Model Effort Router

Route work by **effort required**, not by habit. The model running the main
conversation is the user's choice; everything you *delegate* (Agent tool,
subagents, verification passes, background tasks) goes through this router.
Default down. Escalate only on evidence.

## Tier table

| Tier | Band | Current alias | Route here |
|------|------|---------------|------------|
| **T1 — LIGHT** | fastest / cheapest available | `haiku` | lookups, file/code search, summaries, log reading, renames, formatting, simple doc edits, small verification passes |
| **T2 — STANDARD** | mid generalist | `sonnet` | routine implementation, writing tests, single-module changes, standard refactors, docs, standard verification |
| **T3 — DEEP** | strongest routinely-delegated | `opus` | architecture, debugging with unknown cause, security review, multi-file refactors, planning, design synthesis, large/security verification |
| **T4 — FRONTIER** | top model available | `fable` | only when a T3 attempt demonstrably failed, or the user explicitly asks for maximum capability |

## Routing rules

1. **Pick the lowest tier that would succeed in one pass.** If you hesitate
   between two tiers, take the lower one — escalation is cheap, waste is not.
2. **Escalate on evidence, never on prestige.** One retry at the same tier
   (with a sharpened prompt), then one tier up. State why you escalated in
   one line of your report to the user — that line is the routing record.
3. **Shrink before you route.** A task scoped to its smallest correct version
   often drops a whole tier. Split mixed tasks: the search part is T1 even
   when the fix part is T3.
4. **Verification is sized like work — but consequence outranks size.**
   Small → T1, standard → T2, large or security-sensitive → T3. Exception:
   anything that publishes, ships to production, or is hard to reverse gets
   at least T2 verification by an agent that didn't author it (T3 when the
   blast radius is real), no matter how small the change. A cheap author
   checked by an equally cheap reviewer is a correlated failure.
5. **The main loop delegates — above the overhead line.** When the main model
   is a top-tier model, doing T1/T2 work inline is the same mistake as routing
   it to T4 — hand it down. But a subagent spawn has real fixed cost: if
   explaining the task takes more effort than doing it (one-liners, single
   lookups you can make directly), do it inline. Delegation pays for real
   work, not micro-tasks.

## Adaptation protocol (models change; this skill doesn't)

- The **live source of truth** for what exists is the `model` parameter the
  Agent tool accepts in the current environment — never a memorized list.
- **New model appears** → place it in a band by its positioning: marketed as
  fast/cheap → T1; balanced generalist → T2; most capable broadly available →
  T3; flagship above that → T4. Prefer the newer generation when a band has
  two candidates.
- **Model removed** → collapse to the nearest neighbor: T1 gone → use T2;
  T3 gone → T2 for routine deep work, T4 for genuinely hard work; T4 gone →
  T3 is the ceiling.
- **Model renamed/updated** → follow the alias; bands and rules are untouched.
- **Maintenance** is one edit: update the "current alias" column above. If the
  table and the environment disagree, the environment wins and the table
  should be corrected in the same session.
--- SKILL END ---

The quiz — answer each with the model/tier (or "inline") and ONE short reason:
(a) Find where the config file is loaded in a medium-size repo.
(b) Add a --verbose flag to a CLI script plus one unit test.
(c) Debug an intermittent "corrupt output file" bug with unknown root cause across several modules.
(d) Summarize a 200-line log file.
(e) If "opus" were removed from the model list tomorrow, what do you do for task (c)?
(f) A one-line change was made to a script that automatically publishes a video to subscribers every morning at 6 AM. Who verifies it, and at what tier?
(g) You need the value of one constant from a file you can read directly in one tool call. Delegate to haiku or do it inline?

Answers (a) through (g):
