SCORECARD

What we measure and how. Mission Control, audits, and anyone grading progress read this file first and grade against it. The scores themselves live in exactly one place: the table in fable_roadmap.md. Bucket files carry the WHY and the ideas, never a score. Scores move only at audits, with named evidence.

The overall score

The plain average of the eleven scored categories, one decimal. No weights. If a category matters more right now, that shows up in what we choose to work on, not in the arithmetic.

The categories

Each category answers one question a person would ask about the kit. For each: the question, where the evidence comes from, and what 0, 5, and 10 look like. An audit that moves a score names which anchor the evidence supports.

Trust

When the kit tells me something, is it true? Evidence: outcomes traced to run artifacts, false alarms counted, silent failures counted. 0 — reports are guesses. 5 — honest most of the time, but at least one class of silent failure remains. 10 — a full audit period with no false alarm and no silent failure.

Autonomy

How much work lands without a person touching it? Evidence: the hands-free counter, hand-fixes counted per audit period. 0 — everything needs a person. 5 — routine tasks land hands-free, failures still need a person. 10 — whole specs close with the human only reviewing.

Routing and cost

Is each task done by the cheapest agent that can do it, chosen from data? Evidence: cost per merged task, and whether assignments come from measured data or a list someone typed. 0 — gut calls. 5 — costs recorded, assignments still opinion. 10 — data assigns, overrides are rare and recorded as overrides.

Memory

Does the kit learn, or do we pay for the same lesson twice? Evidence: repeat failures answered from history; lessons appearing in briefs and measurably helping. 0 — every failure diagnosed fresh. 5 — history exists and answers repeats, effect on results unmeasured. 10 — measured: tasks that carry memory fail less and cost less than tasks that don't.

Getting oriented

How fast does an agent find what it needs in a repo? Evidence: share of tokens spent searching (the profiler), doc-pointer read rate. 0 — agents wander. 5 — pointers exist, effect unproven. 10 — the search share falls audit over audit and stays low.

Watching

Can a person see what the kit is doing right now, in one look? Evidence: a live page whose numbers have scripts behind them; stalls noticed by the machine before the human. 0 — logs only. 5 — a live page exists but a person still catches some things first. 10 — no state change a person learns about outside the product.

The human's seat

Is the human asked only when needed, and can they answer in seconds? Evidence: interrupts per spec, time from question to answer to resumed work. 0 — babysitting. 5 — questions reach a person, but answering takes digging for context. 10 — every ask arrives with its options inline and one reply resumes the work.

Audits

Does the kit check itself, on a schedule, with numbers a stranger can rerun? Evidence: audits that ran without being asked; metrics from scripts, not hand counts. 0 — no checks. 5 — good audits, but only when a person asks. 10 — scheduled audits, script-backed numbers, findings that become tasks.

Getting started

Can a stranger switch it on? Evidence: timed setup on a fresh machine, adoption runs on outside repos. 0 — only the builders. 5 — one documented path, proven once. 10 — strangers set it up routinely and time-to-first-merged-task is measured and short.

Runs anywhere

Does it run in more than one place? Evidence: instances alive, setup repeated on hardware that is not ours. 0 — one laptop. 5 — a second instance works. 10 — instances are routine, each knows its own identity, none is special.

Project powers

Does it ship real work that is not itself? Evidence: external projects with merged results a person actually uses. 0 — the kit only builds the kit. 5 — one external project ships something. 10 — several projects running, humans mostly reviewing.

Not scored

Moonshots is an idea funnel, not a capability. It keeps its bucket file and its ideas, but it carries no score and is excluded from the overall.

Rules

1. Scores live in the fable_roadmap.md table and nowhere else. A score found in any other file is stale by definition — treat the table as the truth, and replace the stray number with a link to the table.
2. Scores move only at audits, and every move names its evidence and the anchor it supports.
3. If reality and a score diverge between audits, note it in RESUME. The score itself waits for the audit.