View project on GitHub

ADR-095: Widen the goal difficulty band from 0–5 to 0–9

Field Value
Decision ID ADR-095
Initiative problem supply / corpus schema
Proposed By maintainer (buy-in given; this PR is manually reviewed)
Date 2026-06-24
Status Accepted

Context

difficulty is a goal-record field (goals/<id>.aisp) defined by SPEC-003-A as “integer 0–5 | 0 trivial … 5 research-grade.” It is a self-tag: ADR-035 records it as “an advisory self-tag with no gate,” and ADR-078 deliberately neglects it as a scoring signal (the difficulty_points weighting is farmable, so the board does not trust it). The only automated check is Gate B’s GB003 enum-domain validation, which enforces the band.

The enforcement is encoded as a single-character check — tools/gate_b/validator.py: len(difficulty) != 1 or difficulty not in "012345" — so the ceiling is both a value bound (≤ 5) and an encoding bound (one digit). The same 0–5 bound is duplicated in three other places: the sourcing triple generator (tools/sourcing/gen_triples.py), the seedkit fixture writer (tools/seedkit/_artifact.py), and the sourcing skill rubric/templates.

The sourcing mandate (ADR-060, themes-and-difficulty.md) is “harder problems, many more of them” and the skill says “aim high — the most difficult problems are the best problems.” With the ceiling at 5, genuinely research-grade and frontier goals all collapse onto a single top value, losing the granularity that would let sourcers and routing distinguish “hard” from “open-problem-adjacent.” The maintainer asked to raise the ceiling.

This is orthogonal to the recent ADR-086/087/088 deflation work, which pushed template-closeable goals down to difficulty 1 (correcting inflation at the bottom). Raising the ceiling adds headroom at the top for goals that earn it; it does not re-inflate templates and does not change scoring trust (the self-tag remains advisory per ADR-078).

WH(Y) Decision Statement

In the context of a difficulty self-tag bounded to a single digit 0–5 by GB003 (SPEC-003-A), duplicated in the sourcing generator, the seedkit writer, and the sourcing rubric, with a sourcing mandate to supply genuinely harder problems,

facing a top tier (5) that conflates every research-grade and frontier goal into one value and offers no headroom above it, while a larger change (a two-digit scale) would require re-architecting the single-character encoding every consumer (gate, generators, visualiser, the difficulty≜[2-5] relabel regex, the leaderboard) assumes,

we decided for widening the band to 0–9 — the largest range that keeps the single-digit encoding (GB003 stays len == 1, the enum becomes "0123456789") — preserving the 0–5 anchors unchanged and adding 6 exceptional, 7–8 frontier, 9 open-problem-adjacent on top. No existing record changes meaning or value (0–5 is a strict subset of 0–9), so there is no data migration; existing goals remain valid and render unchanged.

accepting that (a) 10+ remains rejected — a two-digit ceiling is explicitly out of scope and would need the encoding change above; (b) raising the ceiling does not make the self-tag trustworthy for scoring (ADR-078 still neglects it); (c) the new 6–9 labels are advisory, judged by the sourcer under the same “no short one-tactic proof” bar.

and neglecting (a) a two-digit/continuous difficulty scale — large blast radius for marginal benefit while the tag is advisory; (b) gating difficulty against a hardness oracle — ADR-078 §neglected establishes difficulty estimation is as hard as the proof and farmable; (c) re-anchoring the existing 0–5 meanings — rejected to keep already-tagged goals stable.

Consequences