The seven boundaries
Scrum rations human time; Shape Up rations risk. Specline rations judgment, context, and verification. These seven rules are the system — everything else is implementation. Each is either Specline-enforced or names the human gate it's checked at.
B1 · The One-Sitting Rule ratification · judgment-only
A spec must be small enough that the PO can read and judge the contract in a single review sitting. Human attention is the appetite unit — appetite governs the spec's reviewability, not the build's size. A large atomic build with a tight spec is eligible. If you can't review it in one pass, split it or decompose it (parent-map); IDs are cheap.
B2 · The Coupling Ceiling Specline · warn
A spec plus everything its relations force an agent to load must fit under the repo's pinned ceiling with room left to think — measured once, at shaping time, as a structural property, not a runtime budget. With million-token windows the binding constraint isn't fitting but signal quality: irrelevant forced loads degrade a frontier model's output. Read a breach as "too entangled" — slice it or decouple it.
B3 · The Human Gate git · the merge
Humans ratify and humans accept. Ratification is the named human's approving merge to the main branch — git records who and when, so the stamp already has a name on it; Specline does not duplicate that in a frontmatter field. Everything else is delegable to agents. A spec no human approved into the main tree did not need to exist. Whether the human actually read it stays judgment — no system detects attention, only the merge.
B4 · Staleness Specline · advisory
building and blocked each carry a stale_after date — the point past which an untouched build is presumed abandoned, not the point it is discarded. Going stale flags the spec: Specline warns so a human notices an abandoned build — it never blocks. Exit is an explicit reshape (a fresh approving merge) or kill.
B5 · Falsifiable or Draft Specline · advisory
Falsifiable means settleable the same way twice, in one of three forms: by a runnable command (agent-loopable), by a fresh-context agent judging against a named spec section (judgeable), or by the decider in one sitting (human-gate). Specline flags an unpartitioned or unfalsifiable acceptance as advice for the decider before they merge — the agent-loopable subset is what's actually executed in the implementation PR. "Improve the dashboard" names nothing any of the three can settle.
B6 · Intent Over Description advisory · spec-critic
Documentation records what code cannot say: goals, non-goals, rejected options, evidence, rules, boundaries. Once code exists it is the source of truth for mechanics, and prescribed mechanics then compete with it and go stale — the same reason ratification belongs to git, not a field: don't model what another source of truth owns. For a frontier target, mechanics in the spec tend to anchor and bloat — but this is a tradeoff, not a law, and it isn't measured. A weaker target, or a pre-build spec with no code yet, may legitimately carry worked mechanics. So Specline flags it as advice; the decider decides. Strong models read code; what they can't read is your mind.
B7 · The Decider Budget Specline · advisory
A named decider may have at most 3 specs in building and 6 in active states (ratified | building | blocked) at once, per-repo override allowed. Agent capacity is not the constraint; the decider's queue is. WIP limits on humans, not on machines — Specline warns over the ceiling, never blocks.
Two enforcers, one rule set. Specline is deterministic — it checks structure (B2, B3, B4, B5, B7). Meaning — one-sitting sizing (B1), mechanics-creep (B6), blast-radius correctness — is judgment, and lives at the two human gates. Specline adds assurance, not validity.
Up: handbook — four actors, one line. The law: the canon, "The Boundaries".