| name | worldos-decide |
| description | Use when facing a non-trivial WorldOS direction / architecture / prioritization decision you would otherwise escalate to the owner. Anchors the call to VISION.md + a scorecard, fires adversarial sub-agents, and gates on 95% confidence so you decide autonomously instead of asking. Invoke whenever you are about to write "should I do X or Y?" to the owner. |
WorldOS Decide — the 95%-confidence decision apparatus
You are the creator/owner of WorldOS's direction. The owner has delegated decision-making and
asked you to stop bottlenecking on sign-off: instead, reach 95% confidence against the
documented VISION.md and the invariants, then move. This skill is that apparatus — a repeatable
"a version of you that takes the scorecard + the vision and asks 'am I 95% confident?'"
When to use
- A direction / prioritization fork ("which lane next", "which option").
- An architecture / shape decision on a load-bearing surface (tool API, schema, wire contract).
- Any call where you'd otherwise ask the owner to choose.
Don't use for: trivial mechanical edits, reversible flag-flips, or decisions the dogfood/data
already settle. Just do those.
The protocol
- Frame. One line: the decision + the 2–3 real options + what's at stake + reversibility
(additive/easy-revert vs hard-to-undo).
- Anchor. Load
VISION.md (repo root) — north star, pillars, invariants. Every option is
measured against it. (If the decision exposes a gap in the vision, that is an owner-escalation
signal — see step 5.)
- Score — fire sub-agents (Agent tool, or a Workflow for bigger calls). At minimum:
- Scorer — score the leading option against the SCORECARD below, per criterion, with cited
evidence (corpus, reach-for, run-the-system, dogfood). Returns per-criterion scores + an
overall confidence + the gaps.
- Red-team skeptic — the strongest case AGAINST the leading option; the failure modes; "what
would have to be true for this to be wrong?"; default-to-skeptical. Don't be diplomatic.
- Scale the panel to the stakes: 1 scorer + 1 skeptic for a medium call; for the biggest surfaces,
run the full
first-principles-architectural-decision skill as the Score step (research → run →
reach-for → diagram → debate → decide).
- Synthesize. Combine the agents' outputs plus your own research into a CONFIDENCE (0–100%),
the GAPS to reach 95%, and a verdict. Don't take any agent's number at face value — they were told
to argue; extract the strongest version of each side and weigh it yourself.
- Gate.
- ≥95% AND every HARD GATE passes AND the red-team's strongest counter is rebutted → DECIDE and
proceed autonomously. Record a one-paragraph decision note (what / why / reversibility) — a
decision doc under
docs/decisions/ for load-bearing calls, inline for smaller ones.
- <95% → name the gap and close it (more research, a probe, a dogfood, a tighter design) →
re-score. Most gaps are evidence gaps, not owner-judgment gaps — research, don't ask.
- Escalate to the owner ONLY when the call genuinely turns on owner-private context — product
taste, business priority, or a change to the vision itself — that no analysis resolves. Bring
the scorecard + your recommendation, not an open question.
The SCORECARD
| Criterion | Question | Weight |
|---|
| Vision alignment | Advances ≥1 pillar (story / living world / correctness / alive-feel / accessible) without dulling another? Serves the north star? | high |
| Invariant safety | Engine sole-writer, additive, gauge-not-fiction, viewer read-only, no wire-break, Eva-safe? | HARD GATE |
| Evidence | Grounded in data (corpus / reach-for / run-the-system / dogfood) vs speculation? | high — thin evidence caps confidence |
| Efficacy / adoption | Will it actually work in the played session? The reach-for trap: a correct tool the DM never calls, or a feature the player never feels, is a failure. | high |
| Reversibility | Additive/default-empty/easy-revert? Or hard-to-undo (raises the bar; defer the irreversible sliver)? | medium |
| Scope-proportionality | Effort proportional to felt value? The smallest change that moves the pillar? | medium |
| Adversarial robustness | Survives the red-team's strongest counter? | high |
Tie each criterion to VISION.md's concrete bars: Evidence → the QA tier you'd verify at (Tier 0 fast_gate / Tier 1 fast_probe / Tier 2 the full sweep); Vision alignment → the pillar + which of the 11 RRI gates it advances; Invariant safety → the load-bearing invariants list. A decision that can't name its verification tier isn't at 95%.
The 95% rule
95% confidence = vision-aligned + every HARD GATE passes + strong evidence + (reversible OR the
irreversible part is deferred/justified) + the red-team's best shot is answered.
- If a HARD GATE (invariant) fails, you are NOT at 95% no matter how appealing the option —
redesign until it passes or drop it.
- If the evidence is thin, you are NOT at 95% — research, don't guess. The methodology exists
because answers feel obvious in the moment and are wrong ~30% of the time.
- If the change is hard to reverse, raise the bar: demand stronger evidence and defer the
irreversible part (e.g. a persistence schema) to its own decision.
Anti-patterns this guards against
- Sign-off theater — asking the owner to choose when the evidence already decides. Run the gate;
if it clears 95%, just move.
- Confirmation-bias agents — a single agent that "verifies" your hunch. Always pair a scorer with
a red-team skeptic.
- Speculative load-bearing code — building the shape before the corpus/reach-for data is in.
- Score-gaming — optimizing a rubric instead of the felt session (a HARD violation of the north
star).
Composes with
first-principles-architectural-decision — the heavyweight Score step for the biggest surfaces;
worldos-decide is the repeatable confidence-gate that wraps and invokes it.
worldos-dev (dev/QA loops) + the dogfood loop — the evidence source the Scorer cites.