Launch-prompt effort estimates should NOT be the basis for defer-vs-do-now decisions when the work is in the well-trod “batch + Decision Matrix + shared-infra” pattern. The empirical compression factor is stable enough to count on.
Empirical evidence (4 batches as of 2026-05-19):
| Batch | Launch-prompt estimate | Actual execution | Compression |
|---|---|---|---|
| π consumption-automation triplet (tasks 50+45+46) | ~9-13h | ~2.5h | ~4-5× |
| σ annotation-linter MVP + fedlex (tasks 65+74) | ~4-7h | ~55min | ~5-8× |
| χ TIER-2 audit follow-ons (tasks 78+79+80) | ~1.5-3h | ~1.5h | ~1-2× (blocked-on-docker forced ledger row + partial) |
| ρ sprint-planning automation triplet (tasks 47+49+48) | ~7-10h | ~1h | ~7-10× |
(χ is the outlier — Docker WSL blocker forced a re-fire stub for task-80 rather than full execution; the 2 closed tasks 78+79 ran near-instant.)
Why: Rich has observed (2026-05-19T18:40 BST, after BATCH ρ closure) that the consistent 5-10× compression strengthens the case to “do this work immediately” rather than defer to Phase-1.5+. Background:
- Decision Matrix pre-decides forks — 15-25 forks pre-decided in the launch prompt removes “shall I proceed?” pauses + ambiguity-resolution round-trips. Each removed pause = 30-90s saved.
- Thinner-MVP scope corrections at orchestration-time — substrate- corrections registered BEFORE execution (verify-before-author per §13) shrink scope before it inflates.
- Shared infrastructure reuse compounds — BATCH ρ.2 imported BATCH π’s
cascade_q_parser.pyverbatim, saving ~30min of frontmatter-parsing re-implementation. Each reuse pays the prior batch’s investment back. - Banned permission-grab phrasing per global §7 — no more “do you want me to proceed?” pauses at phase boundaries.
- CLAUDE.md §13 verify-before-author — the executing session catches
substrate errors fast (e.g., ρ caught finding-33 collision with
validate-build-plan-sync.shat ~10sec of grep).
How to apply: When considering whether to add a task to Phase-1.5+ deferral OR execute it now:
- Discount launch-prompt effort estimates by 5-10× when the task fits the batch + Decision Matrix + shared-infra pattern (most cascade-Q + scripting + ledger work does). A “MEDIUM ~3-4h” task is often ~30-45min in practice.
- Lower the DO-NOW threshold per MQ-017 Axis-1 α. The cost of doing- now is empirically much lower than the deferral-bias was treating it.
- The compression is mechanism-load-bearing not luck-load-bearing. If the launch prompt has the 4 mandatory techniques (Decision Matrix, banned-phrasing, automatic phase transitions, redirect uncertainty to artifacts) PLUS shared-infra reuse opportunity, the compression empirically materialises.
- Conversely: if a task has NO Decision Matrix opportunity (genuinely novel scope, unclear forks, no shared-infra to reuse) — DON’T expect the compression. Estimates are for that work.
- Don’t blanket-discount estimates — verify the task shape matches the batch pattern first. A research task, a brainstorming session, or a novel ζ-Q lock decision still takes the time the launch prompt says.
Companions:
- MQ-017-defer-vs-do-now-decision-rule-plus-verification-bias-discipline-locked — the existing rule this empirical pattern refines
- feedback_autopilot_prompt_techniques — the 4 mandatory techniques + bonus techniques that produce the compression
- feedback_research_artefact_forward_traceability — verify-before-author at authoring time (the recursive mirror of MQ-015)
Tracked: 7th verify-before-author precedent (cumulative — TA 1925 s.27 + pgvector counts + arch-state chunks + ZGB §521 + 4 χ-orchestration + T81 —deterministic + ρ-orchestration finding-33). Pattern empirically stable.