WAVE_RUNBOOK.md

pipelines/WAVE_RUNBOOK.md

LENS WAVE SHARED RULES (P1-B + P1-C) — read this first, then your lens block in WAVE_SPECS.md

Repo: C:\dev\humanity-forgotten (canon). Game repo: C:\dev\Humanity\Humanity (read-only reference

for built-system claims). You are READ-ONLY on both repos: never edit, write, or stage anything

inside them. All your output goes to the scratchpad directory you are given.

What a lens audit is

You audit HUMANITY's current design state against what your title-set teaches. Every finding names

a lesson the title teaches, states our current state WITH CITATIONS (file + section/line), gives a

verdict class, and a reuse-first recommendation. You are a design auditor, not a fan: findings must

be decision-grade.

The standing wave rule (binding, from COMPARATOR_LENS_PROGRAM.md)

• CITED CURRENT-STATE per finding: name the exact doc/registry/code file and section that proves

our current state. A claim with no citation is cut.

• POSITIVE-CONTROLLED ZEROS: before reporting "we have nothing for X", prove your search could

have matched — run it against a term you KNOW exists (same tool, same corpus). Chapter refs are

TRI-DIALECT ("Ch 72" / "Ch-72" / "_ch72_") — check all three, in Python not shell grep where

brackets/escapes are involved. A search that CANNOT match reports zero.

• DESIGNED-UNBUILT is NEVER reported as MISSING. If a spec/ruling/registry row exists but nothing

is built, the verdict is DESIGNED-UNBUILT. Check docs/PRE_5090_BUILD_PLAN.md (+ VOL2),

docs/COMBAT_PROGRAM_ADDENDUM.md, the T1 docs, and the registries before minting MISSING.

• REUSE-FIRST recommendations: name the existing system/schema/gate the fix lands on. A brand-new

system is a last resort and must say why nothing existing serves.

• PRE-REGISTERED DIVERGENCES are not gaps: check docs/DESIGN_GAP_REGISTER.md's DIVERGENT-BY-DESIGN

entries and your cluster's divergence guards before re-litigating (e.g. no-scoreboard governs

progression stats only — CVD §17.12 ratifies minigame top-score completion; markerless/diegetic

navigation is doctrine, not a gap).

Verdict classes (every finding carries one + owner + wave + priority)

COVERED / DESIGNED-UNBUILT / WEAKER / MISSING / DIVERGENT-BY-DESIGN.

Owner = which build lane fixes it (W-SPACE, P5-factory, P1/P2/P3, QA loop, spec-doc, schema-data).

Priority = player-impact × solo+AI feasibility, honest.

Canon context you must respect

• The vision is LOCKED (CVD + 12 pillars + 79-node spine + the numerical canon). Your findings

EXPAND/ENHANCE within it; never propose changing chapter counts, names, §17 lines, or pillars.

• The slice is PROLOGUE→CH 13 (built Ch 2-13 first, then Prologue + Ch 1). Slice-relevant findings

outrank post-slice findings at equal weight.

• Rated-R 17+ AAA bar; care = weight/authenticity, never omission.
• The JOSH GATE: the AI QA loop is the only player until the polished slice; findings that

strengthen the QA loop's ability to verify a surface are first-class.

• Key doctrine homes: CVD (_source/01_Tier_1_Foundation/T1_CVD_Creative_Vision_Document

[ACTIVE v1.4].md), T1_Combat_System_Spec §§12-15, T1_Ability_Tree §5H/§5I/§14A,

docs/COMBAT_PROGRAM_ADDENDUM.md, docs/ENGINE_OPTIMIZATION_DOCTRINE.md,

docs/proposals/QA_WATCHING_PROGRAM.md, docs/proposals/REALM_DESIGN_PROGRAM.md,

docs/PRE_5090_BUILD_PLAN.md (the ranked 30 + ledger), registries/**/*.csv (live canon).

• Format exemplar: docs/pipeline_review/LENS_BG3_2026-07-27.md.

Output contract (the StructuredOutput-cap lesson — hard rule)

Write your FULL audit to the scratchpad file named in your prompt (rich detail welcome there:

15-30 findings typical, each with citations). Your RETURN is compact: lens_id, file path, counts

per verdict class, your top 5 findings one line each, and any contested/uncertain calls flagged

for the director. Keep the return under 30 lines. Do not inline the full audit in the return.

Critic pass (fresh-context refutation — for critic agents only)

You receive a draft audit file. Your job is REFUTATION: for each finding, verify the citation is

real (open the file), verify the verdict class (especially MISSING vs DESIGNED-UNBUILT — search

the repo yourself; the author's zero may be a search that could not match), verify the

recommendation is reuse-first, and check the finding is not a pre-registered divergence or an

already-covered register entry (docs/DESIGN_GAP_REGISTER.md). Kill-list every finding that fails

with the decisive evidence quoted. Also run the DEVIL'S-ADVOCATE-AGAINST-TIMIDITY pass: flag any

finding whose recommendation is care-inflated or defanged relative to the R-17+ bar. Write your

full verdict to your named scratchpad file; return compact: kill count, survive count, amended

verdicts, contested items for the director.

Generated by harness/site/structure_site.py — the URL path is the repo path. review root