A3 — Cross-Reference Audit (MIZOKICloudRun completion run)

Auditor: Subagent A3 (read-only). Date: 2026-08-08 (UTC). Repo state: /home/user/MIZOKICloudRun, branch claude/mizoki-completion-run-4ms74o @ 6ef0275 (== origin/main). Live target: https://mizoki3.com (fetched 2026-08-08 ~18:33–18:35 UTC via curl through the agent proxy; server Google Frontend, last-modified: Fri, 07 Aug 2026 23:13:48 GMT on /signal/measurement — consistent with the 2026-08-07 content-wave deploy). Method note: every route audited is served by send_from_directory in # MIZ OKI 3.5/app.py (lines 434–494, 887–894) — raw static file serves, no templating — so raw-byte hash comparison is valid and was used as the primary method; headline/§-string checks were used as secondary confirmation. A methodological hazard hit twice during the banned-string sweep: grep -E ".{60}term.{60}" (fixed context padding) silently misses matches near line boundaries. All banned-term results below come from the corrected plain grep -in sweep.

1. drop/signal-story-bank-v1.1.md vs # MIZ OKI 3.5/docs/marketing/signal-story-bank.md

Verdict: IDENTICAL (byte-for-byte).

• diff -u output: empty, exit 0. Zero hunks of any classification.
• md5 both files: 17ca73b7db610171158b205e72745279 (9,481 bytes each).
• Neither file is newer/superset of the other — both were written in the same commit 05bacb7 (2026-08-07, "marketing: signal net-yield content wave — story bank v1.1, positioning doc, listing copy draft, briefing act proposal, roadmap card, claim-ledger QA rules"). The drop copy was created in that commit; the docs copy pre-existed as v1.0 (at 8c49062) and was upgraded to v1.1 in the same commit.

v1.1 version stamp — VERIFIED in both copies

Line 1 (both files): # Signal Intelligence — Customer Story Bank v1.1 Line 2 (both files): *…All numbers illustrative until replaced by verified pilots. v1.1 adds Stories 7–8 (net yield, cold start) per the Shopify/Net Contribution positioning.*

The five governing rules — present in both, ZERO wording drift (byte-identical)

Quoted from ## The rules (non-negotiable) (drop copy lines 5–9 == docs copy lines 5–9):

1. **Label honestly.** — "Every story is a 'composite scenario, illustrative numbers' until a real customer signs off on a named case study. Never invent company names, logos, or quotes attributed to people."
2. **One grammar, every story:** — "their question → the quiet test → what the ledger showed → what changed."
3. **Translate, don't define.** — includes the v1.1 addition "Net contribution = 'what an order actually nets after everything it costs you.'"
4. **Preview capabilities stay preview.** — v1.1 wording: "Intent prediction (Story 6), net-yield bidding (Story 7), and cold-start seeding (Story 8) are told with in-development framing until their cells ship. Research-backed approaches cite the research; published results from other companies (e.g., Airbnb's synthetic-data metrics) are NEVER presented as ours."
5. **Never a guaranteed outcome.** — "Stories end in decisions and receipts, not promised percentages for the listener."

v1.0 → v1.1 lineage (context, not drift between the two audited copies)

git diff 8c49062 HEAD -- '# MIZ OKI 3.5/docs/marketing/signal-story-bank.md' = 23 insertions / 7 deletions: - Version stamp v1.0 → v1.1 (title + intro sentence). - Rule 3: added the "Net contribution =" translation (substantive rule extension). - Rule 4: rewritten from "Intent prediction stories are told in future tense … until Cells 28/33/34/35 are live" to the broader Story-6/7/8 + never-present-others'-results wording (substantive rule change — this is where the Airbnb-metrics ban entered the rules). - Added Story 7 (net yield — PREVIEW) and Story 8 (cold start — PREVIEW) in full (story additions). - Placement map: +2 rows (signal.html roadmap/preview section; Shopify app listing), briefing act row reworked to PROVE→PROFIT→ANTICIPATE (substantive placement change). - Upgrade path: "six composites" → "eight composites" (consistency edit).

2. drop/mizoki-shopify-net-yield-positioning.md vs # MIZ OKI 3.5/docs/marketing/mizoki-shopify-net-yield-positioning.md

Verdict: IDENTICAL (byte-for-byte).

• diff -u output: empty, exit 0. Zero hunks.
• md5 both files: ee77b02af6059e38aaae9216cc3a7f83 (8,044 bytes each).
• Both written by the same commit 05bacb7 (2026-08-07). Title carries v1.0 of the positioning doc in both.

Claim ledger — ZERO divergence (rows identical by byte identity; verified by reading both)

The §2 table "Claim ledger (enforce in ALL copy — extends content_qa rules)" has 9 rows, identical in both copies (drop lines 16–26):

Banned-clause line (line 28) identical in both: presenting any row above its status; the Airbnb numbers as ours; "autonomous profit machine" without the autonomy-ladder caveat.

3. Repo vs LIVE (https://mizoki3.com) — 9 pages

All nine fetches returned HTTP 200 at the primary clean route (no fallback paths needed; app.py also exposes .html aliases). Primary comparison: md5 of raw fetched bytes vs the repo file. All nine match exactly.

Secondary confirmation — title/h1 served live (identity of content, not just hash): - / — "MIZOKI3 — A Nervous System for Your Business" / h1 "A nervous system for your business." - /signal — h1 "Prove which spend caused the customer." - /signal/thresholds — h1 "Below the threshold, spend buys noise." - /signal/budget — h1 "Noise cannot move money." - /signal/creative — h1 "Creatives decay. Measure it before it costs you." - /signal/audiences — h1 "Target who changes their mind — not who was coming anyway." - /signal/measurement — h1 "One clock. One identity. One ledger." - /shopify — h1 "Other tools optimize the number your ad platform reports. / Mizoki optimizes the number your bank account reports." (§01–§06 all present in stripped text) - /blog/doorman-problem — h1 "The Doorman Problem: what your retargeting dashboard can't tell you"

There is no repo copy newer than live and no live copy newer than repo for these nine surfaces: the deployed revision serves exactly the bytes at 6ef0275.

3a. LIVE /signal/measurement — machinery-claim inventory (verbatim live sentences)

All nine listed machinery claims appear on the live page. Per-claim labeling status on the live rendering:

Page-level framing on live (the only umbrella over claims 1–2 and 4–9), §03 note, verbatim:

"Stated plainly: mechanisms on this page are operating machinery, and the windows and thresholds shown are operating defaults — not promised outcomes. Ledger figures shown anywhere on this site are illustrative until they are your own — produced by your experiments, on your data, with the confidence intervals printed."

Findings/flags on the measurement page: - The word "Preview" appears zero times on live /signal/measurement (grep count 0). No machinery row carries a preview or design-target tag. - The blanket note affirmatively asserts, present tense, that "mechanisms on this page are operating machinery." Under the truth discipline ("present-tense deployment claims for unbuilt capabilities" are banned), this sentence is a deployment-status claim for all nine mechanisms, including the five platform writeback/ingest rails (Enhanced Conversions, Meta Conversions API, AEM, GA4 MP, offline conversion upload) and the Meta/Google weight back-propagation. Whether those rails are in fact built/live is outside what this audit measured (no code-to-claim verification was in scope for A3); flagging for cross-check against build state (Cell 35 / connector reality). If any rail is unbuilt, the violating sentence is the §03 blanket note plus the unlabeled §02 rows, exactly as quoted above. - Contrast: the sibling claim ledger (audited doc §2) requires "Net-contribution bidding via CAPI / Value Rules → 'In development · preview'". The measurement page's rails table describes conversion measurement plumbing, not net-contribution bidding writeback, so it is not automatically the same ledger row — but the same rails are the delivery mechanism named in that row, which strengthens the case for the cross-check.

3b. Banned-string sweep on all 9 live pages (plain grep -in, then per-hit context)

3c. Preview-framing distribution on live (supplemental to the banned sweep)

• live /signal — carries the required framing: "…belongs to the anticipatory intent capability in §06 — Preview · in development — and it's marked that way for a reason" (line 355); net-yield card: "Stated plainly: net-yield pricing and writeback are in development, not deployed. The Preview label on this card flips only after a real pilot…" (line 433); §06 note "not mind-reading, and it is in development rather than deployed. Signal uses no microphone or audio signals of any kind — not now, not on the roadmap." (line 416). 6 "preview" occurrences total.
• live /shopify — net-yield story §02 labeled "composite scenario · illustrative numbers" in the § heading; ledger card tagged "composite · illustrative"; future-tense "When net-yield bidding ships, the bidding signal switches…"; "Stated plainly: net-yield pricing and bidding are Preview · in development — nothing here claims they are deployed, and the label flips only after a real pilot writes verified numbers. Returns of 25–40% in apparel are an industry statistic, not a customer result (composite scenario)."; roadmap §05 header tag "Prove → Profit → Anticipate Preview · in development" and closer "Stated plainly: both cards above are Preview · in development, not deployed." Doorman story §03 carries "(illustrative)" on both the 8x and 1.4x figures. Anticipatory-intent and net-yield claims on /shopify and /signal comply with the preview-framing rule as rendered live.
• live / (index) and /blog/doorman-problem — 4 and 3 illustrative/composite/development markers respectively; no anticipatory/net-yield present-tense claims found unlabeled.
• live /signal/measurement — the one surface with zero "Preview" strings; see 3a flag.

4. drop/ folder hygiene

git ls-files drop/ → all 6 files are git-tracked (nothing untracked, nothing ignored):

No stray/junk files; the folder is a clean mix of two governing marketing docs (mirrored into docs/marketing/), three owner prompt/runbook files, and one already-consumed HTML fragment.

Completion gate

• Read-only audit; no repo files created/edited/deleted. Only writes: this report + fetched copies under scratchpad (…/scratchpad/live/).
• python3 scripts/claude_memory.py check --strict → "memory system is structurally valid".
• States distinguished: everything above is measured (diff/hash/grep on repo @ 6ef0275 and live fetches at 2026-08-08 ~18:34 UTC). The one item NOT measured: whether the /signal/measurement rails are actually built — flagged, not asserted, either way.