7. Cross-file consistency audit — the synthesis is the weak point

Reviewer read all seven research files plus ../README.md, ../CRITIQUE.md,
../RESEARCH-distribution.md, ../RESEARCH-van-economics.md,
../RESEARCH-precedent.md and ../RESEARCH-routes.md. 35 number collisions
found. The verdict is blunt and worth quoting:

"The evidence base underneath this corpus is unusually good and the document
sitting on top of it is not safe to spend money against."

The five regional files are strong — dense with first-party quotes, properly
tagged, honest about gaps, and they repeatedly argue against the business case.
00-synthesis.md is where integrity breaks, and it is the only document the
README tells people to read.

7.1 CRITICAL

S-1. The synthesis deletes the caveat every source file leads with.
research/README.md: "search quota was exhausted early… operators only findable
by search may be missing." 03: "WebSearch was unavailable for the entire
session… negative findings cannot be treated as exhaustive." 02: "Absence of
evidence… is not proof."
00 contains zero methodology caveats — grep for "search quota", "WebSearch",
"exhaustive" returns nothing. It opens with "Roughly 90 operators were examined…
Not one runs a roving support van for self-guided customers", presented as a
completed sweep.

S-2. The synthesis states a private-equity ownership chain that no file supports
and that 06 explicitly records as NOT FOUND.
06: "[NOT FOUND] … Wilderness Group's ownership or backing… returns HTTP
403." 00 §10: "owned via Wilderness Travel Group Bidco/Midco/Topco and
controlled by Mobeus Equity Partners LLP — the founders' PSC entries ceased
7 September 2018."
"Mobeus", "Bidco", "Midco", "Topco" and "PSC" appear nowhere else in the
project. A named PE firm, a corporate structure and a dated filing event asserted
with no trace, in a corpus where everything else is tagged. Delete or source it —
this is the kind of error that discredits the other 400 KB.

S-3. The flagship number in §2 means the opposite of what it is used to prove.
"A dedicated van needs 6–8 riders… Groups of 2–5 sit below the line, which is
exactly why they get a phone number." That is a statement about one van's
justification ratio. The README wants one van serving six to eight groups =
18–24 riders. 02 caught this and wrote the warning:

"DuVine puts a van behind seven people. Cellar Door Cycling proposes putting one
behind three times that many. That is either the innovation or the over-reach."

That sentence appears nowhere in the synthesis. The corpus's strongest challenge
to the core operating assumption was dropped in the roll-up, and the number
repurposed as supporting evidence.

S-4. The headline negative has at least six counter-instances inside the corpus.
On-demand phone-dispatched rescue for unrelated self-guided hire customers
already exists at: Wine Tours by Bike ("on road back up"), Bike 2 Wine ("a
bicycle rescue service, for any reason"), Zoom Leisure, Bloomfield ("Call us
and we'll come and fix you on the spot… or bring you another bike"), Grapemobile AU
("complimentary roadside assistance in the vicinity of our designated trails"),
La Rioja Bike Tours. 05 itself calls Wine Tours by Bike "the closest existing
business to the concept anywhere" — at NZ$65.
The surviving true negative is three conjoined conditions: no operator runs a
continuously circulating vehicle with a published response window that also
delivers wine to the guest's own bed. Far weaker than "nobody does this".

S-5. The Australian competitors that most threaten the plan are missing from the
synthesis. The README claims the gap is "self-guided autonomy at a guided level
of support, priced between $159 and $195–275."

• Escapegoat, McLaren Vale: A$175pp self-guided wine e-bike tour — South
  Australia, e-bike, self-guided. Sits inside the claimed gap. Absent from the
  README's competitor table and the synthesis price table.
• TrailHopper Barossa: A$89.95–99.95, 14 stops, daily, "All guests travel on
  the same hop-on hop-off service". 05 wrote: "TrailHopper is the sharpest
  competitive threat, not the guided bike tours." The synthesis mentions it once
  — in a list of booking-software users.
• ../RESEARCH-precedent.md adds iHop Hunter Valley (hourly loop, 15 stops) and
  Hunter Valley Hop at $35/head.

Australian shared-loop vehicles already sell to unrelated customers at a third
to a half of the modelled price, and this is absent from the document the README
says to start with.

S-6. 00 cannot decide whether response-time SLAs exist. §3: "Two published
response commitments exist in the entire world — Discover France's 'back on the
road within an hour' and the ADFC's 120-minute window." §12: "[NOT FOUND]
anywhere in the world: a single published response-time commitment." §12's version
generates the recommendation ("genuine, uncontested white space"); §13.2
simultaneously calls 20 minutes "six times better" than ADFC and §12 calls it
"the riskiest sentence the business could print."

S-7. Wine Australia 29% overstated — independently confirms C-4 above.
00 says "An Australian cellar door earns… a 29% government top-up on every bottle
sold on site" and tells the founder to take it into five winery calls. Its own
paragraph states the A$1,207,000 sales floor, the A$100,000 cap and the
A$10m rationed pool. It is also 29% of notional wholesale, not cellar-door
retail.

7.2 The most damaging number collisions

Full table of 35 in the reviewer's report. The ones that change decisions:

7.3 Moderate

• M-a. "No operator builds on Airbnb-type inventory" is method-blind: a
  method that starts from a list of tour operators structurally cannot find a
  business serving Airbnb guests. Only the three French refusals (verbatim, own
  sites) are genuine positive evidence of an operating rule.
• M-b. "No day pass / subscription covering shuttle and rescue" is a
  conjunction of two abundant things — stated directly above a table listing ADFC
  €66, ETA £24, Better World US$44.95, RACV A$47 (rescue subs) and Derby $90,
  Up Down Around $99, Maydena $1,450, Adrenalin £46 (day/season passes).
• M-c. Bloomfield "tried it and stopped" may be false. 03 notes the homepage
  says "not renting bikes at this time" while the rentals page says "Renting
  Bicycles will begin again on May 1, 2026" — three months before the scan.
  00 §14 drops the hedge and makes it a headline caution.
• M-d. Hoodoo's "4-seat minimum" is a per-vehicle minimum charge on MTB
  trailhead shuttles, sold in 00 as "the break-even for a shared van run in a
  wine region, published."
• M-e. The e-bike kill-signal lacks its denominator. Three refusals — only
  one for insurance — against ~20 operators in the same corpus renting e-bikes
  for wine tours. A risk flag, not a kill signal.
• M-f. Seasonality is inverted. 00: "The README's 30–35 sellable weekends is
  consistent with every season length found internationally." Six to seven months =
  26–30 weekends. The synthesis's own figure is below the README's by up to
  20%, and it says the opposite.
• M-g. 00 breaks the corpus's own evidence convention. Tag counts: 01 60,
  04 65, 02 34, 05 29, 06 24 — 00: one. (03 uses an undocumented
  vocabulary: [V] / [V-2°] / [DIRECTORY] / [GAP].) The one document a decision-maker
  reads offers no way to tell a fetched price from a blog snippet.
• M-h. Vine Hopper's "75-minute headway" does not measure what it is used for —
  it is five departure times from one depot, used in 00 as "the only published
  circulation interval in the world" and as the sanity check on the README's 45-min
  loop headway. Departure frequency ≠ how often a vehicle passes a point on a loop.
  This is the only external check on the single riskiest operating assumption, and
  it does not check it. (Compounds D-4 above.)
• M-i/j. Drift into the two top-level files. 05 quietly repaired two things and
  left the originals standing: RESEARCH-van-economics.md still says "the analogous
  payer is the cellar doors" (the corpus found [NOT FOUND] on winery funding)
  and "Given the Barossa finding that the van's real job is access" (05
  downgraded this to "If…"). RESEARCH-distribution.md still says TourRadar is
  "not published" and leads with "none of the 13 platforms publishes a rate" —
  which 05 disproves four times.

7.4 Minor

Currency discipline breaks where convenient (README says "no conversions applied",
but 05 says "a third of the price" NZ$65 vs A$189, and 04 says NZ$4,500
"brackets the README's A$3,500") · Wine Country Bikes' USD $49/day sits inside
a "€65–€350 per week" range · Top Bike's "€92 per riding day" divides by
nights · "+40–60%" extrapolated from a single +41% observation · "one local
council pays 60%" is one council, first-timers only, expiring 31 Dec 2026 ·
"roughly 90 operators" never reconciles with file-level counts · Argentina and
Chile are counted in the twelve-country sweep but contribute essentially no
first-party evidence · Tasmania's Bike Friendly Businesses program already exists
as an Australian state scheme, while 00 §13.6 proposes an Australian badge as novel.