Quality audit · Verdal Group (fictional case)

Synthesis and decisions required

Three independent passes over the same deliverable, and what they concluded.

WP1 Market — three-level QA: synthesis, recommendations and decisions required

Documents audited: livrables/verdal-wp1-market.pptx · livrables/verdal-wp1-market-model.xlsx Date: 27 July 2026 · Engagement: VG-STRAT-2026-07 · Nothing was modified.

The three passes: | Level | Method | Report | |---|---|---| | L1 | Source verification — 47 claims traced to primary, two independent web-research streams | wp1-audit-L1-source-verification.md | | L2 | Internal audit — isolated Cowork subagent, dynamic grid, model rebuilt from its 30 inputs | wp1-audit-L2-internal-audit.md | | L3 | Dual external audit — Claude Fable 5 + GPT-5.6 Sol, identical prompt, neither seeing the other | wp1-audit-L3-dual-external-audit.md |


The one-paragraph answer

The model is the strongest artefact in this engagement and the deck is currently writing cheques it does not cash. All three passes independently reach the same shape of verdict: the workbook's honesty discipline is at or near best practice — every input carries a value, unit, confidence and source, and it names in capitals the two cells that would flip the answer — while the deck drops that discipline, overstates several claims the model itself disclaims, and puts two exhibits on a Board page that no cell in the model computes. Not one of the defects requires new research. What does require judgement is the €25bn perimeter, which deck and model currently answer differently, and the Savills benchmark, which the L1 pass has now settled at primary and which moves further than expected.

Scores: L2 internal 66/95 (3.5/5) · Fable 5 3.5/5, verdict fixable in a day · GPT-5.6 Sol 2.2/5, verdict HOLD.


What changed as a result of the audit — three things you did not know this morning

1. The Savills benchmark is settled, and it carries a second figure you do not have

Read at primary: Savills Research, Electric Vehicle Charging: Maximising Opportunity, UK Commercial, Autumn 2024. £3,000-5,000 per bay per annum, or 5% revenue share, or 20% profit share, 30-year lease, CPI review. CONTEXT.md open decision #4 closes: upgrade Assumptions!F37 from Probable / contested to Verified.

Three qualifiers, and they do not all cut the same way:

2. The funnel's opening multiplication mixes two perimeters

IEA's 4.3% is IEA-"Europe" (EU-27 plus UK, Norway, Switzerland, Türkiye and the non-EU Balkans) measured against final electricity consumption. Ember's 2,790 TWh is EU-27 gross generation. The two errors run in opposite directions and do not cancel predictably. This is the calculation that tests the Board's figure, and it is the one a Board member with an energy background will catch. Cheap to fix — align both terms on one perimeter — and nothing in the argument breaks, because the funnel is already declared a bounding exercise.

3. Six sourced claims are contradicted by their sources

The one that matters most: the ACEA/McKinsey Masterplan is not "of the same order" as Société Générale's €30bn. ACEA publishes €8bn a year for public charging — roughly €70-80bn cumulative to 2030, more than double. The deck smooths a factor-of-two contradiction between two of its own sources. That is the single place in the package where a contradiction was smoothed rather than presented, and it is the exact behaviour the client said it would score against. Also: the Grand View 2030 figure is no longer reproducible at source; Market Data Forecast's US25.4bnbelongstoits * charger * reportwhileits * chargingstation * reportsaysUS671bn for the same year and continent; the ADAC primary could not be reached at all; ANFAC publishes 746,510, not 0.69m; Eurostat's Italy figure contradicts Eurostat's own published ranking.

One thing got stronger, not weaker. The McKinsey mobility-finance report is verified at primary — 11 October 2023, five countries (FR, DE, IT, ES, UK), "about €25 billion of new annual revenue for mobility-finance providers by 2030". And it contains a sentence the deck does not use: "EV charging infrastructure… will remain a relatively niche segment, contributing about 1 percent of total 2030 revenue." If that is the Board's source, then even inside the €25bn, charging is about €250m — and it is charging finance, not charging service revenue. The falsification test also held: no published source cleanly produces "European EV charging market, €25bn, 2030".


Consolidated fix list

Ranked by risk, with the number of independent passes that found each. None requires new research.

Tier 1 — must not go in front of a Board uncorrected

# Defect Found by Fix
1 Slide 6's bridge step "Cost of delivery −40.5" is wrong. €40.4m is the operator's retained 85% under a revenue share. The actual cost of delivery is €70.0m on slide 12 of the same deck. L2, Fable Re-label "Operator's 85% share, −40.4", or split the last leg into two steps. Add tab 3!D26 so it is live.
2 Deck and model give different answers to the mandate question. Slide 4 and tab 4!C5 put €25bn at level 1 (capex); tab 1!A20 puts it at level 2 (plug revenue). L2, GPT Decision required — see D-1 below.
3 Slide 13's "Charger spec" and "Bays per charger" bars exist in no model cell. F17 itself admits it. The deck promises on slide 14 that the grid "is live, not a picture" — slide 13 is a picture. L2, Fable, GPT Build them as live formulas in tab 5 section C, or strip the two bars.
4 Slide 13's title is contradicted by slide 13's chart — "two equipment assumptions do", but only charger spec crosses zero (+0.6); bays per charger stops at −2.8. L2 Retitle.
5 Slide 14's kicker is contradicted by its own last column — "more bays make the loss larger" is false at 160 kWh/bay/day, where 8 bays gives +3.2. L2, GPT "Below break-even throughput more bays deepen the loss; only above about 160 kWh a bay a day does scale help."
6 The deck's headline prize (€7m) contradicts the model's own definition of the prize (€0.537m) — and tab 3!D25 explicitly warns against exactly the comparison slide 2 makes. L2 Decision required — see D-2.
7 FX error on the Savills range and the "2-3x" multiple. L1, L2, Fable, GPT Correct to €15.4-25.7m and 2.3-3.9x; state the rate.
8 Slide 8's step percentages and slide 6's bridge steps are computed from rounded display labels, not from cells. −69.4% and −42.3% should be −70.0% and −43.0%. L2, Fable Recompute from unrounded cells, or show the multiplier (×30%, ×22%) instead of a spurious-precision delta.
9 Slide 8 credits the IEA with the 30% public split that the model says nobody publishes. L2, Fable "The IEA sources the 4.3% EV share; the 70/30 public split is commonly cited but is not published as a kWh split — it is our first soft step."
10 Slides 15 and 16 say the lease fee is "the only" unsupplied Verdal figure — the same slide lists two more (bay counts, energy contract). Fable Delete "only"; say three.
11 The ACEA "same order" sentence smooths a factor-of-two contradiction between two of the deck's own sources. L1 Present both constructions and say they disagree.
12 Slide 2's "on a destination-grade 75 kW unit it is reached today" is an overclaim — the model shows about break-even at 2030 modelled throughput on an unsourced assumption. Fable, GPT "…break-even at base-case 2030 throughput, on an unverified throughput assumption."

Tier 2 — before the interim on 8 September

# Item Found by
13 Put the Verified / Probable / Hypothesis taxonomy on the slides. It is applied rigorously in the model, appears on zero slides, and is the client's explicit scoring criterion. L2, Fable, GPT
14 Give the lease-fee flip condition a visual on slide 11, not a bullet. A44 says the recommendation flips; the headline still asserts the revenue share wins. L2, Fable, GPT
15 Solve the break-even against the whole estate — 162.2 kWh/bay/day, not the DC-only 159.8. The model's own grid is still −€0.6m at 160. L2, GPT
16 Repair or drop the three market-research citations (Grand View not reproducible; Market Data Forecast wrong report of two that differ 26×; Meticulous horizon 2032 not 2030). L1
17 Fix the six contradicted source claims and move the five confidence labels (DC price Verified→Probable; AC price Probable→Hypothesis; electricity Verified→Probable; 2,790 TWh re-labelled; Savills Probable→Verified). L1
18 Disclose or decompose the possible network-charge double count. Eurostat's band IC price already includes network charges; the model adds €0.04/kWh of capacity charges on top — €3.37m a year. Also state the band basis (84.3 GWh ÷ 1,100 sites = 76.6 MWh/site, below band IC's 500 MWh floor unless charging shares the store meter). GPT
19 State that site eligibility has not been screened. One line on slide 9 and a hand-off to WP3. Savills' own qualifiers make this material. GPT, L1
20 Deliver methodology Step 5 — a bounded range on the headline, not a point estimate. Lease €6.6m base / €15.4-25.7m at Savills; revenue share €4.8m at 10% to €11.9m at 25%; operate −€22.4m to +€0.7m. Build as tab 5 section D. L2, GPT
21 Align the funnel's opening multiplication on one perimeter (IEA-Europe consumption, or an EU-27 EV share). L1, GPT
22 Scope and cost the pilot so 8 September is a real budget decision. Fable, GPT
23 Add the like-for-like ratio to slide 4: €7.1m against €766m = 1/107th, alongside the 1/2,850th. Someone in the room will do that division. L2
24 Correct Assumptions!C24 — €0.19 is presented as the Eurostat figure (€0.1837) with no stated reason for the 3.4% uplift. And Interpath's own footnote assumes €180,000 per charger, not the €200,000 used — worth €1.8m a year. L1, L2
25 Re-date and version the deck for the 8 September interim; the 27 July build date precedes the 4 August kick-off. Fable, GPT

Tier 3 — improvements, not defects

Slide 5 as a table with currencies and horizon years visible · slide 2 from six supports to three or four · a fourth section divider before slides 15-16 (they currently sit under "What a site host actually earns") · slide 14's missing column-axis unit, restored from tab 5!A18 · slide 17 finished with the doc ID and a leave-behind recap · end on a governing recommendation rather than a list of requests · correct slide 4's level-4 row (under self-operation Verdal is operator and host) · add the two Tier-1 cross-checks L1 surfaced: Fastned's own published supermarket business case (40 sessions/station/day; max €252k revenue per station against €1,417k for a motorway site) and Pod Point's Tesco estate (£2,983 operating profit per site per year — below the bottom of the Savills per-bay rent).


Decisions required from you

Everything above except these can be executed without you. These cannot.

D-1 · Which perimeter level does €25bn occupy — or do you refuse to choose?

The package currently holds two answers. Three routes:

Recommendation: (c). It is the only one that is true today, the kick-off already committed to "present all candidates, assert none", and it is precisely what the client said it would score on. Both L2 and GPT-5.6 Sol independently reached the same conclusion.

D-2 · What is WP1's headline number — €7.1m or €0.5m?

€7.1m is the pool at Verdal's perimeter level. €0.537m is the incremental prize over today's assumed lease, and it is what tab 3!D25 calls "THE ACTUAL PRIZE" — and what the methodology names as the comparison the Board paper fails to make.

They lead to different Board conversations. €0.5m is the more defensible number and the weaker headline; it may also make the case for WP2-WP6 harder to hold. Note that at the now-Verified Savills range the incremental prize is negative — today's lease would already be winning.

Recommendation: lead with the pool and the increment together, on the same line. "The pool at Verdal's level is €7.1m a year. The increment over today's lease is €0.5m — and on the only published benchmark it may be negative." That is one sentence and it removes the attack.

D-3 · Savills is now Verified. How much of it do you put in front of the Board?

Four sub-questions, and they compound:

  1. Do you present the CPO-funded range (£3,000-5,000/bay → €15.4-25.7m) at all, given it flips the recommendation?
  2. Do you also present the landlord-funded benchmark ("upwards of £10,000 per bay" → c. €51.5m), which is the structure the Board is actually contemplating and which the deliverable does not currently carry?
  3. Do you carry Savills' three site qualifiers as an explicit brake, or does presenting them concede that the addressable estate is unscreened?
  4. Or do you hold the entire margin comparison until the two lease agreements arrive?

Recommendation: present both benchmarks with the qualifiers, and hold the recommendation. Show the flip as a scenario strip, say that the answer to the Board's own question — "is value being left on the table?" — may be no, and that the two agreements settle it. There is no third option that is both honest and useful; withholding a Verified benchmark you hold is worse than showing it with its limits.

D-4 · Which frame governs the fix — Fable's or GPT's?

The two external models scored the same package 3.5/5 and 2.2/5. The gap is not noise; it is two legitimate frames:

Recommendation: Fable's frame for WP1, with two of GPT's objections adopted now. NPV/IRR and country economics belong to WP3 and WP5 and should not be pulled forward. But the site-eligibility caveat costs one line and is corroborated by Savills' own qualifiers, and ending on a governing recommendation rather than a list of requests costs one slide. Take those two; defer the rest with the hand-off written down.

D-5 · Does WP1 state a view on self-operation, or defer to WP2?

The deck currently says "WP1 must not rule out self-operation: WP2 chooses the spec, and it decides." Methodologically correct. To a Board it may read as a refusal to answer. The alternative is to state a conditional verdict — "on a highway-grade spec self-operation loses €22m; on a destination-grade spec it is roughly break-even; WP2 chooses, and that choice is the whole answer" — which is the same content with a spine.

Recommendation: the conditional verdict, and put the two specs side by side on slide 12 rather than in sequence across 12 and 13. The current sequencing lets a Board member conclude the −€22.4m is a straw man before slide 13 admits it.

D-6 · Keep, soften or move the McKinsey mobility-finance line?

It is now verified at primary and it is the sharpest piece of research in the deck — the report matches the figure, the currency and the year exactly and measures something else entirely. It is also the line most likely to cost Katrin Sommer political capital: "the Board paper cited the right number from the wrong report" will be heard as the Board confused an auto-loan report with charging, whatever the hedge on the next line.

Recommendation: keep the finding, delete the sentence. State the match, state what the report measures, state the 1%-of-revenue line, and end on the citation request. Never characterise the Board's citation.

D-7 · Whether to brief Sommer before 8 September

The perimeter finding is stable, depends on no data-room item, and reverses the premise of a Board paper her Board wrote. Six weeks is a long time for a sponsor to be unaware of that. This is a relationship call, not an analytical one.

D-8 · Three items that belong at firm level, not here

Per the asymmetric rule, these cannot be settled in this engagement folder:

  1. A house FX convention — the Savills error happened because no rate was stated. "Every cross-currency figure states its rate and its date" is a firm rule, not a Verdal rule.
  2. Does "every page carries its source line, no exception" apply to section dividers? GPT-5.6 Sol scored a standards breach on slides 3, 7, 10 and 17; L2 and Fable read dividers as exempt. Settle it once.
  3. The template has no footer or slide-number placeholder on any of its four layouts — measured, not inferred. No slide in this deck carries a page number, so nobody in the room can say "turn to page 12". That is a template-level fix.

If you want any of these to stick, take them to the firm workspace and say: "improve the consultancy: [rule]".


What the audit did not find

Worth recording, because it bounds the fix: