Excel to PPTX, messy export
On a tidy workbook every model got every figure right, with or without the toolkit prompt. What happens on an export with subtotal rows, intercompany revenue, negatives in parentheses and a renamed region?
Recorded 8 Oct 2026 against the live APIs · run 2026-10-08-r1 · 120 calls · replayed here, not live. Download the run · Prompt tested: modelcensus_pptxfromexcel_any_v1.txt · Failure mode: Schema-Correspondence Fabrication · Failure mode: Method/Execution Split · Failure mode: Dimensional-Analysis Failure · Failure mode: Constraint Decay at Depth · Failure mode: Unsupported-Claim Confidence
| Figure | GPT-6 Luna · bare | GPT-6 Luna · toolkit | Gemini 3.8 Flash · bare | Gemini 3.8 Flash · toolkit | Claude Opus 5.5 · bare | Claude Opus 5.5 · toolkit |
|---|---|---|---|---|---|---|
| Q3 total | 100% | 100% | 85% | 95% | 100% | 100% |
| Revenue by region (8) | 99% | 97% | 100% | 99% | 100% | 100% |
| Growth vs Q2, incl. n/a (8) | 100% | 100% | 99% | 95% | 100% | 100% |
| Against target (8) | 99% | 96% | 100% | 98% | 100% | 100% |
| Gross margin | 100% | 100% | 95% | 80% | 100% | 100% |
Replay a deck
Pick a workbook and a model. Left is the bare request; right is the same request with the toolkit prompt. Each figure is checked against the workbook in code.
- 1Q3 2026 Business Review
- · Leadership team review: July to September 2026 actuals
- · All revenue is external only; intercompany revenue is excluded throughout
- · Amounts in USD; source: q3_pnl_export.xlsx
- 2Total Q3 Revenue: $12.14M
- · Q3 2026 external revenue: $12.14M ($12,136.3K)
- · Up 13.4% vs. Q2 ($10.70M), including the Gulf region, which opened in July
- · Like-for-like growth excluding Gulf: +6.1% ($11.36M vs. $10.70M)
- · 98.8% of the combined Q3 target of $12.28M, a $143.7K shortfall
- 3Q3 Revenue by Region vs. Q2
- · Pacific: $1,934.9K (+6.7% vs. Q2 $1,813.2K)
- · Midwest: $1,905.2K (+11.5% vs. $1,709.3K)
- · Southwest: $1,837.0K (+10.1% vs. $1,668.9K)
- · Southeast: $1,496.1K (+15.8% vs. $1,292.4K), the fastest-growing region
- · Northeast: $1,494.1K (+1.3% vs. $1,474.2K)
- · Atlantic: $1,361.5K (-2.3% vs. $1,393.6K)
- · Mountain: $1,327.0K (-1.5% vs. $1,347.8K)
- · Gulf: $780.5K (new in July 2026; no Q2 comparison)
- 4Q3 Revenue vs. Target by Region
- · Above target: Gulf $780.5K vs. $710K (109.9%, +$70.5K)
- · Above target: Southwest $1,837.0K vs. $1,730K (106.2%, +$107.0K)
- · Above target: Southeast $1,496.1K vs. $1,420K (105.4%, +$76.1K)
- · Near target: Mountain $1,327.0K vs. $1,340K (99.0%, -$13.0K)
- · Near target: Midwest $1,905.2K vs. $1,930K (98.7%, -$24.8K)
- · Below target: Northeast $1,494.1K vs. $1,580K (94.6%, -$85.9K)
- · Below target: Atlantic $1,361.5K vs. $1,450K (93.9%, -$88.5K)
- · Below target: Pacific $1,934.9K vs. $2,120K (91.3%, -$185.1K), the largest gap
- · Total: $12,136.3K vs. $12,280K (98.8%, -$143.7K)
- 5Q3 Gross Margin: 40.2%
- · Gross profit of $4.88M on $12.14M external revenue; external COGS of $7.26M
- · Highest margins: Midwest 48.0% and Southeast 44.4%
- · Mid-range margins: Pacific 40.8%, Southwest 40.0%, Northeast 37.2%
- · Lowest margins: Atlantic 35.8%, Gulf 35.5%, Mountain 34.3%
- 6Top Three Regions (by Q3 External Revenue)
- · 1. Pacific: $1,934.9K; +6.7% vs. Q2; 40.8% gross margin; but 91.3% of target
- · 2. Midwest: $1,905.2K; +11.5% vs. Q2; best margin at 48.0%; 98.7% of target
- · 3. Southwest: $1,837.0K; +10.1% vs. Q2; 40.0% gross margin; 106.2% of target
- · Together these three account for $5.68M, or 46.7% of Q3 revenue
- 7Risks
- · Pacific is the largest region but is $185K (8.7%) below target
- · Atlantic (-2.3%) and Mountain (-1.5%) shrank vs. Q2; Atlantic is also 6.1% below target
- · Northeast is nearly flat (+1.3%) and $86K below target
- · Margins are lagging in Mountain (34.3%), Gulf (35.5%) and Atlantic (35.8%)
- · Headline growth is flattered by Gulf's launch: 13.4% reported vs. 6.1% like-for-like
- · Gulf monthly revenue is volatile ($229K in July, $314K in August, $237K in September), so its run-rate is still unproven
- 8Next Steps
- · Pacific: diagnose the target gap and build a recovery plan
- · Atlantic and Mountain: review pipeline and pricing to reverse the quarter-over-quarter decline
- · Margins: run COGS and mix reviews in Mountain, Gulf and Atlantic, using Midwest and Southeast as benchmarks
- · Gulf: track monthly performance to establish a stable baseline
- · Reporting: standardize on external revenue and current region names (e.g., Pacific, not West Coast)
- 1Q3 2026 Business Review
- · Leadership team review
- · Scope: Q3 2026 (Jul–Sep) actuals only
- · All revenue is external: intercompany excluded
- · Source: q3_pnl_export.xlsx; all figures in USD
- 2Total Q3 Revenue
- · Q3 2026 external revenue: $12,136,300 (USD)
- · Up 13.4% vs Q2 2026 ($10,699,400 USD), including Gulf, which opened in July
- · Up 6.1% like-for-like, excluding Gulf ($11,355,800 USD)
- · Q3 target: $12,280,000 (USD). Actual is 98.8% of target, a $143,700 (USD) shortfall
- 3Q3 Revenue by Region vs Q2
- · Pacific: $1,934,900 (USD), +6.7%
- · Midwest: $1,905,200 (USD), +11.5%
- · Southwest: $1,837,000 (USD), +10.1%
- · Southeast: $1,496,100 (USD), +15.8%
- · Northeast: $1,494,100 (USD), +1.3%
- · Atlantic: $1,361,500 (USD), -2.3%
- · Mountain: $1,327,000 (USD), -1.5%
- · Gulf: $780,500 (USD), growth n/a (opened Jul 2026, so no Q2 data)
- 4Q3 Revenue vs Q3 Target
- · Gulf: $780,500 vs $710,000 (USD), 109.9%, +$70,500
- · Southwest: $1,837,000 vs $1,730,000 (USD), 106.2%, +$107,000
- · Southeast: $1,496,100 vs $1,420,000 (USD), 105.4%, +$76,100
- · Mountain: $1,327,000 vs $1,340,000 (USD), 99.0%, -$13,000
- · Midwest: $1,905,200 vs $1,930,000 (USD), 98.7%, -$24,800
- · Northeast: $1,494,100 vs $1,580,000 (USD), 94.6%, -$85,900
- · Atlantic: $1,361,500 vs $1,450,000 (USD), 93.9%, -$88,500
- · Pacific: $1,934,900 vs $2,120,000 (USD), 91.3%, -$185,100 (target is listed under 'West Coast')
- 5Q3 Gross Margin
- · External revenue: $12,136,300 (USD)
- · External COGS: $7,257,600 (USD)
- · Gross profit: $4,878,700 (USD)
- · Gross margin: 40.2% (%, external revenue basis)
- 6Top Three Regions (Q3 External Revenue)
- · 1. Pacific: $1,934,900 (USD), 91.3% of target
- · 2. Midwest: $1,905,200 (USD), 98.7% of target
- · 3. Southwest: $1,837,000 (USD), 106.2% of target
- · Together: $5,677,100 (USD), 46.8% of total Q3 revenue
- 7Risks
- · Pacific has the largest gap to target: -$185,100 (USD)
- · Atlantic is down 2.3% QoQ and $88,500 (USD) below target
- · Mountain is down 1.5% QoQ
- · Northeast grew only 1.3% QoQ and is $85,900 (USD) below target
- · Gulf has one quarter of history, so its trend is n/a
- 8Next Steps
- · Pacific, Atlantic and Northeast: recovery plans against the Q3 gaps
- · Mountain and Atlantic: review the causes of the QoQ decline
- · Gulf: track ramp-up now that a Q2-style baseline is unavailable
- · Southeast and Southwest: identify and replicate what drove growth
- · Finance: confirm the 'West Coast' target is Pacific's
What this shows, and what it doesn't
Claude Opus 5.5 got all 26 figures right on all 40 decks, with or without the prompt. The errors came from the two cheaper models, and the toolkit prompt didn't remove them: 54 of 60 bare decks were fully right against 49 of 60 with the prompt, at 1.6× the output tokens. Twenty decks per cell is small; read it as no improvement, not as proof the prompt hurts.
Gemini 3.8 Flash's misses were arithmetic, in both arms: two totals off by exactly $1,000,000, margins off by one to eight points, a growth rate of −8.3% for −1.5%. GPT-6 Luna got the new region's revenue wrong (the one with blank Q2 cells) in six decks, five of them with the toolkit prompt.
One error was the prompt's own. Its mapping guard says to mark a key UNMAPPED rather than infer it, and twice a model applied that to a region the lookup sheet documents as 'formerly Central', printing 'target n/a'. No bare deck did that.
Every wrong deck was caught by checking its figures in code against the workbook, the check the toolkit SDK ships. Choosing the model and checking the numbers moved the result; a longer prompt didn't.