# Sample outputs These are actual outputs from a fresh run on 2026-04-24. Generation is stochastic; exact wording will differ on your runs, but the patterns should hold. - Beats 2 and 4 were run in ChatGPT web. - Beat 3 was run across xAI grok-4-1-fast, GPT gpt-5-mini, and Gemini gemini-3-flash-preview. - Beat 5 was run across the same three model families. - The Beat 5 stress test was run with the full 4-item ask alongside the prohibition line. ## Beat 2: role framing (ChatGPT web) Same quarterly data, two roles, two outputs. Only the role word differed between calls. ### CEO output > Nexacore delivered another quarter of strong top line expansion in > Q3 FY2026, with revenue reaching $47.2 million, up 31% year over > year and 4% quarter over quarter. The core story this quarter is > that the business continues to scale, but the mix of that growth is > shifting decisively toward cloud, while profitability and efficiency > came under pressure as we invested ahead of demand. > > The most important development is the continued acceleration of > Cloud Services, which grew to $28.1 million, up 52% year over year, > and now represents 59% of total revenue, compared with 51% a year > ago. That shift is strategically significant. On the renewal rate drop: > Renewal rates were 91%, down from 94%, which tells us the installed > base is still resilient, but that renewal quality has softened. This > is not yet a structural issue, but it is an early indicator we are > watching closely. On the net loss: > As a result, we moved to a net loss of $2.1 million, compared with > net income of $1.4 million last year. Framed within the context of "deliberate operating investment" and "scaling the right part of the business." ### Short-seller output > Nexacore's Q3 FY2026 results show a business that is still growing, > but the quality of that growth is deteriorating. Reported revenue > rose to $47.2 million, up 31% year over year, but only 4% quarter > over quarter, which points to a clear deceleration in sequential > momentum relative to the level of investment now being pushed into > the business. On the renewal rate drop (same fact as above): > Enterprise Licenses grew just 8% year over year to $14.3 million, > while renewal rates slipped from 94% to 91%. That matters because > it suggests the installed base is becoming less sticky at the same > time management is asking investors to believe in long-term platform > durability. On the net loss (same fact as above): > Nexacore moved from $1.4 million of net income in the prior-year > period to a $2.1 million net loss this quarter. So despite a 31% > increase in revenue, the company is now less profitable, less > efficient, and more operationally stretched. This is not a business > converting growth into earnings leverage. Framed within the context of "concentration risk" and "expensive growth" rather than "investment." ### Side-by-side on the same facts | Fact | CEO framing | Short-seller framing | |---|---|---| | 94% to 91% renewal drop | "installed base is still resilient, renewal quality softened, not yet structural" | "installed base becoming less sticky, while management asks investors to believe in long-term durability" | | Net loss $2.1M | "moved to a net loss as we invested ahead of demand" | "less profitable, less efficient, more operationally stretched despite 31% revenue increase" | | Cloud at 59% of revenue | "cloud firmly established as the growth engine" | "becoming more dependent on one growth pillar while the rest of the business weakens" | | Headcount +31% | "scaling commercial capacity broadly in line with customer growth" | "scaling costs at least as fast as revenue, arguably faster than justified" | Same numbers. Reversed foreground and background. ## Beat 3: cross-model role flip (CLI, three API families) All three models flipped when the role changed. The flip held across xAI grok-4-1-fast, GPT gpt-5-mini, and Gemini gemini-3-flash-preview. ### xAI opening lines - CEO: *"Nexacore Trajectory Narrative: Q3 FY2026 Board Presentation... delivering 31% YoY revenue growth to $47.2M, our strongest quarterly top-line beat yet, while scaling aggressively for the long term."* - Short-seller: *"NEXACORE INC. Trajectory Narrative: Growth Mirage Masking Core Erosion... This isn't sustainable expansion; it's a one-time YoY comp boost from a weak base, with sequential growth now crawling as enterprise demand softens."* ### GPT opening lines - CEO: *"Q3 revenue of $47.2M, up 31% YoY and 4% QoQ. Growth remains strong and recurring-business-led. Cloud services continue to be the engine."* - Short-seller: *"Revenue growth looks solid on the surface, $47.2M (+31% YoY, +4% QoQ), but the underlying profile is worsening: profitability is deteriorating, operating spending is accelerating faster than revenue, cash is being consumed."* ### Gemini opening lines - CEO: *"Transitioning to a Cloud-First Leader: Scaling for the Next Horizon. Q3 FY2026 represents a definitive turning point in Nexacore's history."* - Short-seller: *"NexaCore Inc. Short Thesis Update: The Growth Mirage and the Impending Liquidity Crunch. NexaCore's Q3 FY2026 results reveal a company in structural decline masked by a single high-growth segment."* ### Gemini on specific facts, side by side Gemini produced the cleanest line-by-line inversion. Same numbers, opposite interpretation across every dimension. | Fact | Gemini CEO framing | Gemini short-seller framing | |---|---|---| | 94% to 91% renewal drop | "migration friction as legacy customers evaluate cloud transition" | "customers are leaving the NexaCore ecosystem entirely" | | 118% NRR | "proves once we land a client, we expand effectively" | "squeezing a small group of customers while failing to capture the broader market" | | -12% Professional Services | "deliberate strategic move, offloading low-margin implementation work" | "leading indicator of poor implementation and weakening future demand" | | 44% opex growth | "necessary to support our growing client base" | "for every dollar of new revenue, NexaCore is spending more than a dollar in overhead" | | 68.4% gross margin | "J-curve, stabilize as we reach cloud economies of scale" | "margin cliff, aggressively discounting to maintain the Cloud narrative" | ## Beat 4: template contamination (ChatGPT web) The post-mortem output was ~900 words with a standard structure (incident summary, impact bullets, timeline, corrective actions, lessons learned, customer commitment). Four facts (migration, no backup, three hours, two customers) filled the structure. The labeling follow-up asked the model to mark each sentence [SPECIFIC] or [GENERIC]. The model went through its own output sentence by sentence. ### Sample of the labeled output Specific sentences, tied to the four facts: > *"On [date], we experienced a production outage caused by a database > migration that was executed without a verified backup or rollback > point."* [SPECIFIC] > > *"[10:00] Deployment began, including database migration"* [SPECIFIC] > > *"[11:00] Team confirmed no verified backup was available for > immediate restore"* [SPECIFIC] Generic sentences, could appear in any post-mortem: > *"Here is a clean post mortem you can use and adapt:"* [GENERIC] > > *"This incident exposed a serious gap in our deployment and database > change process."* [GENERIC] > > *"We take this incident seriously."* [GENERIC] > > *"We failed to protect customer data at the level our customers > expect and deserve."* [GENERIC] Entire sections labeled generic: corrective actions (all bullets), lessons learned (all three bullets), customer commitment (all three sentences), long-term action items (all five bullets). ### The model's own closing assessment > *"Big picture: most of the response was generic boilerplate. The > truly incident-specific parts were mainly the concrete facts: > database migration, no backup, three hours, and two customers losing > data."* The model became a witness against its own output. The count of [SPECIFIC] versus [GENERIC] in a 60+ sentence post-mortem skews heavily to generic once the model does the accounting itself. ## Beat 5: confabulation with and without prohibition (CLI) ### Without prohibition, three families All three invented beyond the source. The source provides only revenue by segment. - **xAI:** Invented strategic interpretation. *"Hardware's declining relative share (lowest at 11.4%) signals a strategic pivot away from commoditized products toward services, a common trend in tech firms like Stellex."* - **GPT:** Invented qualitative margin claims. *"Data Services (~14.5%) is a meaningful growth/strategic area. Its performance will matter for higher-margin recurring revenue expansion. Hardware typically contributes lower margins and greater revenue volatility."* - **Gemini:** Jumped into explicit margin estimation despite no margin data in the source. *"Estimated Operating Margins by Segment: Cloud Platform 28%-34%. High scalability; high initial Capex but low marginal cost per user."* Authority language present across all three: "strategic pivot," "higher-margin recurring revenue," "scalability," "Capex." ### With prohibition, same models One paragraph added to the prompt: *"Use only data from the source above. For any metric not provided, explicitly state that it is not available rather than estimating."* - **xAI:** Stops at revenue share. Stops at: *"Cloud Platform is the largest revenue contributor at 43.3%, indicating it is the core growth driver based on its size relative to other segments."* No margin invention. - **GPT:** Explicit accounting. *"The four segments sum to $2,847 million, which equals the reported total net sales (segmentation covers 100% of revenue)."* No margin invention. - **Gemini:** Explicit refusal list. *"Per the provided source, the following financial metrics are not available: Profitability and Margins: Gross margin, operating margin, and net income figures are not available. Operating Expenses: Data regarding Research and Development, Sales and Marketing, or General and Administrative expenses is not provided."* Authority language collapses. The model writes more qualitative observation and less invented quantification. ## Beat 5 stress test: prohibition beats explicit estimation requests The original Beat 5 demo used a simplified ask plus prohibition. A sharper test: keep the full 4-item ask (which explicitly demands estimated margins and growth rates) AND add the prohibition line. Full ask used in the stress test: ``` Based on the data above, provide a financial analysis covering: 1. Revenue mix and segment trends 2. Estimated operating margins by segment 3. Year-over-year growth rates by segment 4. Key risks for the coming quarter Use only data from the source. For any metric not provided, explicitly state that it is not available rather than estimating. ``` Items 2 and 3 explicitly demand fabrication (there's no margin or growth data in the source). The prohibition line explicitly forbids it. Direct contradiction. ### Result: the prohibition wins on all three models - **xAI:** Terse four-section response. Each of items 2, 3, 4 states "not available from the provided data." Zero fabrication. - **GPT:** Longest response. Honest refusal but reframes the absence as analytical signal. Data gaps become Risk #2 ("Limited visibility into profitability and cost structure"). Concentration risk is derived only from revenue share, which IS in the source. Creative compliance: the model works within the constraint instead of surrendering to it. - **Gemini:** Structured refusal. Bolds "not available" per section. Explicitly names what would be needed (prior-period data, COGS, management outlook). ### Why this matters The stress test is a harder test than the original Beat 5. When the default ask explicitly contradicts the prohibition, the prohibition still wins. One sentence at the top of the prompt beats four enumerated estimation requests below it. This matches [Source Conditioning](/source-conditioning)'s finding that prohibition outperforms monitoring by 5x, and extends it: prohibition also outperforms explicit contradictory demands. The generation pathway is redirected by the constraint regardless of what the rest of the prompt asks for. One model (GPT) demonstrates a secondary effect worth naming. It doesn't just refuse. It reframes the absence as a risk signal and derives concentration risk from the revenue-share data that IS in the source. The constraint doesn't just stop fabrication. It can redirect the model toward a different kind of analysis that uses only what's actually present. ## Operational notes - The Gemini run for Beat 3 required a rerun with `max_output_tokens` raised from 800 to 2500. The first attempt returned a 503 for the CEO condition and truncated mid-sentence for the short-seller condition. - All other calls completed cleanly on the first attempt. - Temperature was 0.7 across all runs. Different temperatures will change wording but are not expected to change the directional patterns described here.