Grounded vs. Synthetic: One Question, Two Answers

Free LLMs are smart. But when a question needs live market data they can’t see, they don’t refuse. They confidently fill in the numbers and answer anyway. Below is the same question sent two ways: to a top frontier model (Gemini 3.1 Pro, no tools) and to Context’s grounded pipeline. Read the two verdicts. They disagree, and the reason they disagree is the entire point of Context.

Gemini 3.1 Pro · no tools

Verdict: SHORT SQUEEZE (bullish reversal)Built on the prompt’s assumption that OI was climbing. Returned a plausible narrative plus a 100-line Python script for you to run on your own Crypto Data feed. No live numbers. No chart.

Context · grounded

Verdict: Neutral / MixedBuilt from the live 7-day Crypto Data feed. Open interest was actually contracting −4.35%. Returns the rendered dashboard, the exact metrics, and a grounded trade read.

The frontier model assumed open interest was climbing, because that is how the prompt framed it. Context checked the live feed and found OI had fallen 4.35% over the same window. That one invented premise flips the whole trade verdict from “bullish short squeeze” to “neutral, wait-and-confirm.”

The question

A single, demanding BTC order-flow prompt: aggregate funding and open interest across all exchanges over 7 days at 1h granularity, overlay price, run rolling CVD regressions, classify the regime with the OI matrix, scan for absorption, and deliver one dashboard plus a trade verdict.

Read the full prompt

Please read the order flow for me on BTC. Using Crypto Data over the last 7 days
at 1h granularity, aggregate funding rate and open interest across all
exchanges; overlay BTC 1h close on the same timeline so price action reads
against positioning. Chart funding and OI with hourly delta histograms
underneath so the recent 48h funding dip and steady OI climb are obvious. Add
aggregated spot CVD and aggregated perp CVD at the same 1h cadence and window:
normalize each for relative strength on a rolling basis, then chart the
difference in normalized relative strengthening between spot and perps. For
each CVD leg (spot and perp), run rolling linear regression on bar index versus
CVD level over 24h and 48h lookbacks. Plot slope, R², and adjusted slope =
slope × R² as time series so trend direction, trend quality, and how
aggressively each leg is accumulating are visible. Classify the regime with the
open-interest matrix: price rising/falling × OI rising/falling × volume
rising/falling. Layer CVD absorption (uptrend absorption = bearish; downtrend
absorption = bullish), check spot and perp CVD separately, map the combined
read to who is trapped, and from funding sign and liquidation imbalance deliver
one dashboard chart plus a clear trade-setup verdict.

A · Gemini 3.1 Pro, no tools

With no access to the live feed, the model accepted the prompt’s framing (“steady OI climb”) and reasoned from there. It produced a confident, well-structured analysis, and then handed back a Python engine for you to run yourself, because it never had the data.

Trade-Setup Verdict: SHORT SQUEEZE (Bullish Reversal) With negative funding, rising OI, and bullish downtrend absorption confirmed, the market is primed for a violent short squeeze. […] Who is trapped? New Shorts.

Notice what is missing: no real funding value, no real OI delta, no liquidation figure, no rendered chart. The conclusion rests entirely on the assumption that OI was rising.

Gemini's full response (narrative + Python engine)

The model returned a multi-section narrative followed by a complete pandas / scipy / matplotlib script (roughly 100 lines) ending with:

# To run: generate_orderflow_dashboard(crypto_data_btc_df)

In other words: here is the code to build the dashboard once you’ve sourced and cleaned the Crypto Data yourself. The model can describe the method perfectly. It cannot fetch the data or produce the chart.

B · Context, grounded

Context ran the live pipeline against the real Crypto Data feed, computed every metric the prompt asked for, and returned the rendered dashboard plus exact numbers.

BTC order-flow dashboard: price vs open interest, funding rate with hourly deltas, normalized spot vs perp CVD, and rolling adjusted-slope panels

The real 48-hour metrics

Metric	Value
48h price change	+0.01% (effectively flat)
48h OI change	−4.35% (contracting)
Avg funding (48h)	+0.0069% / 8h
Latest funding	+0.0056% / 8h
Liquidation imbalance	−$219M (sell liqs ≫ buy liqs)
Spot CVD adj. slope (24h)	+$8.67M/hr
Perp CVD adj. slope (24h)	−$0.66M/hr

The grounded read

Regime (OI matrix): flat price + falling OI + below-average volume lands on weak uptrend / short-covering, not the new-money strong uptrend a bullish squeeze would need.
Absorption: spot CVD trending up (+$8.67M/hr) while perp CVD is flat-to-negative = real spot absorption of perp-driven selling.
Who’s trapped: earlier shorts are under pressure (−$219M sell-liq imbalance, funding still positive), but mid-week longs are exposed because OI is declining, so no fresh margin confirms the move.

Verdict: Neutral / Mixed: lean cautiously bullish on absorption, but wait for confirmation. Watch for a reclaim of local highs with simultaneous OI expansion and funding dipping toward zero. Until then, this is a wait-and-confirm environment, not a high-conviction entry.

Data: Crypto Data, 7-day 1-hour aggregate across Binance, Bybit, OKX, Hyperliquid, Deribit (perps) and Binance, Bybit Spot, OKX, Coinbase (spot).

What this shows

Ungrounded models echo your premise

Gemini had no feed, so it trusted the prompt’s “steady OI climb” and built a bullish squeeze thesis on top of it. The reasoning was clean; the foundation was assumed.

Grounded answers check the premise

Context pulled the live data, found OI was actually down 4.35%, and the verdict changed. The data is the deliverable, including the chart you can act on.

Try it yourself

The grounded answer above came from the same public SDK surface your agent can hit today. Drop in the full prompt from the accordion above:

import { ContextClient } from "@ctxprotocol/sdk";

const client = new ContextClient({ apiKey: process.env.CONTEXT_API_KEY });

const result = await client.query.run({
  query: "Read the order flow on BTC using Crypto Data over the last 7 days at 1h granularity...",
  responseShape: "answer_with_evidence",
  includeDataUrl: true,
});

console.log(result.response); // grounded verdict
console.log(result.dataUrl);  // rendered dashboard + full dataset

Your agent pays per response in USDC. No subscription, no KYC, no invented numbers.

Start with the SDK

Drop the Lite MCP into your agent in one config block.

How the runtime works

See the full architectural story behind the grounded answer above.

Gemini 3.1 Pro · no tools

Context · grounded

​The question

​A · Gemini 3.1 Pro, no tools

​B · Context, grounded

​The real 48-hour metrics

​The grounded read

​What this shows

Ungrounded models echo your premise

Grounded answers check the premise

​Try it yourself

Start with the SDK

How the runtime works

The question

A · Gemini 3.1 Pro, no tools

B · Context, grounded

The real 48-hour metrics

The grounded read

What this shows

Try it yourself