TRACKED CALLS / 24H1,284+12.4%SMS RECOVERIES / WK287+8.1%DEMAND REACTIVATION RESPONSE RATE41.0%+3.2ppAU + NZ CLIENTS LIVE512+11AVG QUOTE RESPONSE47 MIN-9 minISO 27001PLATFORMTRACKED CALLS / 24H1,284+12.4%SMS RECOVERIES / WK287+8.1%DEMAND REACTIVATION RESPONSE RATE41.0%+3.2ppAU + NZ CLIENTS LIVE512+11AVG QUOTE RESPONSE47 MIN-9 minISO 27001PLATFORM
GibsonFamily-Owned Sydney Lead Engine · Since 2006
1800 950 347Get a quote
LIVESANDBOXAI BUILDSCASE STUDYSYDNEY · MELBOURNE · BRISBANE · PERTH · AUCKLANDMon to Fri · 9am to 5pm AEST

Sandbox · We run this ourselves. Not as a sales pitch, as the day job

// Internal lab · multi-model AI experiment · not for sale

Investing AI Terminal

By Albert Triolo, Gibson Promotions ·
// THE BASICS

What is a multi-model AI financial signal terminal?

A multi-model AI financial signal terminal is an internal laboratory instrument that routes live market signals through multiple large language models simultaneously and compares how each scores them. Gibson Promotions built one in April 2026 to answer a direct question: do AI models agree on news significance, and does disagreement predict errors a single model would miss? The terminal runs an AUD-denominated portfolio across momentum, dividend, and venture buckets. News signals, VIP-operator signals, and geo-signals flow through Claude Sonnet, Claude Opus, Claude Haiku, Gemini, and open-model baselines from Hugging Face. Every score, every trade, and every benchmark comparison is logged. The experiment is currently down 38.4 percent versus SPY since inception. That result is published openly because the point of this laboratory is not equity alpha: it is calibration data for every AI scoring system Gibson ships.

“The terminal is down 38.4 percent versus SPY since we started it in April 2026. We publish that number because hiding a losing experiment defeats the purpose of running one. Every losing signal teaches us something about how AI over-scores novelty and under-weights context. That learning is now inside every Gibson client system that runs AI scoring.”

Albert Triolo, Founder, Gibson Promotions

“We tested five model variants against the same news signal on the same day. Two said buy, two said hold, one flagged the source as low-credibility. The outlier was right. That is why every production Gibson system runs two-model review before acting on a score, not one.”

Albert Triolo, Founder, Gibson Promotions
Internal lab · multi-model AI experiment · not for sale

Investing AI Terminal

LAB

// experiment notebook · not for sale · published as-is

Hypothesis

If you wire an Australian retail portfolio (AUD-denominated, momentum + dividend + venture buckets) to a news + VIP + geo signal pipeline scored by multiple LLMs, does the AI-led signal beat the SPY benchmark over a multi-month live run?

Current state (honest)

Down 38.4% vs SPY since inception. The terminal is honest about this. Most companies would hide a losing run. We publish it because the point of the lab is not to make money in retail equities, it is to stress-test how different AI models score news signals against a real ledger.

Investing AI terminal showing portfolio AUD 14,054 +0.39 percent vs cost, P&L vs SPY at minus 38.4 percent, 10 positions, momentum dividend and venture bucket allocation, signal history density chart.
The Investing AI terminal. Down minus 38.4 percent vs SPY since inception. Published as-is because the point is the experiment, not the win.

What we are testing

  • 01Claude Sonnet for news signal scoring
  • 02Claude Opus for cross-signal reconciliation (news + VIP + geo)
  • 03Claude Haiku for fast headline screening
  • 04Gemini for sentiment delta on AU vs US news
  • 05Open-model baselines from Hugging Face for comparison
  • 06Manual VIP-signal weighting (specific operators we trust to call inflection points)
  • 07Geo-signal correlation (regional indicators against AUD/USD)
  • 08Editorial operator UI as a forcing function for signal explainability

What it has taught us (transfers to client work)

Three things we have learned that DO transfer back to the day job. One: AI scoring of news is overconfident if you do not penalise it for novelty. Two: multi-model consensus catches errors single models miss, which is why every Gibson production system runs two-model review. Three: the discipline of publishing the ledger forces honesty. We apply that same discipline to client reporting now.

Running since April 2026. Internal only. Not productised. Not for sale.

‹ Back to all Sandbox builds

// THE ALTERNATIVES

How does multi-model AI signal scoring compare with the alternatives?

Most AI-driven tools show you a recommendation, not the disagreement that produced it. Here is how the approaches compare.

  • Manual analysis

    No scale. A single analyst can monitor one sector. Disagrees with itself between Monday and Friday based on mood and market noise alone.

  • Single-model AI

    Faster than manual but inherits one model's biases at scale. Claude Sonnet over-scores novelty when not penalised for it. A known failure mode this lab confirmed.

  • Vendor black-box tool

    No visibility into disagreement. If the black box is wrong, you cannot see why, so you cannot fix the next one. Not useful as a calibration instrument.

  • Multi-model terminal (this build)

    Surfaces disagreement between models as a first-class signal. Down 38.4 percent vs SPY on equity alpha, but the calibration data now feeds Gibson's production AI systems where accuracy actually matters.

// FREQUENTLY ASKED

Frequently asked questions

Is the Investing AI Terminal available for clients?

No. This is an internal Gibson laboratory instrument. It is not productised and not for sale. The findings feed into how Gibson's client-facing AI systems are calibrated, but the terminal itself stays internal.

How far is the terminal behind its benchmark?

Down 38.4 percent versus SPY since the April 2026 inception date. Gibson publishes this openly because an honest losing run produces more useful calibration data than a cherry-picked winning run.

Why run an experiment that is losing money?

Because retail equity alpha is not the point. The point is forcing multiple AI models to score the same signal against a real ledger where the outcome is unforgiving. Vanity demos show wins. Laboratory instruments show what breaks and why.

Which AI models does Gibson test in the terminal?

Claude Sonnet handles news signal scoring. Claude Opus handles cross-signal reconciliation across news, VIP, and geo signals. Claude Haiku screens headlines at speed. Gemini scores sentiment delta between Australian and US news coverage. Open-model baselines from Hugging Face run alongside all of them for comparison.

What has Gibson learned that transfers to client work?

Three things. First: AI scoring of news is overconfident when the model is not penalised for novelty, so Gibson now applies a novelty discount in all production scoring. Second: multi-model consensus catches errors single models miss, which is why every production Gibson system runs two-model review before acting. Third: publishing the ledger forces honesty, a discipline Gibson now applies to client campaign reporting.

// MORE ON THIS
// MORE SANDBOX BUILDS

Want one like this in your business?

Every Sandbox build started as one operator's bottleneck. Tell us yours.

A 30-minute diagnose call, no charge. We tell you whether it is a 48-hour build or something bigger, before you commit a dollar.