AI LAB · FIRST FIELD APPLICATION
ARC01.5
Continuous research.
Explicit authority.
ARC is the financial-research application within Reinventy AI Lab. It brings market observation, quantitative analysis and user-scoped controls into one environment. IDRA is its multi-strategy research engine: several ways to investigate the same market, with evidence required before a hypothesis can advance.
Research software, not investment advice. No guaranteed returns. Financial activation remains a separate, personal action.
01 / THE APPLICATION
One environment. Clear responsibilities.
ARC
The applicationMarket data, research and an isolated environment for each user. Financial permissions and activation remain personal and explicit.
IDRA
The AI research agentFour quantitative strategy families that can branch into new candidate specialisations, with a traceable history of experiments.
Nemotron
The generation modelProposes structured research variants. It does not hold wallets or authorise financial execution.
02 / INSIDE ARC · IDRA
Four heads.
More perspectives.
Meet IDRA, ARC’s AI research agent. Four specialised heads investigate the market; Nemotron helps generate new candidate branches. One architecture connects their ideas to measurable experiments.
Conceptual representation. The heads represent research families, not physical hardware.
IDRA · THE AI RESEARCH AGENT OF ARC
Intelligence that multiplies perspectives.
IDRA is ARC’s AI research agent. Like the Hydra that inspired its name, it approaches the same market through several specialised heads and, with Nemotron, can generate new variants to put to the test.
Each head represents a different analytical perspective. One investigates the direction of a movement; another studies recovery after a decline; a third explores oscillations; a fourth examines changes of regime. These are not four decorative labels or four independent wallets. They are four research families within a shared architecture, designed to ask different questions of market data.
Calling IDRA an agent means that its role extends beyond a conversational answer: it connects structured proposal generation, validation and repeatable experiments. The language model helps explore new hypotheses; quantitative rules make those hypotheses measurable. The aim is to build a research process that can broaden its analytical repertoire without losing the ability to explain what was tested and why.
The Hydra metaphor becomes an engineering principle: specialisation can branch. A base head can give rise to more focused candidates, each with declared parameters and its own experimental record. More heads do not automatically mean better decisions. They mean more perspectives that must earn their place through evidence.
One agent. Several analytical perspectives. A shared discipline of evidence.
THE FOUR FOUNDING HEADS
One market. Four ways to investigate it.
A market can move directionally, recover after a decline, fluctuate within a range or enter a new regime. IDRA investigates these different hypotheses rather than treating every movement as the same opportunity. Its strength as a research architecture is this plurality: the same data can support questions about continuity, recovery, equilibrium or change.
TREND
Recognise direction, rather than chase every movement.When a market takes a direction, the question is not only how far it has moved, but whether that movement shows continuity. The Trend head studies persistence within the observation window and forms hypotheses about continuation. It seeks to distinguish directional behaviour from a temporary fluctuation. Its specialisations can explore different windows and thresholds, comparing faster or more selective readings of the same phenomenon.
DIP RECOVERY
Look for recovery, rather than guess the bottom.A decline is not automatically an opportunity: it can precede a recovery or a further fall. Dip Recovery studies recovery conditions after a decline, seeking evidence beyond the observation that the price has fallen. Its role is to turn the intuition of a rebound into a testable question. It does not assume that every loss will be recovered or claim to know the lowest price in advance.
RANGE REVERSION
Make oscillations a subject of research.Not every market follows a sustained direction. Prices can fluctuate around a reference range. Range Reversion studies deviations and potential return conditions, offering a different perspective from trend analysis. The range is a hypothesis to test, not a boundary the market must respect: when behaviour changes, the interpretation must be questioned too.
BREAKOUT
Investigate the points where behaviour changes.A departure from a previous range may signal a new dynamic, but it may also fade quickly. Breakout examines these transitions and tests whether the change persists. Where other heads seek continuity or a return, this one explores discontinuity. Comparing variants helps study how more sensitive or more selective readings affect results, without treating every break as a confirmed opportunity.
03 / CONTROLLED SPECIALISATION
New heads are hypotheses.
Not shortcuts.
Nemotron can propose up to two additional candidate specialisations for each base family: at most eight candidates. Each varies permitted parameters inside a defined research space. IDRA does not create unrestricted code or independently expand its authority.
4 base families → up to 8 additional candidates. A research candidate is not an authorised trading agent.
NVIDIA NEMOTRON · GENERATIVE RESEARCH
Not just analysing. Proposing new ways to analyse.
Nemotron gives IDRA a generative dimension. Instead of relying exclusively on variants chosen manually in advance, the system can request structured proposals from the model within a permitted family of strategies. The result is a candidate configuration that the application can inspect, validate and compare—not a market opinion to accept on trust.
For example, a Trend head can be explored through two candidate variants using different permitted observation windows or thresholds. The question becomes concrete: does the change improve the observed balance between selectivity, costs and drawdown, or merely make the strategy look better on a particular sample? This example illustrates the research mechanism; it is not a claim that any specific variant has demonstrated an advantage.
The separation of responsibilities matters. Nemotron proposes; the application enforces the format and boundaries; the quantitative engine evaluates. The model does not decide that its own output is valid. It cannot turn an experimental result into a financial permission, and it does not obtain wallet keys or the authority to sign transactions.
The AI proposes. The engine measures. Controls define what can advance.
04 / FROM GENERATION TO EVIDENCE
Every new head has to earn its place.
Generating more variants is only the beginning. IDRA preserves the relationship between the base family, the proposed change and the outcome of the experiment. A promising narrative is not enough: the candidate must be evaluable, comparable and open to rejection. Recording unsuccessful attempts matters as much as recording encouraging ones, because it prevents the research story from being built only around selected successes.
- 01
Generation
Nemotron proposes a structured candidate in the permitted format.
- 02
Validation
The application checks permitted parameters and research boundaries.
- 03
Comparison
Replays compare variants with their base family, including modelled costs and drawdown.
- 04
Review
Rejected candidates and insufficient evidence remain traceable. No automatic promotion to live trading.
PREDICTIVE RESEARCH
Explore possible scenarios. Measure what follows.
For IDRA, predictive research means formulating testable hypotheses about subsequent market behaviour. It does not mean knowing the future. A directional hypothesis, a recovery hypothesis and a range hypothesis can lead to different expectations; their usefulness depends on how they perform on data that did not determine their design.
The development path is therefore as important as the model: comparisons with simple baselines, evaluation on independent periods, realistic costs and scrutiny of performance when the market changes. These are the conditions for assessing whether additional specialisations contribute a durable advantage. They remain validation requirements, not results implied by the presence of an LLM.
05 / EVIDENCE AND LIMITS
A distinctive architecture. A testable ambition.
IDRA connects quantitative strategy families, model-generated variants and controlled evaluation. This combination is the focus of our research—not a claim that the system knows future prices, market bottoms or market tops.
Implemented research capabilities include structured generation with Nemotron, bounded candidate validation, experiment tracking and replay. These technical results do not demonstrate a profitable predictive edge.
Predictive performance requires independent data and realistic execution costs. User availability and financial execution have separate qualification and permission gates.