// DELIBERATIVE ENGINE
Casandra thinks. The products execute.
Casandra is the deliberative cognitive engine of the Ágora ecosystem. It is not the database. It is not the identity layer. It is not the shared memory. It is the engine that reasons when an ecosystem system needs deep analysis. Not all systems need to invoke Casandra all the time, but the products that require deep thinking, like Sibila and Némesis, rely on Casandra to produce more solid analysis.
The shared memory remembers.
// FOUR AGENTS, ONE VERDICT
Four agents deliberate before each answer.
Verifier, challenger, synthesizer and director. Each with a distinct role, each with a distinct model. If the four don't reach consensus, the system says 'I'm not certain'. That property, knowing when it doesn't know, is the difference from a wrapper over Claude.
// WHAT IT SOLVES
Three pains it addresses.
AI that answers with false certainty
LLMs trained to please give confident answers on topics where they shouldn't have certainty. That convincing fluency with weak backing is what is called hallucination. Casandra solves the problem from the root: every conclusion is attacked before it leaves.
Analysis without explicit confidence level
An answer that doesn't come with its confidence level is a half-answer. Casandra delivers each dictum with a unique, calculated and auditable confidence level. The user knows if they can act without hesitation or if they need more data.
Internal contradictions that go unnoticed
A sophisticated analysis can contradict itself internally without anyone noticing. Casandra is designed to detect internal contradictions, hidden assumptions and logical leaps before the result reaches the user.
// CORE FUNCTIONS
What it does concretely.
Deliberation between specialized agents
Multiple agents with differentiated profiles (analytical, skeptical, contextual, technical) deliberate in parallel on the same question and contrast their conclusions.
Active refutation before delivery
A refuting agent attacks each conclusion before it reaches the user. If the refutation succeeds, it forces a new round. What reaches the user survived the attack.
Internal contradiction detection
Cross logical analysis to identify contradictions between assumptions, data and conclusions within the same dictum. What is not consistent is not delivered.
Real confidence estimation
Each dictum arrives with a unique confidence level, calculated from source quality, logical chain robustness and hypothesis solidity.
Explicit and auditable reasoning
Casandra does not deliver black-box conclusions. Each dictum comes with the explicit reasoning chain: hypothesis, evidence, refutations considered, refuter's counterweight.
Avoids flattering responses
By design, Casandra is not oriented to please the user. If the honest conclusion is contrary to what the user wanted to hear, that is the conclusion it delivers.
// WHO IT IS FOR
Target audience.
- ›It is Sibila's internal engine for producing deep dossiers
- ›It is Némesis' internal engine for critical decision audits
- ›It can be invoked by any ecosystem system requiring deep analysis
- ›It is not sold or licensed separately, it is part of Ágora's infrastructure
// FAQ
Frequently asked questions about Casandra
Can Casandra be hired separately? +
No. Casandra is the deliberative cognitive engine of the Ágora ecosystem, not an independent commercial product. Casandra is accessed through the systems that invoke it, mainly Sibila and Némesis. It is not licensed or sold as an isolated engine.
What AI models does Casandra orchestrate? +
Casandra orchestrates commercial models from Anthropic (Claude), OpenAI (GPT) and Google (Gemini) in parallel according to the task. Each specialized tribunal agent can use a different model based on its role, and the deliberation between them is what produces the final dictum.
How is Casandra different from a simple wrapper over Claude or GPT? +
A wrapper makes one call to a model and returns the response. Casandra makes multiple parallel calls to agents with different profiles, makes them deliberate with each other, actively attacks conclusions with a refuting agent, measures confidence level and discards what doesn't survive the attack. It is cognitive architecture, not direct API consumption.
// THE EIGHT CAPABILITIES
What makes Casandra categorically different.
Every architectural decision in Casandra responds to these eight capabilities. If a technical decision violates them, the decision is wrong even if it seems to save time or money.
Different team for different problem
Casandra assembles a different team of specialists depending on what you ask. For Sibila it calls commercial and industry profiles. For Nemesis it calls CEO, CFO, COO and field specialists. For Polaris it calls predictive profiles. Same engine underneath, the deliberation team changes by problem. A single team cannot be good at everything.
Memory between analyses
Persistent memory that identifies recurring error patterns, injects them as context at the start of each new analysis, and grows sharper with each case. If Casandra recommended something on a similar company and the deal was lost for a specific reason, next time a similar case appears it takes that into account. No competitor in the commercial space has this.
Binding audit, not decorative
The auditor agent warning is binding. If it detects a critical problem, the engine does not export the dossier until the problem is resolved, explicitly marked, or deactivated by conscious human decision. There is no way to ignore the auditor. An audit you can ignore is not an audit.
Honesty with foundation
Every confidence number comes with the formula that produced it. "Confidence 73 percent, we started at 100, took 10 off because there is an unresolved internal contradiction, added 3 because the team reached consensus in the first round, took 20 off because we could not verify two critical facts". Anyone can request the breakdown. A sophisticated client never trusts a number they cannot audit.
One true datum per concept
For each important concept (decision-maker, company, founding year, address, score) there is a single canonical datum read by every place in the document. If the canonical decision-maker changes, the whole document updates automatically. By design it is impossible for page 4 to say one thing and page 27 say another.
Client memory, not just analysis memory
When Casandra lives inside the Agora ecosystem, it sees the full client history. Omnira conversations, Cosmos orders, Polaris alerts, Nemesis verdicts, all available for the next analysis. Intelligence becomes dramatically deeper with real operational history. This benefit grows over time.
Alternative scenarios, not one conclusion
Casandra delivers multiple scenarios with estimated probability. "Most likely interpretation is company in active expansion, 62 percent. But the evidence is also consistent with mature company with loyal customer base, 24 percent. These are the signals that tell them apart". The client makes better decisions when seeing alternative scenarios.
Learning from real outcomes
It registers outcomes at 30, 90 and 180 days. The user records what happened (won, lost, dropped, reason). That information feeds the engine memory and calibrates future analyses. Without an outcome loop, the engine does not improve. With a closed loop, it improves with each case that passes through it.
// THE PIECE THAT CLOSES EVERYTHING
Coherence Gate, coherence check before export.
When the analysis is ready and about to be delivered to the client, Casandra does a final pass before exporting the PDF. It reads the whole document looking for internal contradictions, that the cover and the text match, that the canonical datum appears in every place, that the auditor warnings appear in the recommendations, that no homonym-contaminated profile is used. If it finds contradictions, it does not export. It returns to deliberation or asks for human intervention. It is the difference between AI that sometimes makes mistakes and AI that controls itself before speaking.