Fundamentals
- What it reads
- Reported financials, revenue, earnings quality, forward EPS, and the assumptions behind valuation.
- What it changes
- Defines what the company has earned and which valuation cases remain defensible.
Financial research · Evidence integrity
Kairosys
A research operating system that keeps a report tied to the evidence beneath it.
Kairosys brings fundamentals, industry, macro conditions, positioning and flows, news and sentiment, and technical context into one research question. Source, period, unit, falsifiers, and the decision trail stay attached; stale or unsupported statements stop before a labeled draft reaches a human analyst.
Designed and built solo by Titus Lai.
Inspectable proof6 deterministic scenariosHUMAN ANALYST · FINAL DECISION
One question · Six financial lenses · Cross-lens controls
Can the filing period, source record, and displayed EPS support this forward P/E?
01 / 05Question opened
The work starts with the decision the research must support.
Can the filing period, source record, and displayed EPS support this forward P/E?
Cross-lens controls
Source authorityPeriod + unitFalsifierDecision trace
Question opened
Six lenses read
Claims challenged
Final prose audited
Draft admitted
Six research dimensions
Different evidence answers different parts of the same investment question.
These are six real areas of coverage, not six names for the same signal. Valuation is formed from the surviving evidence during synthesis; source checks, falsifiers, and traceability apply across every dimension.
Industry
- What it reads
- Peers, industry structure, demand drivers, competitive position, and supply-chain relationships.
- What it changes
- Shows whether company performance belongs to a durable position or a temporary industry move.
Macro
- What it reads
- Rates, inflation, currencies, liquidity, and cycle data with an explicit period and basis.
- What it changes
- Sets the economic regime in which the company, industry, and valuation must be interpreted.
Positioning & flows
- What it reads
- Institutional activity, financing balances, broker behavior, and other positioning evidence.
- What it changes
- Shows who is moving and whether market participation supports or challenges the current reading.
News & sentiment
- What it reads
- Regulatory filings, company events, earnings calls, news, and sentiment with recency attached.
- What it changes
- Separates current catalysts from repeated headlines, stale narratives, and model-made claims.
Technical
- What it reads
- Price, volume, market regime, and pattern evidence across the relevant time horizon.
- What it changes
- Tests timing and market context without overriding an unsupported business or valuation thesis.
What coverage changes
Missing evidence changes the delivery decision—not just the completeness score.
Six research lenses establish breadth once source integrity is intact. A supported, complete draft may proceed; a coverage gap requires review; an untrusted source blocks delivery.
Coverage → delivery
Six lenses present
Current evidence and meaningful coverage support an editable analyst draft.
One lens missing
Supported work remains useful, but it cannot present itself as complete research.
Source cannot be trusted
Breadth cannot rescue a statement whose source trail fails.
Five responsibilities
Different layers ask different questions before the handoff.
No layer gets to blur another layer’s failure. The human analyst keeps the final decision.
- 01Chief
What decision must this research support?
- 02Intel
Which sources are current, relevant, and traceable?
- 03Trader
Which market behavior belongs only as context?
- 04Analyst
Which financial and valuation statements do the periods support?
- 05Auditor
Do the final words still agree with the evidence and math?
Kairosys governance lineage
Soul Canon began as Kairosys’s operating constitution—and kept evolving with the system.
In 2025, I designed Soul Canon to give the Kairosys deputy persistent identity and context, research and decision rules, permission boundaries, and a record of failure. Successive Kairosys rewrites changed the implementation, but kept the same question: how can an AI system evolve without losing its governing principles or human authority?
Kairosys evolution stage · APR 2025
APR 2025
Kairosys Soul Canon
Define the boundaries before adding more intelligence.
The first Kairosys design treated identity, context, permissions, traceability, and recovery as one governing layer—not as cleanup after the model responded.
- Persistent context
- Explicit authority
- Traceable failure
- APR 2025Kairosys Soul Canon
Define the boundaries before adding more intelligence.
- MAY 2025Runnable Kairosys modules
Turn the control ideas into modules that can deny, record, and restore.
- SEP 2025Kairosys Multi-Agent
Carry the same boundaries across cooperating agents.
- FEB 2026 → NOWK_ReAct Soul Codex
Let rules evolve without letting the system authorize itself.
A Kairosys governance lineage, not one continuous codebase: successive implementations and rewrites carried the same governing problems forward.
Four Kairosys stages · Four governance shifts
The dates mark distinct Kairosys implementations; each card shows how the governance layer changed.Context survives interruption
On interruption or restart, preserve a snapshot of rules, module state, and conversation context; ask before recovery.
A restart becomes a recoverable state instead of a silent reset.Authority and failure share one boundary
Without explicit mutation authority, deny the action; when execution fails, retain the decision and trace instead of silently proceeding.
The system can explain both why it acted and why it refused.Multi-agent work keeps ownership
Separate roles and bounded handoffs carry the task, evidence, and authority forward without letting collaboration erase responsibility.
More agents do not mean less accountability.Rules can evolve without self-approval
The current Soul Codex separates human-owned principles, AI-proposable rules, and process metadata. Proposed changes remain proposals until an authorized human approves them.
The system can learn and change without quietly becoming its own authority.Two separate gates
First ask whether it is sound. Then ask whether it is useful.
Integrity can block a draft outright. Utility separately catches work that is accurate but too shallow to present as complete research.
GATE 1Integrity gate
May this draft move forward?
- PASS
- Evidence, financial state, and final wording agree.
- FAIL
- Block, withhold, or keep only clearly labeled context.
GATE 2Utility review
Is the surviving work broad enough?
- PASS
- The required research lenses are meaningfully complete.
- FAIL
- Require review; shallow work cannot present as analyst-ready.
A polished error is still an error. A sound fragment is still incomplete.
Flagship engineering case
The final report can invent confidence that the evidence never earned.
I audit the words the reader actually sees—not only the inputs that existed before generation.
Source spoof
The model writes a confident attribution with no trusted source behind it.
CONTROLRebuild the source trail from tool records and discard model-only statements.
Unsupported target
The report gives a target even though a required financial period is missing.
CONTROLRemove the target and keep only context that cannot mislead.
Displayed math conflict
The final target, forward EPS, and displayed multiple disagree.
CONTROLCheck the rendered relationship, lower confidence, and block final use until corrected.
Runnable public evidence
Six fixed cases show six different delivery decisions.
The public slice uses synthetic data and fixed outcomes so reviewers can inspect the gates without seeing the private system.
ready_reportEvidence, state, and report math agree.
DeliveryEditable analyst draftADMIT
spoofed_provenanceA model promotes its own source statement.
DeliveryReview-capped draftREVIEW
incomplete_financialsA missing quarter makes the target unsupported.
DeliveryContext-only draftCONTEXT
contradictory_financialsFinancial facts conflict.
DeliveryWithheldWITHHOLD
rendered_math_conflictFinal wording contradicts displayed math.
DeliveryDraft retained; final use blockedBLOCK
shallow_but_soundSupported facts cover only two lenses.
DeliveryReview required for completenessREVIEW
Evidence boundary
The behavior is public. The private research system is not.
The linked repository is a curated clean-room slice with synthetic inputs. It demonstrates admission and report-audit behavior; it is not the full system or production software.
Runnable cases, architecture notes, fixed examples, tests, diagrams, and explicit boundaries.
Production coordination, proprietary prompts, providers, live data, operating history, and commercial source.
All Rights Reserved · Portfolio evaluation only.Boss0320 / kairosys-agentic-research
Inspect the public repository