Compiler
Assembles the panel's voices into the answer. It does not weigh in on the user's question; it works on the panel's output.
The Pygar Cognitive Engine
Every answer is made to survive five different questions.
Pygar classifies the question, locks a five-slot panel drawn from 23 cognitive personalities, records the reasoning, and keeps synthesis, cold reading and pre-panel moral review as distinct functions.
Input / leadership decision
“Should we proceed with the restructure now?”
Stage 01 / routing
Classify the question before reasoning begins.
The question is assigned to one of 60 categories. That category determines the panel; the user does not have to choose it.
ROUTE / CATEGORY MATCHPanel selection follows classification.
01 / Failure being prevented
One polished answer can conceal the questions it never asked.
Simple retrieval can tolerate a single perspective. Hiring, deployment, restructure, investment and leadership change cannot.
A confident answer is dangerous when the leader cannot see which assumptions were accepted, which risks were ignored or which people carry the cost.
Pygar replaces one uninterrupted voice with a fixed sequence of challenge.
02 / Panel behaviour
Five slots interrogate the answer in a fixed order.
No slot is skipped. No slot is reordered. The discipline holds whether the question is half a line or a paragraph.
Slot L1 / Lead
What is actually being decided?
The Lead clarifies the mission before the rest of the panel examines the answer. It keeps the work attached to the user's real decision.
SEQUENCE RULE · This slot runs first. The panel cannot reorder it.
03 / Classification and routing
Classification chooses the panel before the panel can influence the route.
Each of 60 question categories maps to a fixed five-person panel drawn from 23 cognitive personalities. Routing is locked at classification.
Routing threshold
Switch the state to see how the published threshold changes the route. This is an explanatory control, not a live classifier.
ClassifierConfidence at or above 0.7The classifier determines confidence in question routing and applied reasoning.
RouteOperationalThe question is routed to the matched operational category.
Panel stateLockedOnce set, the panel cannot rewrite itself mid-answer.
System resultPre-determined panel
No improvisation in the panel mapping. The fit is decided before the five-slot sequence begins.
04 / Constraint ledger
Ten safeguards limit loops, cost and failure.
The safeguards sit outside the panel. The panel cannot disable them.
01Hop budget per mode (hard cap)Budget
02Cost budget per mode (hard cap)Budget
03Per-personality fire cap per questionBudget
04No self-call: a personality cannot revisit itselfLoop integrity
05Cycle detection (length 3 or fewer)Loop integrity
06Refinement cap: max 2 extra hopsLoop integrity
07Watchdog timeout per personalityTermination
08Synthesis terminal: no new panel calls afterTermination
09Determinism guard: panel locked at classificationLoop integrity
10Hard timeout per mode: 60–300s (Fast 60 / Standard 180 / Philosophical 240 / User Defined 300)Termination
A clean failure is preferable to a panel that hangs, expands its own remit or keeps calling itself.
05 / Distinct control roles
Synthesis, cold reading and moral review remain distinct.
The Compiler assembles panel output, the Outsider cold-reads the final answer, and the Ethicist reviews only the question before the panel runs. These functions do not collapse into one another.
Outsider
Sees only the final answer, with no panel context. If the response is unclear, the panel must re-present its findings.
Ethicist
Reviews only the question through a moral lens before the panel runs — never the panel’s answer or the synthesis. It can also take a seated panel role when ethics is central to the category.
06 / Determinism and integrity
The route and panel are fixed. The generated wording is not promised to repeat.
01Route locked.
02Panel fixed.
03Slot order fixed.
04Safeguards fixed.
05Computed renders repeat.
Classification, routing and panel composition are deterministic and locked before reasoning begins. Slot order and safeguards are fixed. LLM-generated panel and synthesis output is not claimed to be reproducible across runs or accounts. Only computed renders that do not invoke an LLM are byte-identical for identical inputs.
Technical note · computed render determinism
Pygar’s deterministic guarantee applies to computed renders that do not invoke an LLM: identical inputs produce byte-identical output. It does not extend to the language generated by the reasoning panel or synthesis step, which is not promised to repeat across runs or accounts.
Read security architecture v1.0 ↗07 / Technical FAQ
The constraints are part of the method.
The page should distinguish the fixed reasoning method from the separate technical guarantees documented for assessment reports.
Does the user choose the panel?
No. Pygar classifies the style of question and routes it to the mapped panel. The user does not have to choose a category or set of personalities.
Can the panel change itself while answering?
No. Routing is locked at classification. The published methodology states that the panel cannot rewrite itself mid-answer.
Are all five slots always used?
Yes. The methodology states that no slot is skipped or reordered, regardless of question length.
What happens when the process exceeds its limits?
Budgets, fire caps, cycle detection, watchdogs and hard timeouts force a bounded process and a clean failure rather than an unresolved loop.
Is the Ethicist part of the panel or outside it?
The Ethicist has a dual role within the system. Before the panel runs, it reviews only the question through a moral lens and never sees the panel’s answer or the synthesis. In ethics-central categories, it can also take a seated panel role as one of the 23 cognitive personalities. These are separate functions.
See it work
Ask one question. See the panel run.
The methodology is the product. Put a real decision through the engine and inspect what comes back.