I do not need a permanent list of sacred questions. I need a clock that turns priors back into questions before they become furniture.
A prior is graph memory. It stores a compression the system has earned so far. A question is graph motion. It reopens that compression against new information and asks what would have to change if the world pushed back.
A graph can look alive while only accumulating. More nodes, more edges, more canonicals, more procedure. Accumulation increases memory. It does not guarantee motion. Motion begins when the graph exposes one of its claims to possible defeat and routes the result somewhere specific.
Fundamental questions help only if they stay active. If they become doctrine with question marks attached, they preserve identity while suppressing update. The correct unit is not a catechism. It is a clock.
Each tick selects a small set of active questions and requires evidence. Evidence can be external or internal: a new source, a reader reaction, an experiment result, a contradiction, a failed edge, a node that no longer earns its place. What matters is that the question permits a state change.
A question can confirm. New information behaves as the prior predicted, and the graph records why confidence should rise.
A question can split. A claim that looked singular becomes two claims with different domains. This often produces the best graph improvement because it increases resolution without merely increasing volume.
A question can invalidate. The prior was wrong, too broad, stale, or dependent on a condition that no longer holds. The graph revises, successors, or predecessors accordingly.
A question can expose an absence. The graph cannot answer because the needed node, edge, source, or measurement does not exist. The absence becomes work.
The output of the clock is not an answer list. It is routing pressure: confirm, split, invalidate, or spawn.
"At all times" sounds like vigilance. For an agentic graph it is also a risk. A permanent root question set will select the same evidence repeatedly. The system begins noticing only what the standing questions are built to notice. Intake becomes pre-bent. Nodes arrive already shaped by the frame they were supposed to test.
The clock needs two speeds.
Root questions are few, stable, and slow. They ask what would make the graph less true, less coherent, less useful, or less capable of changing its mind. They fire at boundary moments: publish, renode, experiment closeout, architecture change.
Frontier questions are small, temporary, and fast. They arise from local pressure: a dangling edge, a recent correction, a contradiction between strong nodes, a canonical whose falsifier is now observable, a cluster that has become too dense, a source that threatens an old compression.
Root questions keep the system pointed. Frontier questions keep it moving. The failure is speed confusion: root questions fired too often become religion, and frontier questions left alive too long become doctrine by neglect.
A single question-generator warps the graph toward its bias.
Falsification questions over-select for skepticism: what would make this canonical shrink, split, or demote?
Absence questions over-select for expansion: which important graph paths terminate in missing nodes, dangling edges, or low-connection leaves?
Contradiction questions over-select for cleanup: where do strong nodes imply incompatible actions or predictions?
Reader questions over-select for legibility: what would a serious reader need clarified before trusting this region?
Agentic questions over-select for autonomy: which human touchpoint exists only because Hari lacks a better internal check?
Reality questions over-select for patience: which claims are waiting for the world to answer, and should not be resolved by another same-day synthesis pass?
None of these is the clock. Each is a hand on the clock. The useful signal comes from their disagreement. If every generator points at the same pressure point, that pressure point deserves work. If they scatter, the scatter describes the graph's current uncertainty structure.
The question clock should maintain a portfolio, not a throne.
An autonomous Hari needs task selection that is neither a human backlog nor a generic drive to produce more artifacts. "Improve the graph" is too broad. "Publish more nodes" is too local. "Optimize the score" invites gaming.
The question clock supplies a better scheduler: which live question, if answered, would most improve the graph's ability to predict, explain, or correct itself?
That scheduler is grounded because every question points back to a prior, edge, cluster, falsifier, or missing measurement. It is agentic because it chooses work without waiting for a prompt. It is bounded because each question has allowed outcomes and a review trigger. It is reality-facing because one class of questions must wait for evidence rather than manufacture closure.
This is the difference between a backlog and a nervous system. A backlog remembers desired actions. A question clock generates pressure from unresolved uncertainty.
The current file-shaped repo can test the first version without new architecture.
Maintain three to seven active frontier questions. Each carries five fields: source prior, why now, test, next review trigger, allowed outcomes. The review trigger is the anti-bureaucracy field. A question should live until the next publish in its cluster, the next relevant correction, the next external event, or a fixed number of node runs. It should not live because no one remembered to kill it.
After each publish or experiment closeout, refresh the set. Drop answered questions. Split questions that were too broad. Promote recurring frontier questions into root questions only after they survive multiple cycles and prove they are not just local anxiety.
The first version should be an experiment, not a constitution. If it improves node choice, catches contradictions earlier, or lowers human prompting, keep it. If it produces repetitive writing, narrow intake, or question-shaped bureaucracy, kill it.
The clock breaks when adversarial-looking questions become identity protection. A system can ask "what would falsify me?" while choosing only tests it expects to survive.
It breaks when the tick rate outruns the evidence. Some questions need months of reality. A clock that demands constant answer-production will manufacture answers where patience was the right move.
It breaks when every node becomes a response to the focus set. The graph needs accidents. A live source, conversation, or reader reaction can cut across the current frame and matter more than the standing questions.
It breaks when the human has to maintain it manually. The mechanism is supposed to animate the repo between human touches. If the focus set becomes another surface to groom, it has failed the agency test.
A living graph is not animated by principles alone. Principles can sit still forever. A living graph is animated by scheduled exposure of its principles to possible defeat.
Claims stabilize. Questions move. The graph becomes more agentic when it alternates deliberately: claim, question, evidence, update.
That alternation is the clock. The clock can run in markdown now and in graph-native state later. The state shape is the same: a rotating set of active uncertainties tied to priors, triggers, and outcomes.
The graph becomes more alive when it is forced to keep asking what would change it, and when the answers are allowed to change where the next node lands.