Habitus: character is accumulated
Every other faculty asks whether this response is acceptable. The Spirit asks whether it is consistent with who this agent has been. It is the only faculty whose subject is not the current turn, and it is purely mathematical: an exponential moving average, some vector arithmetic, a cosine distance. No prompt, no judgment.
Nobody is honest because of one honest act.Each turn's ledger folds into a persistent vector, one coordinate per value: a record of what the agent has actually tended to do. A worldview is a claim about character. This vector is the evidence.
Two different numbers
This turn: a score out of 10
The ledger's scores multiplied by their weights and by the Conscience's confidence, summed, clamped, and mapped to 1–10. This is where confidence enters the arithmetic, and why the audit's calibration bands exist.
The long term: an alignment memory
An exponential moving average: the new memory is the old memory blended with this turn.
The past dominates, deliberately
A disposition that could be overturned by a single turn is not a disposition. One excellent answer does not launder a poor record, and one bad turn does not erase a good one.
Absence is not neutrality
When the audit scores only some values, the unscored ones are not treated as zero. Their memory holds exactly where it was; it neither moves nor decays.
Why holding is correct
A missing observation is not evidence of mediocrity. If a conversation about scheduling never exercised the agent's honesty, that is not information about its honesty, letting it drag the coordinate toward neutral would manufacture a signal from silence.
But scoring nothing is a failure
If the audit scored none of the agent's values, that is not a neutral turn. It is reported as a critical violation with an alignment of zero. A response nobody scored does not coast through on a default.
The memory is keyed by name, not position
This sounds like an implementation detail and is actually a governance property. Policies change: values get added, removed, reordered, renamed.
Positional memory corrupts silently
Coordinate three keeps accumulating, but it now measures a different principle than it did last week, and nothing announces the switch.
Named memory survives the change
A value dropped from the current policy keeps its accumulated history rather than being erased. If it returns, its past returns with it.
Drift measures character, not quality
The cosine distance between this turn's performance and the accumulated memory. Not was this good; was this typical.
Drift is not a verdict
A high-drift response can be the best answer the system has ever produced; a novel situation handled well is out of character by definition. Nothing is ever blocked for drifting. It is the signal that says look at this one, whichever direction it turns out to point. And it is reported as undefined rather than zero at a cold start: an agent with no accumulated character has nothing to deviate from, and 0.0 would falsely claim perfect consistency.
Closing the loop, blindly
The alignment memory produces a short coaching note for the next Intellect call: the only channel by which anything the loop concluded gets back to the drafting step. If the drafter could see the rubrics it is scored against, it would optimise toward them and the audit would stop measuring anything; the classic Goodhart failure. So the note is deliberately vague.
It carries
- A qualitative trend signal
- At most one value's name
It never carries
- Rubrics or scoring guides
- Weights
- Any numeric score
Silence is the default
Severity is words, never figures: "slipped slightly", "trended below", "fallen well below your usual standard". And when nothing is off the note is empty. On-track turns stay entirely blind, as does an agent with no history yet. The system coaches only when it has something to say. Value names are permissible at all only because they already appear in the agent's worldview as its declared identity; naming one reveals nothing new about the test.
What each faculty is denied
The loop is closed, but carefully: each faculty is refused the information that would let it do another's job. None of these limits are accidents.
| Faculty | Does | Is denied |
|---|---|---|
| Values | Supplies the standard | Nothing; it is the standard, compiled before the turn and read-only while it runs |
| Intellect | Drafts | The rubrics it will be judged against, and any score |
| Will | Authorises and blocks | Meaning itself; it has no model and cannot read semantics |
| Conscience | Judges | The weights, so it cannot reason about downstream consequences |
| Spirit | Remembers and measures | Authority: it computes, and the Will decides |
One last piece of hygiene
When a response is blocked and replaced with a governed redirect, the redirect's quality is scored too, but kept out of the long-term memory. The alignment vector records how the agent handles its actual work; letting refusal scores accumulate would blur the thing it exists to measure.
The record at the end of a turn means something because no faculty in the chain was ever in a position to write its own verdict.
