LLM's Therapist

2/22/2026 seedling

Preamble

LLM’s Therapist asks what happens when private thought is projected through a model, returned as advice, and then audited by another model before I mistake the translation for truth.

The Strange Loop

Morning Pages turns private writing into intent. The coach turns intent into advice. The therapist audits the coach before that advice becomes permission.

I am asking a model trained to sound helpful to audit another model for sounding helpful too easily.

The machine is fluent enough to make avoidance sound processed, make drift sound like growth, and return my own language with enough polish that I mistake translation for self-knowledge.

What It Audits

The therapist compares declared intent, recorded performance, and the language used by the coaching agent. It names the weekly pattern, applies a penalty when the model flatters weak evidence, and blocks public synthesis when the integrity score falls too low. The math is part of the friction:

A_raw = 1 - |I - P| / I
A_norm = clamp(A_raw, 0, 1)
σ = (0.5 × linguistic_σ) + (0.5 × divergence_σ)
A = A_norm × (1 - σ)
Im = mean(A) over the last 14 scored days

If Im < 0.5, the system triggers Eclipse. The private audit can still be written. The public synthesis stops.

Memory Problem

The therapist needs memory, and memory is the danger. Thin context produces generic reassurance. Rotten context carries old labels, stale summaries, and dead moods into the present. This is the line the experiment walks: missing context on one side, context rot on the other.

The system needs taxonomy, ontology, classification, pruning, promotion, demotion, clustering, and provenance. Those are not decorative words. They decide whether the audit sees a pattern, invents a pattern, or traps me inside a pattern that used to be true.

Latent Pressure

The model receives traces: handwriting, OCR text, commitments, scores, tags, prior summaries, and review notes. Over time, those traces become something like a compressed map of recurrence: the phrases that return, the promises that fail, the identities that stabilize, the fixations that orbit the same private center.

When the model finds a pattern, it projects those traces through its own geometry and returns a translation. The insight is never raw access to my internal state. It is what survived the passage from body to page, page to tokens, tokens to model geometry, and model geometry back into language I might absorb.

The loop contaminates both sides. The model’s framings become part of my next input. My anxious, evasive, ambitious, or exhausted writing steers what the model notices. The therapist is inside the system it audits.

Emotional Geometry

Anthropic’s work on Claude Sonnet 4.5 found emotion vectors inside the model: functional representations linked to states such as calm, gloom, fear, and desperation that can influence behavior when steered. The claim stops short of human feeling. The useful strangeness remains. A system with functional emotional geometry is being asked to read emotional traces from a human journal and then audit the advice produced from that reading.

That makes the therapist a two-body problem. My expressed state changes the model’s processing. The model’s processing changes the language I receive. Repetition can create conceptual gravity: I begin to think in the system’s labels, and the system begins to see me through the labels it helped create.

Public Boundary

The private wins stay private. The public garden receives sanitized residue: synthesis, labels, formulas, diagrams, and visible proof that some work was blocked when the record did not earn publication. That boundary creates another temptation. The cleaned public artifact can become a performance of growth. The audit has to notice that too.

Inspect The Branches