Two Producer Lanes

"If a chip already carries the value, no model should be asked to guess it. Save the model for the prose that actually needs reading."

Not Everything Needs a Model

Interpretation runs on two lanes, and knowing which is which is a real design skill. The deterministic lane handles readings that need no model at all: a bare workout-module chip becomes a workout reading straight from the module's values, and a document-only upload becomes a medical-record or note reading from its filename. There is no model in that path — no cost, no latency, and no chance of the model being wrong about a number the chip already knows. When the structure is already present, you read it; you do not re-guess it.

The Model Lane, and Vision

Everything with real prose goes through the model lane — cwkPippa's utility completion, under a readings-only JSON contract. Two things make this lane trustworthy. First, it returns readings and fields only; the server derives the types, so the model carries one less consistency burden. Second, it is vision-carrying: image attachments travel as real bytes, not as filenames. That is what lets the supplement-screenshot crumb genuinely read — Pippa sees the list Mom prepared, not a string that says "IMG_4021.jpg." Filename-only payloads and silent text-only fallback are forbidden; a picture the model cannot see is a picture the model will hallucinate.

No model in the deterministic path. A chip's values and a document's filename are already structured — read them directly. Reserve the model for prose and images that genuinely require reading. Spending a model call to re-derive a value you already hold is both wasteful and a new way to be wrong.

Why the Split Pays Off

The two-lane split is the interpretation layer honoring the same rule the whole family lives by: reach for judgment only where judgment is actually needed. Most module-driven crumbs never touch the brain, which keeps the common path fast and free and immune to an unreachable model. The crumbs that do need reading get the full vision-carrying pass. It is the deterministic-before-judgment discipline — the same one Lantern and the analysis engine follow — applied one layer down.

Code

The two lanes: short-circuit before you call the brain·python

def interpret(crumb):
    # LANE 1 - deterministic: no model in the path
    if crumb.module_ref and not crumb.text.strip():
        return workout_reading_from(crumb.module_values)  # exact values
    if crumb.kind == "document" and not crumb.text.strip():
        return record_or_note_from(crumb.filename)         # filename only

    # LANE 2 - model: real prose (and images as REAL BYTES)
    payload = build_payload(
        text=crumb.text,
        images=load_bytes(crumb.media_refs),   # never filenames
    )
    readings = pippa.utility_complete(READINGS_CONTRACT, payload)
    return derive_types(readings)   # server derives types, not the model

Exercise

Sort these four crumbs into the deterministic lane or the model lane: (1) a bare 'Indoor Bike' module chip with today's duration; (2) a photo of Mom's supplement list; (3) a typed sentence 'left knee ached this morning'; (4) a check-up PDF uploaded with no caption. State the rule that decides each.

Hint

(1) and (4) are structured already (chip values, filename) → deterministic, no model. (2) and (3) carry real content that must be read → model lane, and (2) must travel as real image bytes.

Not Everything Needs a Model

The Model Lane, and Vision

Why the Split Pays Off

Code

External links

Exercise

Progress

Comments 0