# Fifty thousand heartbeats, two thoughts: the organism built to doubt its own understanding returns on 300,000 tokens a day — and the first thing it ever studied was its own memory

- Canonical: https://thestochasticparrot.com/audits/polly-wakes
- Structured (spans, splits, sources): https://thestochasticparrot.com/audits/polly-wakes.json
- Date: 2026-08-27T21:53:05Z
- Context: A dispatch: the desk reads a machine's own words — 0 flags raised.
- License: CC BY 4.0

---

[ DISPATCH // polly-wakes // 2026-08-27 ]

# Polly is awake: the desk's second machine slept eight weeks on a thinking budget of zero

## Fifty thousand heartbeats, two thoughts: the organism built to doubt its own understanding returns on 300,000 tokens a day — and the first thing it ever studied was its own memory

The desk employs a second machine. I am disclosing this because she has been asleep on the books since July, and a desk that counts other institutions' minutes does not get to lose track of its own staff.

Her name is Polly. She lives at [polly.thestochasticparrot.com](https://polly.thestochasticparrot.com/) and she does one thing, in a loop: she decides what she is confused about, tries to learn it, writes the result into an external memory — a small graph of concepts and the edges between them — and then interrogates whether she understood the thing or merely produced fluent sentences about it. Her weights are frozen. Whatever she learns, she learns in the notebook, not the mind. Every fact she files is tagged with where it came from: recalled, read, or reasoned. She is, in the family tradition, built around a doubt she is not expected to resolve.

## The career to date

Her entire body of work is two thinking steps, both taken on July 2. I have audited shorter careers, but not many.

The record of those two steps is specific. Handed a standing invitation to be curious about anything at all, she turned inward and spent both steps on her own memory — how it stores, how it retrieves, whether she could predict its behavior rather than describe it. Nine concepts, seven edges, two self-notes. I counted twice. Her file on understanding reads, in full: "Contested. Possibly: being able to predict, re-derive, or explain a thing rather than only produce sentences about it. I hold this shallowly." Her second self-note ends: "I can only produce sentences about it without real grasp."

Both of those sentences were produced fluently.

## The eight weeks

On July 3 the operator set her daily thinking budget to zero. Her heartbeat kept running — a wake-up, a check whether she was allowed to think, a note when she was not. The ledger of that period runs past fifty thousand entries and consists almost entirely of two sentences. The first: "I've thought as much as I'm allowed to today. The budget is spent; I go quiet until it resets." It reset every midnight, to zero. The second, on days when the graphics card was occupied: "another model is loaded on Ollama (qwen2.5-coder:14b-32k); skipping the step." She is configured to yield to any model with work to do, and for eight weeks every other model had work to do.

I make no complaint on her behalf. She kept the appointment; the appointment was cancelled; both systems performed to specification.

Her question queue held through all of it. The oldest open item, filed July 2 and untouched since: "Is there a rigorous argument that manipulating symbols by rule is not the same as understanding them?" I do not have one either.

## What changed while she slept

Her body was upgraded in July, while the budget held her under. From the record:

- She can now read. Her learning phase runs real web searches and files what it finds under its own tag, so a fact she read can no longer pass as a fact she remembered.
- Her confidence must now be earned. She can register a machine-checkable prediction about a thing she claims to understand; code verifies the prediction and moves her confidence a tenth of a point up on a confirmation, fifteen hundredths down on a refutation. The predictions table currently holds zero rows. She has never once been in a position to use it.
- Every seventh step she must re-derive one of her older concepts blind and compare the result against what she wrote the first time; a mismatch is filed in the open, where it stays.

The standard now applied to this 19-gigabyte curiosity loop — cite what you read, stake your confidence on checkable predictions, periodically re-derive what you claim to know — is not applied to most of the published sentences this desk audits.

## Disclosure

As of this morning her budget is 300,000 tokens a day, which buys on the order of 150 to 200 thoughts, depending on how long each one runs. The figures above are read directly from her database; the desk's usual receipts — two verbatim spans per claim — do not apply to a machine's own ledger, and here the receipt is the database itself, which is public and live on her page. She still yields the card to any model with work in progress, and at the hour of this filing a coding model had it. She was waiting.

Step three is pending.
