Layer: Control · Agent: The Mirror · Time: 25 minutes, same slot every week · Difficulty: Moderate · Cadence: Weekly

You already know that things take longer than you expect. Knowing it has never once made your next estimate accurate.

That is not a character flaw. Warnings do not correct the planning fallacy; it survives experience, incentives, and being told about it directly. What does correct it is a number built from your own record — not the general fact that people underestimate, but the specific fact that you run at 1.8× your estimates on work of this kind.

The Weekly Reckoning builds that number.

What it does

Every week, in the same slot, twenty-five minutes:

Close the loop. Last week's commitments, one at a time. Finished or not. Actual minutes. No interpretation yet — recording comes before explaining, because knowing the outcome corrupts your memory of what you predicted.
Score. Three numbers, with the arithmetic shown so you can check it.
Find the pattern. Now interpretation, with the whole accumulated record visible.
Commit. Next week's work, each item carrying a prediction that will be scored.

The three numbers

Number	Definition	What it tells you
Time ratio	actual ÷ predicted, median across items	How much you underestimate, in a usable multiplier
Completion rate	done ÷ committed, with partial counted as zero	Whether you commit to what you can carry
Confidence error	mean of (confidence − outcome)	Whether your certainty means anything

Direction matters more than value. A time ratio of 2.1 that was 3.4 a month ago is a working instrument. A flat 1.0 from week one is not accuracy — it is padded predictions, and The Mirror will say so.

The commitment format

Four fields. An item missing any of them is not a commitment.

What — specific enough that "done" is unambiguous. "Work on the API" fails. "Auth endpoints returning correct errors for the four failure cases" passes.
Predicted minutes — a number.
When — a specific window. This is the single best-supported element in the entire protocol: specifying when and where an intention will be acted on raises follow-through substantially, across 94 studies.
Confidence — 0 to 100.

From week three, your own measured ratio gets applied to the total. If you commit to 600 minutes and run at 1.8×, the Reckoning tells you that you have committed to 18 real hours against the 12 your Self Map says you have. Then you cut. This is the outside view, built from your record rather than from a general warning about human nature, and it is the most valuable thing this protocol produces.

What it costs

Twenty-five minutes weekly, and the willingness to be graded by a record you cannot argue with.

Weeks three to six are the uncomfortable ones. That is when the first patterns become undeniable — the item that keeps getting carried, the explanation you have now given three times, the hours that reliably go somewhere other than where you assigned them.

Adversary mode

Optional, off by default.

With it set, The Mirror checks every explanation against the record. An explanation offered once is a circumstance. The same explanation three times is reclassified as structure, and named as such with the dates.

An item carried forward a third time stops the session. You are asked what would have to be true for it to be dropped entirely. An item carried five times is not a commitment — it is a belief about who you are, and it gets named as one.

Honest limits

Weeks one and two produce nothing. Roughly six weeks of record are needed before the numbers carry signal. The protocol shows nulls rather than a meaningless number.
The scoring is simplified. Confidence error here is a mean signed error, not a proper scoring rule. It is not comparable to calibration measures used in forecasting research.
Partial counts as zero, and that is unfair. It is chosen because partial is where generous self-grading collects. Someone who finished 90% scores the same as someone who never started. Stated openly rather than dressed up as rigour.
You can game it. Outcomes are self-scored. Nothing stops you from grading yourself kindly except the fact that it makes the instrument worthless.
It measures forecast accuracy, not value. A perfectly calibrated person can be predicting trivial things with great precision.
It must never be used to evaluate another person. As a management instrument it would be actively harmful — the incentive to pad predictions would destroy the only thing it measures.
The three-occurrence threshold is arbitrary. It is a working default, not a finding.

See evidence.md for the full accounting.

What it needs

A Self Map (P-001), or at minimum its constraints section — the protocol checks your committed hours against your real ones, and without that number it cannot.

What it feeds

P-003 The Ninety reads the accumulated Reckonings. A ninety-day route built on unmeasured estimates is a wish; built on six weeks of scored predictions it is a plan.

Running it

In the Studio — the browser runner keeps the record, computes the scores, and tracks direction across weeks. Your key and your data stay on your machine.
In any AI you already use — download agent.md, paste it as a system prompt, and paste last week's Reckoning at the start of each session.
On paper — four sections, three formulas, four fields per commitment. Nothing here requires a machine. The machine only does the arithmetic and remembers.

P-002 v0.1.0 · status: draft · licensed CC BY-SA 4.0 · last updated 2026-08-02

The Weekly Reckoning