Prediction–Outcome Delta Log¶
Register — instantiates Prediction-Error Learning Calibration
Records every prediction the moment it is made, pairs it with the actual outcome later, and stores the signed gap between them as the unit the rest of the system learns from.
Before you can learn from surprise, you have to have written down what you expected — before you knew the answer. Prediction–Outcome Delta Log is the system of record that captures each prediction at the moment it is committed (timestamped and locked), joins it to the outcome once that arrives, and computes the signed delta between the two. Its whole reason for existing is that a prediction remembered after the outcome is worthless: memory quietly edits "I thought it might go either way" into "I knew it all along,"[1] and the surprise — the only thing worth learning from — evaporates. By committing the prediction up front and making it immutable, the log turns a vague sense of "that went about as expected" (or not) into a hard, hindsight-proof number: received minus expected, with a sign. Everything downstream — how much to learn, which cue to credit, whether to review — runs on that number.
Example¶
A retailer's demand-planning team commits a weekly forecast for each SKU before the week begins: this promoted cereal will sell ≈4,000 units. The forecast is locked with a timestamp. When the week closes, actual sales (≈5,200) are joined to that record and the log stores the delta: +1,200, a positive surprise. Do this across a season and a pattern the planners never felt becomes visible in the record — promoted items run a persistent positive error, plain-shelf items a small negative one. Nobody argued their way to that insight; the committed-before, joined-after structure simply made the signed gaps impossible to rationalize away. The log does not decide what to do about it — it just makes the surprise legible and durable enough that a decision can be made on evidence rather than recollection.
How it works¶
- Commit the prediction before the outcome. Each prediction is recorded with a timestamp and locked, so it cannot be revised once reality is known.
- Join the outcome to its prediction. When the result lands, it is matched to the exact prior record by a stable key (which SKU, which week), not to a fuzzy memory of "roughly what we said."
- Compute and store the signed delta. Received minus expected, keeping the sign — a positive surprise and a negative one are different facts, not just "an error."
- Retain the series. Deltas accumulate into a history, so systematic bias (always over by 20%) becomes visible where a single delta would look like noise.
The distinguishing move is the pre-commitment: the value of the record comes entirely from its being written before the answer was known and never edited after.
Tuning parameters¶
- Logging granularity — per-decision versus batched. Finer capture surfaces more structure but costs discipline and storage; coarse capture is cheap but hides patterns.
- Commitment firmness — how hard the prior is locked. A truly immutable record kills hindsight editing but is unforgiving of typos; a softer lock is friendlier but re-opens the door to after-the-fact revision.
- Join key precision — how strictly an outcome is matched to its prediction. Loose matching captures more pairs but risks pairing the wrong outcome with the wrong forecast, which silently corrupts every delta.
- Prediction format — point value, range, or probability. Richer formats let you score calibration, not just error, but demand more of whoever is predicting.
- Retention horizon — how far back the series is kept, trading storage and drift against the ability to see slow, systematic bias.
When it helps, and when it misleads¶
Its strength is that it neutralizes hindsight bias and makes surprise measurable — it is the substrate every other mechanism in the archetype consumes, and without it there is no prediction error to calibrate at all. It is also the cheapest honesty device available: a locked prior is very hard to argue with.
Its failure modes are all upstream and downstream of the number itself. Garbage predictions in — vague, unfalsifiable, or un-timestamped — yield deltas that cannot be scored. Ambiguous joins (which outcome belongs to which prediction?) produce deltas that are precise-looking and meaningless. And the log can only measure the gap; it says nothing about why it happened or how much to update — read as if it did, a tidy delta invites overreaction to a single draw. Its classic misuse is back-filling: entering a prediction after the outcome is known so the record can say "we called it." The discipline that guards against this is the one the mechanism is built around — commit before, lock it, and never edit.
How it implements the components¶
Prediction–Outcome Delta Log realizes the capture-and-record side of the archetype — the components that must exist before any learning can be calibrated:
prior_prediction_record— its defining act: the committed, timestamped, immutable statement of what was expected.received_outcome_record— the realized result, captured and joined to the specific prediction it answers.signed_error_signal— the computed delta, received minus expected, sign preserved; the atom the whole archetype learns from.prediction_history_memory— the retained series that lets systematic bias emerge from what would otherwise look like scattered one-off errors.
It does not judge which cue *caused an error (that is Credit Assignment Trace), set how much of the error to absorb (that is Learning Rate Schedule), or establish the expected baseline itself (that is Reward Baseline Dashboard). It stores the gap; others interpret and act on it.*
Related¶
- Instantiates: Prediction-Error Learning Calibration — the log supplies the committed predictions, outcomes, and signed deltas the rest of the appraisal runs on.
- Sibling mechanisms: Credit Assignment Trace · Learning Rate Schedule · Expectancy-Calibrated Feedback Form · Negative Prediction Error Review · Positive Surprise Capture · Calibration Curve Review · Reward Baseline Dashboard · Reward Signal Red Team · Shortcut Probe Holdout Set · Surprise Threshold Alert · Temporal-Difference Update Rule
Editorial Notes¶
Form Classification¶
Form family: Record, Log & Register
Rationale: Prediction–Outcome Delta Log operates as a persistent ledger, log, register, or case record that preserves history and traceability because it records every prediction the moment it is made, pairs it with the actual outcome later, and stores the signed gap between them as the unit the rest of the system learns from.
Independent corroboration: The frozen evidence defines Prediction–Outcome Delta Log as 'Records every prediction the moment it is made, pairs it with the actual outcome later, and stores the signed gap between them as the unit the rest of the system learns from', so its operative form is Record, Log & Register.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Pairing forecasts with outcomes and storing signed errors is a statistical calibration and residual-analysis practice.
Related originating lineages:
- Data Science & Analytics — Data science contributes production logging and model-learning pipelines.
Review resolution: Both blind reviewers agree that statistics experimental design is the primary origin. Reconciliation resolves encyclopedia synthesis disagreement. Formative alternate lineages are retained as data_science; later breadth of use is recorded separately as domain_reach=multi_domain, while origin_mode=cross_disciplinary_synthesis describes the relationship among origin lineages.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The log is deliberately dumb: it captures and stores, and refuses to interpret. That is a feature, not a gap. Keeping the record of what-happened separate from the judgments of what-it-means and what-to-do is exactly what lets a team improve its predictions — sharper forecasts, cleaner joins — without re-litigating every downstream decision, and it is why nearly every other mechanism here consumes this one rather than duplicating it.
References¶
[1] Fischhoff, Baruch. "Hindsight ≠ Foresight: The Effect of Outcome Knowledge on Judgment Under Uncertainty". Journal of Experimental Psychology: Human Perception and Performance 1(3): 288–299, 1975. Shows that outcome knowledge increases perceived prior likelihood and distorts judgments about what was foreseeable. registry ↩