Reward Baseline Dashboard¶
Reference dashboard — instantiates Prediction-Error Learning Calibration
Establishes and displays the expected-reward baseline so a result is read as above or below what was already anticipated — not as raw good or bad news.
Reward Baseline Dashboard is the standing instrument that answers "what did we already expect here?" before anyone reacts to an outcome. Its defining move is to compute and display an expected-reward baseline — a context-adjusted level the system had reason to anticipate — and to plot every actual result against it, so the salient quantity on the screen is the gap, not the raw number. A result that merely matches its baseline is shown as unremarkable even when the absolute figure is large; only departures from expectation stand out. That framing is what stops a system from re-celebrating fully-expected wins and re-punishing fully-expected losses. It is upstream of any trigger or update: it defines the reference frame, and leaves the firing and the learning to other mechanisms.
Example¶
A subscription app's growth team watches daily signups. Raw, the numbers are misleading: Mondays and post-launch weeks always spike, so a big day gets over-celebrated and a quiet Sunday sparks needless alarm. The dashboard instead maintains a baseline conditioned on day-of-week, active campaigns, and seasonality, and draws each day's actual signups against that expected band. On a Monday that comes in at ≈1,900 against an expected ≈1,850, the bar sits on the baseline — a visibly ordinary day, no matter how large the headline. Two weeks later a plain Wednesday lands ≈40% above its baseline with no campaign running; that gap is what the display makes impossible to miss. The team's conversation shifts from "was today good?" to "which days genuinely departed from what we already knew to expect?" — the only days worth learning from.
How it works¶
The distinguishing element is a context-conditioned baseline, not a flat average. The dashboard keeps a rolling history of outcomes, fits an expected level that adjusts for the covariates that ordinarily move the metric (calendar, campaign, cohort), and renders the residual — actual minus expected — as the primary visual. It deliberately produces no verdict and no threshold: it neither decides that a gap is "real" nor prescribes how much to learn from it. Its whole contribution is to make expectation itself legible, so that surprise can later be judged against an honest reference rather than against zero.
Tuning parameters¶
- Baseline window — how much history the expected level is fit from; shorter tracks regime shifts, longer smooths noise but lags real change.
- Context conditioning — which covariates the baseline adjusts for; more conditioning removes "expected" variation but risks absorbing genuine signal into the baseline.
- Aggregation grain — per-event, daily, or cohort; finer grain surfaces local surprise but is noisier.
- Band vs point — whether the expected level is shown as a single line or an expected range; a band communicates ordinary variation but invites the display to be read as an alert (which it is not).
- Refresh cadence — how often the baseline is refit as new outcomes land.
When it helps, and when it misleads¶
Its strength is that it operationalizes the archetype's first rule — expected outcomes should mostly maintain the model — by making the expected level something you can see, so an on-baseline result stops masquerading as information. This mirrors the reinforcement-learning practice of learning on reward minus a baseline: an outcome equal to its baseline carries no signal.[n1]
Its failure mode is that the baseline is itself a model and can be wrong. A baseline fit on too much history, or conditioned on the wrong covariates, will either flatten genuine surprises into "expected" or manufacture departures that are really noise. The classic misuse is retrofitting the baseline to recent data so that whatever just happened lands on-target — a moved goalpost that launders a miss into a hit. The discipline is to fix the baseline methodology in advance, validate it against held-out periods, and keep it visibly separate from the outcomes it is judging.
How it implements the components¶
value_reference_frame— its core output: the context-adjusted expected level that defines what counts as "as anticipated."received_outcome_record— it ingests and displays the stream of actual outcomes plotted against that reference.prediction_history_memory— it retains the rolling history of outcomes the baseline is continuously refit from.
It does not pair each specific prior prediction with its outcome to compute the signed gap — that is Prediction–Outcome Delta Log — nor does it fire when a gap is large; that trigger is Surprise Threshold Alert.
Related¶
- Instantiates: Prediction-Error Learning Calibration — it supplies the expectation reference the whole calibration depends on.
- Sibling mechanisms: Surprise Threshold Alert · Prediction–Outcome Delta Log · Learning Rate Schedule · Temporal-Difference Update Rule · Credit Assignment Trace · Reward Signal Red Team
Editorial Notes¶
Form Classification¶
Form family: Monitoring, Sensing & Alerting
Rationale: Reward Baseline Dashboard operates as ongoing observation, sensing, or alerting that detects and surfaces state without itself executing the response because it establishes and displays the expected-reward baseline so a result is read as above or below what was already anticipated — not as raw good or bad news.
Independent corroboration: The frozen evidence defines Reward Baseline Dashboard as 'Establishes and displays the expected-reward baseline so a result is read as above or below what was already anticipated — not as raw good or bad news', so its operative form is Monitoring, Sensing & Alerting.
Nearest alternative: Analysis, Modeling & Optimization — Reward Baseline Dashboard includes features of an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution, but its defining operation is ongoing observation, sensing, or alerting that detects and surfaces state without itself executing the response.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Neuroscience
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: The defining comparison of received reward with predicted reward is canonical reward-prediction-error neuroscience, where fully predicted outcomes establish the baseline and deviations drive learning. Psychology, cognition, behavioral economics, finance, and statistics provide parallel expectation-relative interpretations.
Related originating lineages:
- Behavioral Economics — behavioral_economics contributes experimental evidence on incentives, expectations, and strategic behavior to the mechanism’s formative or independently convergent form; that contribution does not displace the primary neuroscience lineage.
- Cognitive Science — cognitive_science contributes working memory, attention, expectation, and representation to the mechanism’s formative or independently convergent form; that contribution does not displace the primary neuroscience lineage.
- Economics & Finance — economics_finance contributes cost, allocation, repeated-game, expectation, and risk-analysis traditions to the mechanism’s formative or independently convergent form; that contribution does not displace the primary neuroscience lineage.
- Psychology — psychology contributes social learning, group regulation, trust, behavior rehearsal, and expectation to the mechanism’s formative or independently convergent form; that contribution does not displace the primary neuroscience lineage.
- Statistics & Experimental Design — statistics_experimental_design contributes prospective protocols, uncertainty, longitudinal follow-up, and model validation to the mechanism’s formative or independently convergent form; that contribution does not displace the primary neuroscience lineage.
Review resolution: The blind reviewers disagreed on primary lineage (economics_finance versus psychology); authoritative or primary research supports neuroscience as the best historical origin. The defining comparison of received reward with predicted reward is canonical reward-prediction-error neuroscience, where fully predicted outcomes establish the baseline and deviations drive learning. Psychology, cognition, behavioral economics, finance, and statistics provide parallel expectation-relative interpretations. The cited Schultz, Dopamine Reward Prediction Error Coding directly supports the defining operation used in that choice. All independently supported contributing domains are retained without an arbitrary cap, while domain_reach=multi_domain records later applicability separately from provenance.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
Notes¶
The baseline is a claim, not a fact. Because everything downstream reads surprise relative to this reference, a quietly stale or over-conditioned baseline corrupts every gap computed against it — so the dashboard's own baseline deserves the same periodic validation it enables for everything else.
[n1] Subtracting a baseline from the reward (the advantage) is a standard variance-reduction step in policy-gradient reinforcement learning; the learning signal is the amount by which an outcome beats or misses its baseline, so an outcome equal to its baseline contributes nothing to the update. ↩