THE CANDY FACTORY

Alpha Research Pool

Running — nothing needs you

Read-only

Front Desk

Map Room

Kitchen

Q.C

Shelf

Assortment

Boiler Room

Records

Front Desk

Is it alive, what is it doing, and what needs me?

Panel state:

System state · refreshed 8s ago

SELECT worst_state, plane FROM health_rollup WHERE ts = latest_cache_tick();

Stale — last known rollup, retrying in background.

Running. Searching in context momentum-4h. Nothing needs you.

Research

nominal

Evaluation & Gate

nominal

Portfolio & Live

nominal

Infrastructure

elevated · host at 78%

No health rollup for this tick.

The cache hasn't been populated yet — check again shortly.

Could not load system state.

health_rollup query timed out after 3000ms

What it's doing now

SELECT campaign, generation, elapsed, evals_per_min FROM campaign_status WHERE active = true;

Campaign

momentum-4h

Generation

214

Elapsed

6h 12m

Throughput

1,840 /min

No active campaign

Could not load campaign status

Stale — last known status

What changed since you last looked

SELECT * FROM events WHERE ts > last_viewed_at() ORDER BY ts DESC;

3 new survivors promoted to paper

a1b2c3d4 ⧉ f00dbabe ⧉ 9c8d7e6f ⧉ 12m ago

Incident gate-timeout-217 acknowledged

41m ago

Calibration canary passed — deflation threshold recalibrated

2h ago

No events since your last visit.

Nothing has been promoted, invalidated, or flagged in the last 6 hours.

3 new survivors promoted to paper

12m ago

Data is 47 minutes old — the refresh cache is overdue. Treat this panel as historical, not live.

Could not load recent events.

events_cache query timed out after 3000ms

Open runbook: events-cache-timeout →

1 incident acknowledged

41m ago

Promotions — unavailable this refresh

3 new survivors promoted to paper

measured via fallback method 12m ago

What needs you

SELECT * FROM backlog WHERE status != 'resolved';

Nothing needs you.

Last incident acknowledged 3d ago · next postmortem review in 9 days.

Backlog table is empty

Could not load the backlog

Stale — last known backlog

Resource loops

Our usage — 24% of 30–35% budget

Whole machine — 61%, normal band

Boiler Room →

Map Room

Where has the factory prospected, how hard, and with what luck — and where hasn't it looked?

Panel state:

The Catalogue Grid

refreshed 8s ago

SELECT family, timeframe, trials, survivors, yield_per_1k, deployed FROM catalogue_coverage;

size = trials (log-scaled) color = yield (survivors / 1,000) ▨ hatched = never tried ★ = has a currently deployed survivor

No coverage recorded

The catalogue hasn't been populated for this view.

Could not load this panel

timeout · catalogue_coverage

Stale — last known data, retrying in background

Unopened Doors

SELECT region FROM catalogue_coverage WHERE trials = 0 ORDER BY neighbour_yield DESC LIMIT 5;

Stale — last known list

Unexplored, next to productive regions.

{{ door.family }} × {{ door.timeframe }}

No thin-but-adjacent-to-productive regions in this view.

No unopened doors right now

Everything adjacent to a productive region has been tried.

Could not load this panel

timeout · catalogue_coverage

Saturation curves

SELECT trials_cum, distinct_strategies FROM discovery_curve WHERE region IN (:selected) ORDER BY trials_cum;

● Momentum × 4h — still climbing ● Microstructure × 1m — flattened ● Volume / Flow × 15m — thin, wide projection

Dashed = projected richness · shaded band = its uncertainty (wider where data is sparse).

No discovery curve for this selection

Select a region on the map to see its saturation trend.

Could not load this panel

timeout · discovery_curve

Stale — last known data, retrying in background

The Prospect Map

SELECT trials, yield_per_1k, atoms_touched FROM catalogue_coverage WHERE trials > 0;

Rich and untouched

Hot seam

Unprospected

Mined out

Bubble size = distinct atoms touched · dashed lines = median split.

Nothing prospected yet

No regions in this view have recorded trials.

Could not load this panel

timeout · catalogue_coverage

Stale — last known data, retrying in background

The Atom Inventory

SELECT dimension, n_exist, n_tried, n_productive FROM atom_inventory;

Stale — last known inventory

Exist → ever tried → ever productive, per catalogue dimension.

{{ row.summary }}

Atom inventory is empty

Could not load the atom inventory

timeout · atom_inventory

Kitchen

What is it cooking now, and is the search actually working?

Panel state:

The production line — generation 214

refreshed 8s ago

SELECT stage, count, pct_of_generated FROM generation_funnel WHERE generation = 214;

Stale — last known funnel, retrying in background.

{{ row.count }}

→

Losses are drawn at full weight — most candidates fail here, and that is the gate working correctly.

No candidates generated this cycle

The generator hasn't produced a new batch for generation 214 yet.

Could not load this panel

timeout after 8000ms · generation_funnel

Elite fitness trajectory

SELECT generation, island, best_fitness, median_fitness FROM population_fitness ORDER BY generation;

● Island A — best (solid) / median (dashed) ● Island B ● Island C ▬ evaluation budget /gen (right axis)

No lineage recorded yet

Islands haven't reported fitness for this generation range.

Could not load this panel

timeout · population_fitness

Stale — last known data, retrying in background

Budget burn

SELECT generation, evals_cumulative, envelope FROM evaluation_budget ORDER BY generation;

Dashed line = the pre-registered evaluation envelope — a hard ceiling, not a suggestion.

No evaluations recorded yet

This generation hasn't burned budget — check back once the run starts.

Could not load this panel

timeout · evaluation_budget

Stale — last known data, retrying in background

Diversity

SELECT generation, niche_diversity_index FROM population_diversity ORDER BY generation;

How spread the population is across niches. A collapse here is a genuine pathology — premature convergence.

No diversity samples in this range

The niche index hasn't been computed for these generations.

Could not load this panel

timeout · population_diversity

Stale — last known data, retrying in background

Islands

SELECT generation, best_fitness FROM population_fitness WHERE island = :island ORDER BY generation;

Island A · 3 migrations

Island B · 2 migrations

Island C · 3 migrations

Dots mark migration events between islands.

No island activity recorded

No migrations or fitness reports yet for this generation range.

Could not load this panel

timeout · population_fitness (island)

Stale — last known data, retrying in background

Quality Control

Is what we found real?

Panel state:

The gate funnel — generation 214

refreshed 8s ago

SELECT stage, count, reject_reason FROM gate_funnel WHERE generation = 214;

{{ row.count }}

→

Rejection is the gate working. A candidate surviving every check is the rare exception, not the norm.

No candidates evaluated this generation

Could not load this panel

timeout · gate_funnel

Stale — last known funnel, retrying in background

The deflation adjustment

SELECT raw_sharpe, raw_p, n_attempts, deflated_p, deflated_sharpe FROM deflation_ledger WHERE candidate = 'a1b2c3d4';

The more strategies tried, the better the best one looks by luck alone — so the bar rises with the number of attempts. This is why a candidate can post an outstanding raw number and still fail.

Raw result

Sharpe 2.14

p = 0.0002 (looks decisive)

Attempts judged against

214

strategies tested this generation

Deflated result

Sharpe 1.42

deflated p = 0.031 — PASS

No candidate awaiting judgement

Could not load this panel

timeout · deflation_ledger

Stale — last known ledger, retrying in background

Certificate of analysis — a1b2c3d4

SELECT check_name, statistic, threshold, verdict FROM certificate WHERE candidate = 'a1b2c3d4';

Checks

Walk-forward folds — 6/6 passed

Equity curve — backtest / holdout

Cross-validation distribution

No candidate under review

Could not load this panel

timeout · certificate

Stale — last known data, retrying in background

The calibration canary

SELECT generation, claimed_reject_rate, actual_reject_rate FROM calibration_canary ORDER BY generation;

▬ claimed rejection rate (5%) ● actual rejection rate

Nominal — the gate is rejecting noise at the rate it claims.

No canary runs recorded

Could not load this panel

timeout · calibration_canary

Stale — last known data, retrying in background

The holdout ledger

SELECT ts, candidate, purpose, slice, pct_spent FROM holdout_access_log ORDER BY ts DESC;

Cumulative holdout spent

6.2% of budget

{{ row.hash }}

{{ row.spent }}

Holdout is spent, not renewable. Every access here is permanent.

No holdout accesses recorded

Could not load the holdout ledger

timeout · holdout_access_log

Stale — last known ledger, retrying in background

The Shelf

What do we actually have, and how distinct is it?

Panel state:

The jars — survivors by behavioural class

refreshed 4m ago

SELECT behavioural_class, status, count(*) FROM survivors GROUP BY behavioural_class, status;

{{ jar.total }} total {{ jar.live }} LIVE

Live Paper Onramp Retired

No survivors on the shelf yet

Could not load this panel

timeout · survivor_jars

Stale — last known jar counts, retrying in background

The distinctness map

SELECT candidate, embedding_x, embedding_y, behavioural_class FROM distinctness_embedding;

Each point is a surviving strategy, positioned so near-duplicate ideas sit close together. Tight clusters are redundancy — the same idea kept multiple times; scattered points are genuinely distinct edges.

Momentum Mean Reversion Volume / Flow Breakout Volatility Arb Pairs / Relative Value

Not enough survivors yet to map distinctness

Could not load this panel

timeout · distinctness_embedding

Stale — last known positions, retrying in background

Inventory

SELECT dna_hash, behavioural_class, regime_profile, holdout_status, live_status, decay_state FROM shelf_inventory;

DNA

Class

Regime profile

Holdout

Live status

Decay

{{ row.hash }}

No survivors recorded

Could not load the inventory

timeout · shelf_inventory

Stale — last known inventory, retrying in background

Decay watch

SELECT dna_hash, expected_metric, live_metric, drift_pct, severity FROM decay_watch ORDER BY drift_pct ASC;

DNA

Class

Expected

Live actual

Drift

Severity

{{ row.hash }}

{{ row.expected }}

{{ row.actual }}

{{ row.drift }}

Drift is measured against each survivor's own certified backtest expectation, not against zero.

No survivors are currently drifting from expectation

Could not load decay watch

timeout · decay_watch

Stale — last known drift readings, retrying in background

The Assortment

Is the box better than its best single candy — and how is it doing live?

Panel state:

The box — current portfolio

refreshed 2m ago

SELECT dna_hash, behavioural_class, weight_pct FROM portfolio_members ORDER BY weight_pct DESC;

{{ member.hash }}

{{ member.weightLabel }}

No members currently in the portfolio

Could not load this panel

timeout · portfolio_members

Stale — last known weights, retrying in background

Portfolio versus best member

SELECT ts, portfolio_equity, best_member_equity FROM portfolio_vs_best ORDER BY ts;

● portfolio (box) ● best single member

Portfolio Sharpe 1.68 vs best member's 1.42 — the improvement is statistically significant (deflated p = 0.012). The box is genuinely better than its best part.

No live track record yet to compare

Could not load this panel

timeout · portfolio_vs_best

Stale — last known comparison, retrying in background

Why it's better — diversification

SELECT class_a, class_b, correlation FROM behavioural_class_correlation;

Cross-class correlation

Regime complementarity

Momentum and Breakout members lean on trending regimes; Mean Reversion and Pairs lean on range-bound ones. Low cross-correlation between these groups is what lets the box smooth over regime shifts that would otherwise hurt any single member.

Not enough members yet to estimate correlation

Could not load this panel

timeout · class_correlation

Stale — last known correlations, retrying in background

The live lifecycle board

SELECT dna_hash, behavioural_class, stage FROM deployment_lifecycle;

{{ card.hash }}

No deployments currently tracked

Could not load the lifecycle board

timeout · deployment_lifecycle

Stale — last known board, retrying in background

Risk

SELECT drawdown_pct, kill_switch_pct, gross_exposure_pct, capacity_headroom_pct FROM live_risk_snapshot ORDER BY ts DESC LIMIT 1;

Drawdown vs kill-switch

-6.4%

Kill-switch trips at -18.0% — currently 36% of the way there.

Gross exposure

138%

Normal band (below 170%). Elevated and danger bands sit above.

Capacity headroom

61%

Room left before market impact degrades expected returns.

No live risk snapshot recorded yet

Could not load the risk snapshot

timeout · live_risk_snapshot

Stale — last known risk readings, retrying in background

The Boiler Room

Is the machine healthy and behaving itself?

Panel state:

The health tree — worst-of rollup

refreshed 6s ago

SELECT component, plane, state FROM health_rollup ORDER BY plane, component;

Good (Green) Warning (Amber) Unknown — More alarming than Warning Reserved for genuine danger (Red)

Alpha Research Pool {{ systemWord }}

No health rollup reported yet

Could not load this panel

timeout · health_rollup

Stale — last known rollup, retrying in background

The two resource dials — never merged

SELECT own_usage_pct, whole_machine_pct FROM resource_snapshot ORDER BY ts DESC LIMIT 1;

Our own usage — 24% of a 30–35% budget

"Am I greedy?" — our share of the whole machine.

Whole machine — 58%, normal band

"Is the machine in danger?" — whoever's fault, everyone's problem.

No resource reading yet

Could not load resource dials

timeout · resource_snapshot

Stale — last known readings, retrying in background

The interlock timeline

SELECT ts, interlock_state FROM interlock_history ORDER BY ts;

Normal Elevated Brake Critical

Entry threshold 80% triggers the brake; release threshold 65% lifts it — the gap keeps it from flapping.

No interlock history recorded

Could not load the interlock timeline

timeout · interlock_history

Stale — last known timeline, retrying in background

Governor decisions

SELECT generation, worker_count, reason FROM governor_decisions ORDER BY generation;

{{ step.genLabel }} {{ step.workers }} workers {{ step.reason }}

No governor decisions recorded this run

Could not load this panel

timeout · governor_decisions

Stale — last known decisions, retrying in background

Reclamation & leak watch

SELECT pass_id, mb_reclaimed FROM cleanup_passes ORDER BY pass_id; SELECT pass_id, baseline_mb FROM memory_baseline ORDER BY pass_id;

Reclamation per pass (MB)

2 of the last 7 passes reclaimed nothing — flat is a warning, not a non-event.

Leak watch — baseline after cleanup (MB)

Baseline has drifted up ~15% over 7 cleanups — a slow leak, not yet flagged.

No cleanup passes recorded yet

Could not load reclamation data

timeout · cleanup_passes

Stale — last known memory data, retrying in background

Incidents & breakers

SELECT id, kind, description, status, owner, postmortem_owed FROM incidents_and_breakers ORDER BY ts DESC;

Description

Status

Owner

Postmortem

{{ row.id }}

No open incidents or tripped breakers

Could not load incidents & breakers

timeout · incidents_and_breakers

Stale — last known status, retrying in background

Actions — the only place in the product where a click changes anything

Each requires a deliberate confirmation before it takes effect.

Acknowledge this incident?

Run a cleanup pass now?

Stop research safely? The live book is unaffected.

The Records Office

Why did it decide that, and can I trust it?

Panel state:

Universal search

Returns that thing's complete life: every trial, verdict, promotion, invalidation.

Lineage — {{ rdSubjectLabel }}

SELECT dna_hash, parent_context, birth_archetype, generation FROM lineage WHERE dna_hash = '{{ rdSubjectHash }}';

{{ step.label }}

Birth archetype: {{ rdArchetype }}

No lineage on record for this subject

Could not load this panel

timeout · lineage

View runbook: lineage-service-timeout

Stale — last known lineage, retrying in background

Version timeline

SELECT version, deployed_at, kind, invalidated_from, invalidated_to FROM code_versions ORDER BY deployed_at;

Invalidation window: {{ ev.invalidFrom }} → {{ ev.invalidTo }}

No version history recorded

Could not load this panel

timeout · code_versions

View runbook: version-service-timeout

Stale — last known timeline, retrying in background

Invalidation ledger

SELECT id, ts, scope, reason FROM invalidations ORDER BY ts DESC;

Date

What it touched

Reason

{{ row.id }}

No invalidations on record — nothing has been revoked

Could not load the invalidation ledger

timeout · invalidations

View runbook: invalidation-service-timeout

Stale — last known ledger, retrying in background

Foundations

Style foundation and component set for The Candy Factory.

Ground & brand palette — wayfinding only, never data

Ground

Sugar pink

Caramel

Pistachio

Liquorice

Data palette — colourblind-safe, never brand

Sequential (magnitude / yield)

Diverging (above / below expectation)

Categorical (series)

Type

Aa — Chakra Petch, angular tech-terminal type

13px body · dense, legible at small sizes

a1b2c3d4e5f6 — JetBrains Mono, for hashes, ids, timestamps, SQL

page question

hero sentence

stat value

panel title

body

label / caps

Spacing scale

A 4px base. Density is the point: 6 and 10 do most of the work inside panels, 18 separates panels, 22 is panel padding. Nothing larger — whitespace is spent on the hero visual, not the page.

4

6

10

14

18

22

Elevation, focus and hover

Three levels only. Hover brightens a border, never moves anything. Focus is a visible cyan ring — this instrument is driven from the keyboard as much as the mouse.

e0 · flush

Sub-panels, table rows. No shadow.

e1 · panel

The universal container.

e2 · drawer

Drill-down only. Nothing else floats.

Rows highlight on hover:

momentum-4h · 12,480 trials

Absence — drawn, never omitted

One treatment, used for every unmeasured value in the product: a 45° violet hatch on the ground. It cannot be read as a low value, and it is deliberately more arresting than amber.

Not measured

Measured, low

Inline in a series — the gap states itself

Gen 3 — no measurement recorded (reporting stopped, not zero).

Health-state tokens — unknown reads as more alarming than amber

Good

Amber

Unknown

Red — reserved for genuine danger

Panel — the universal container

Title, freshness stamp, query affordance (</>), body — every panel in the product is this same shell. Loading has nothing to be fresh about yet, so the stamp is skipped entirely; stale and error keep it front and center, because knowing when a number last actually worked is the point.

Default

Region yield

refreshed 4s ago </>

Loading — no stamp yet

Empty

Family × Operator

refreshed 4s ago </>

No trials in this view

Widen the timeframe or family filter to find something.

Stale — served, flagged

Budget burn

stale · 47m ago </>

Last known data — retrying in background.

Error

Diversity index

fetch failed </>

Could not load this panel

timeout · population_diversity

Hash chip — monospace, truncated to 8, click to copy, hover for full

a1b2c3d4 ⧉

Stat block

Caps label, monospace value, and a qualifier line that is never optional for a statistical figure — sample size, interval, or the absence treatment.

Survivors

84

counted, not estimated

Yield /1,000

3.1 ± 1.4

n = 12,480 trials

Yield /1,000

8.0 ± 7.9

n = 125 — thin evidence

Yield /1,000

unknown

never tried here

Status pill and the query affordance

The pill always carries the state in words, never colour alone. The query glyph sits in every panel corner and is identical everywhere.

Running Braked — host at 91% Unknown — not reporting Stopped — kill-switch tripped Read-only

</>

SELECT family, timeframe, trials, survivors FROM catalogue_coverage WHERE campaign = 'momentum-4h';

Data table

Caps header, 1px rules, no zebra striping — density comes from tight rows, not from fill. Numerals monospace and right-aligned. A row opens the drawer; unknown cells get the hatch inline.

DNA

Class

Trials

Yield

Live status

a1b2c3d4

Momentum breakout

12,480

3.1

LIVE

f00dbabe

Mean reversion band

6,204

1.8

PAPER

9c8d7e6f

Volume / flow divergence

418

ONRAMP

Chart style specification

Every chart in the product is built from this. It is written to be transcribable straight into an ECharts theme.

Reference chart — band, series, absence

estimate 95% band not measured

axisLine

1px oklch(35% .03 240); no arrows, no bevels.

splitLine

Horizontal only, 1px dashed oklch(28% .03 240). Never vertical — it competes with the data.

ticks

4–5 on y, 6–8 on x. 11px, oklch(58% .032 240), monospace for numerals. Log axes label decades only.

legend

Outside the plot, above it, as HTML — not ECharts' own. Keeps every panel's legend identical.

tooltip

Panel ground, 1px border, no shadow, no transition. States the sample size on any statistical value.

series

Categorical set in order; 1.5px lines, no smoothing, symbols only where points are the datum.

bands

Custom polygon series at 0.26 opacity with dotted upper and lower edges — drawn as one shape, never two stacked lines.

absence

The violet hatch, as a mark. Lines break at a missing point; they are never interpolated across it.

animation

false. Data changes; it does not move.

Confirmed-action button — Boiler Room only

Shown here as a style reference; the only place in the product where a click changes anything.

{{ drawer.family }} × {{ drawer.timeframe }}

Not found

Trials

{{ drawer.cell.t }}

Survivors

{{ drawer.cell.s }}

Yield /1,000

{{ drawer.yieldVal }}

★ This region has a survivor currently deployed.

Never tried. No trials recorded for this region.

No record matches {{ drawer.query }}.

Check the hash and try again, or open the Records Office for a broader search.

Exist

{{ drawer.exist }}

Tried

{{ drawer.tried }}

Productive

{{ drawer.productive }}

{{ toastText }}