RelQL Studio — Pages canvas

01 · Project overview

Loaded — live operations

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Projects ▸ acme-supply-chain

Acme Supply Chain

Production

DH

MK

JS

+4

Live ingest · records/s last 15m

Kfkalshi_ws 3,412/s

BQkalshi_markets 1,240/s

Pgorders_pg 214/s

S3news_wire 86/s

View all 14 connectors →

RelQL queries

128 queries/s

p50 41ms p95 187ms

How fast predictions come back.

Inference pipelines

6 active

oil_impact_stream→ kafka

reorder_forecast→ table

Predictions writing continuously.

Model serving

Healthy

rtj-oil-ft fp16 38ms p50

2 checkpoints serving · RelativeDB Cloud

IngestionHealthy Query engineHealthy PipelinesHealthy Model servingHealthy StorageHealthy Full observability →

Recent activity

▶Run finetune_oil_impact completed · AUROC 0.874run

12m ago→

⚠Alert oil_impact_1sigma fired · published kafka://oil-alertsalert

26m ago→

▤Dataset oil_impact_scores grew +48,120 rowsdataset

1h ago→

📥kalshi_ws backfill window closed · 2.1M recordsingestion

2h ago→

ƒDerived headline_embeddings rebuiltderived

4h ago→

View all activity →

💬

The overview is an operations dash, not an inventory: what is flowing right now, how fast queries answer, what changed last. Every activity row ends in a → jump to its object.

Loading skeleton

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Projects ▸ acme-supply-chain

Acme Supply Chain

Production

Live ingest · records/s

RelQL queries

Inference pipelines

Model serving

IngestionChecking… Query engineChecking… PipelinesChecking… Model servingChecking… StorageChecking…

Recent activity

💬

Rates, sparklines, health words, and the feed are the async regions; chrome and card labels render immediately.

Degraded — news_wire Failed, QPS dips degraded

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Projects ▸ acme-supply-chain

Acme Supply Chain

Production

Live ingest · records/s last 15m

Kfkalshi_ws 3,398/s

BQkalshi_markets 1,238/s

Pgorders_pg 214/s

S3news_wireFailed 0/s

View all 14 connectors →

RelQL queries

96 queries/s

p50 44ms p95 412ms

Dip follows the news_wire stop at 13:41.

Inference pipelines

6 active

oil_impact_streamLagging

reorder_forecastHealthy

oil_impact_stream waits on headlines.

Model serving

Healthy

rtj-oil-ft fp16 39ms p50

2 checkpoints serving · RelativeDB Cloud

IngestionFailed Query engineHealthy PipelinesLagging Model servingHealthy StorageHealthy Full observability →

💬

One connector failing shows its blast radius on one screen: rate flatlines to 0/s, QPS sparkline dips, the dependent pipeline goes Lagging, and the health band names the subsystem. Color stays on state only.

Helper panel open — nav help state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Projects ▸ acme-supply-chain

Acme Supply Chain

Production

Live ingest · records/s last 15m

Kfkalshi_ws 3,398/s

BQkalshi_markets 1,238/s

Pgorders_pg 214/s

S3news_wireFailed 0/s

View all 14 connectors →

RelQL queries

96 queries/s

p50 44ms p95 412ms

Inference pipelines

6 active

Model serving

Healthy

rtj-oil-ft fp16

IngestionFailed QueriesHealthy PipelinesLagging ModelsHealthy StorageHealthy

Studio helper ✕

where do I see why ingestion dropped?

Rates over time live in Observability ▸ Ingestion. The drop traces to news_wire — Failed since 13:41 (S3 credentials rejected). Open the connector →

Continue in workspace →

💬

The FAB opens the same chat + history as the workspace; here it navigates rather than analyzes, and hands off with "Continue in workspace →".

Tablet · 834px responsive

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Projects ▸ acme-supply-chain

Acme Supply Chain

Production

Live ingest · records/s last 15m

Kfkalshi_ws3,412/s

BQkalshi_markets1,240/s

Pgorders_pg214/s

S3news_wire86/s

View all 14 connectors →

RelQL queries

128q/s

p50 41msp95 187ms

Inference pipelines

6active

2 streaming → kafka

Model serving

Healthy

rtj-oil-ft fp16

IngestionHealthy QueriesHealthy PipelinesHealthy ModelsHealthy StorageHealthy

Recent activity

▶ Run finetune_oil_impact completed12m→

⚠ Alert oil_impact_1sigma fired26m→

▤ Dataset oil_impact_scores +48,120 rows1h→

📥 kalshi_ws backfill window closed2h→

View all activity →

💬

Mobile · 390px responsive

acme-supply-chain

Acme Supply Chain

Prod

Live ingest · records/s

Kfkalshi_ws3,412/s

BQkalshi_markets1,240/s

Pgorders_pg214/s

View all 14 connectors →

Queries

128 q/s

p50 41ms · p95 187ms

Pipelines

6 active

Healthy

IngestHealthy ModelsHealthy StorageHealthy

Recent activity

▶ finetune_oil_impact completed→

AUROC 0.874 · 12m ago

⚠ oil_impact_1sigma fired→

kafka://oil-alerts · 26m ago

▤ oil_impact_scores +48,120 rows→

1h ago

View all activity →

💬

⌂Home

⛁Sources

◫Graph

💬Chat

▶Runs

02 · Sources

Loaded — top instances

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

acme-supply-chain ▸ Sources

Sources

Connector instances feeding this project — a type can have many instances.

Instance	Type	Sharing	Rate	Health
kalshi_ws	KfKafka	All projects	3,412/s	Healthy	→
orders_pg	PgPostgres	This project	214/s	Healthy	→
news_s3	S3S3	All projects	86/s	Healthy	→
market_bq	BQBigQuery	This project	58/s	Lagging	→
filings_watch	LoLocal files	This project	4/s	Healthy	→

Sorted by records/s · shared connectors are marked "All projects".

💬

Instances, not types: three BigQuery connections are three rows. The table stays bounded to the five busiest; the catalog of connector types lives behind "+ New connector".

Catalog open — connector types state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

acme-supply-chain ▸ Sources

Sources

New connector

Pick a type — you can run as many instances of each as you need.

✕

PgPostgres3 instances

MyMySQL0 instances

BQBigQuery3 instances

SfSnowflake0 instances

S3S32 instances

GCGCS0 instances

FTSFTP0 instances

LoLocal files1 instance

ESElasticsearch1 instance

SlSlack1 instance

PXPMXT0 instances

KfKafka1 instance

KgKaggle1 instanceImport datasets

{}Custom REST/Python1 instance

14 types · 14 running instances across this org

💬

The full type catalog is an overlay, not the page: monogram tiles carry per-type instance counts. Kaggle is highlighted mid-selection and hints that imports land as ready-to-use datasets.

Row expanded — sharing controls state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

acme-supply-chain ▸ Sources

Sources

Instance	Type	Sharing	Rate	Health
kalshi_ws	KfKafka	All projects	3,412/s	Healthy	▾

Available to all projects

Other projects mount this connector read-only. Credentials stay here.

Used by: oil-desk · retail-forecast · this project

Copies host, topic, and sync settings into a fresh instance you own.

orders_pg	PgPostgres	This project	214/s	Healthy	→
news_s3	S3S3	All projects	86/s	Healthy	→

💬

Sharing expands inline: the toggle names its consequence in plain verbs, the clone action makes "many instances of one type" a one-click path.

Loading skeleton

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

acme-supply-chain ▸ Sources

Sources

Instance	Type	Sharing	Rate	Health

💬

Five skeleton rows match the bounded table; filters and the New connector action render immediately.

Tablet · 834px responsive

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

acme-supply-chain ▸ Sources

Sources

Instance	Rate	Health
Kfkalshi_wsAll projects	3,412/s	Healthy	→
Pgorders_pg	214/s	Healthy	→
S3news_s3All projects	86/s	Healthy	→
BQmarket_bq	58/s	Lagging	→
Lofilings_watch	4/s	Healthy	→

💬

Mobile · 390px responsive

acme-supply-chain ▸ Sources

Sources

Kfkalshi_ws →

HealthyAll projects 3,412/s

Pgorders_pg →

Healthy 214/s

S3news_s3 →

HealthyAll projects 86/s

BQmarket_bq →

Lagging 58/s

💬

⌂Home

⛁Sources

◫Graph

💬Chat

▶Runs

03 · Connector setup

Postgres — full form, in page

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

acme-supply-chain ▸ Sources

Sources

Sources ▸ New connector ▸ Postgres

Pg

Postgres

Connector name

Host

Port

Database

User

Password

SSL mode

Sync mode

Full refresh Incremental · cursor CDC · logical replication

CDC streams inserts, updates, and deletes as they commit — best for live rates.

Replication slot

Schedule

Share to all projects

💬

Setup is a centered drawer over Sources — shown here in the 1280px page frame. The artboards below crop to the drawer itself, one per connector type. Every form ends the same way: Test connection, Create connector, and the org-wide share checkbox. No skeleton state: the form has no async region.

BigQuery connector type

Sources ▸ New connector ▸ BigQuery

BQ

BigQuery

Connector name

GCP project id

Dataset

Service-account key

Drop the JSON key file here, or browse

service-account.json · needs bigquery.dataViewer

Location

Incremental column

New rows are pulled by watermark on this column.

Share to all projects

S3 connector type

Sources ▸ New connector ▸ S3

S3

Connector name

Bucket

Prefix

Region

Authentication

IAM role ARN Access keys

Role ARN

File format

New-file watch

Ingest new objects as they land

Share to all projects

Kaggle — dataset import connector type

Sources ▸ New connector ▸ Kaggle

Kg

Kaggle

Import datasets

Connector name

Dataset slug or competition

Paste a dataset URL or slug — competitions work too.

API token

From kaggle.com ▸ Account ▸ Create API token.

Target dataset name

Refresh

Manual Daily

Share to all projects

Kafka connector type

Sources ▸ New connector ▸ Kafka

Kf

Kafka

Connector name

Brokers

Topic

Consumer group

Start offset

earliest latest

Schema registry URL (optional)

Share to all projects

Slack connector type

Sources ▸ New connector ▸ Slack

Sl

Slack

Connector name

Workspace

Authorizedacme-hq.slack.com

Channels

#trading ✕ #ops-alerts ✕ #supplier-updates ✕ Add channel…

History depth

Messages become timestamped rows — anchor-time rules apply like any other table.

Share to all projects

Elasticsearch connector type

Sources ▸ New connector ▸ Elasticsearch

ES

Elasticsearch

Connector name

Endpoint

Index pattern

API key

Query filter (optional)

Share to all projects

Local watched folder connector type

Sources ▸ New connector ▸ Local files

Lo

Local watched folder

Connector name

Path

Recursive

Watch subfolders too

Format

Schema inference

Infer column names and types from the first files

You can correct inferred types in the dataset view before anything trains on them.

Share to all projects

Custom REST/Python connector type

Sources ▸ New connector ▸ Custom REST/Python

{}

Custom REST/Python

Connector name

Entrypoint

A Python script path or a REST URL to poll.

Auth header (optional)

Polling interval

# scripts implement three calls:

discover() # → tables + columns

read(cursor) # → rows since cursor

checkpoint() # → durable cursor

Share to all projects

The contract is the documentation: three mono function names, shown where the script path is typed.

Test connection — success state

Sources ▸ New connector ▸ Postgres

Pg

Postgres

HealthyConnection succeeded

connected to db.internal.acme.com:5432 in 84ms
found 12 tables · logical replication available · slot relql_orders_slot free

Everything checks out. Create the connector to start the first sync.

Share to all projects

Test connection — failed error

Sources ▸ New connector ▸ Postgres

Pg

Postgres

User

Check this user exists and allows password auth.

Password

Share to all projects

Failure quotes the driver error in mono, marks only the fields to fix, and disables Create until a test passes.

Tablet · 834px — Kaggle responsive

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Sources ▸ New connector ▸ Kaggle

Kg

Kaggle

Connector name

Dataset slug or competition

API token

Target dataset name

Refresh

ManualDaily

Share to all projects

💬

Mobile · 390px — Kaggle responsive

Sources ▸ New ▸ Kaggle

Kg

Kaggle

Connector name

Dataset slug or competition

API token

Target dataset name

Refresh

ManualDaily

Share to all projects

💬

⌂Home

⛁Sources

◫Graph

💬Chat

▶Runs

04 · Observability

Loaded — Ingestion tab

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸Observability

Observability

Ingestion

Queries

Pipelines

Models

Infrastructure

Total ingest · records/s (24h)

peak 6,480/s · now 5,206/s

2026-07-27 14:00

02:00

now

14:02

Connector	Rate	24h volume	Lag	Health
Kfkalshi_ws	3,412/s	281.4M	0.3s	Healthy	→
BQkalshi_markets	1,240/s	102.9M	4.1s	Healthy	→
Pgorders_pg	214/s	17.8M	0.9s	Healthy	→
S3news_s3	86/s	7.2M	12s	Healthy	→
BQmarket_bq	58/s	4.9M	38m	Lagging	→

View all 14 connectors →

💬

One page, five lenses: ingestion, queries, pipelines, models, infrastructure. The area chart rides the shared violet time spine; the table stays bounded to the five busiest connectors.

Queries tab — QPS, latency, slow queries state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸Observability

Observability

Ingestion

Queries

Pipelines

Models

Infrastructure

RelQL queries/s

128

peak 312/s at 09:30

Latency · how long a prediction takes to answer

p50 41ms p95 187ms p99 640ms

Half answer under 41ms; the slowest 1% take up to 640ms.

Errors (24h)

7

0.002% of 3.1M queries

Slowest queries (24h) by p95

Query	Task	Calls	p95
PREDICT SUM(transactions.price) OVER (90 DAYS FOLLOWING) FROM customers…	ltv_90d	41,208	1.84s	→
PREDICT ARRAY_AGG(transactions.article_id) OVER (30 DAYS FOLLOWING RANK TOP 12)…	reorder_rank	18,441	1.12s	→
PREDICT SUM(sales.qty) OVER (7 DAYS FOLLOWING HORIZONS 4) FROM stores	store_forecast	9,006	870ms	→
PREDICT NOT EXISTS(orders.*) FROM customers WHERE c.plan = 'basic' AS OF :anchor…	churn_basic	204,112	412ms	→
PREDICT oil_impact_score FROM headlines AS OF :t RETURN EXPECTED VALUE	oil_impact	96,733	388ms	→

View all queries →

💬

Slow-query rows show the real first line of RelQL in mono with syntax color; each jumps to the workbench with the query loaded.

Pipelines tab — streaming inference state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸Observability

Observability

Ingestion

Queries

Pipelines

Models

Infrastructure

Streaming inference pipelines score new rows continuously and can fire triggers into Kafka or pub/sub.

Pipeline	Source → target	Throughput	Lag	Trigger fires (24h)	Health
oil_impact_stream	news_s3 → oil_impact_scores	86/s	1.2s	14	Healthy	→
kalshi_edge_stream	kalshi_ws → market_edges	3,412/s	0.4s	212	Healthy	→
reorder_forecast	orders_pg → reorder_predictions	214/s	0.9s	3	Healthy	→
churn_daily_score	customers → churn_scores	12/s	6.0s	0	Healthy	→
filing_risk_stream	filings_watch → filing_risk	4/s	51s	1	Lagging	→

View all 8 pipelines →

Trigger fires count conditions met, e.g. prediction moves > 1σ → publish to kafka://oil-alerts.

💬

Models tab — serving checkpoints state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸Observability

Observability

Ingestion

Queries

Pipelines

Models

Infrastructure

Checkpoint	Precision	Tokens/s	Serving p50	Where	Health
rtj-oil-ft	fp16	18,400	38ms	RelativeDB Cloud	Healthy	→
rtj-oil-ft	int8	31,200	22ms	EKS us-east-1	Healthy	→
rt-j-reference	fp16	16,900	44ms	on-prem k8s	Healthy	→
churn-head-v3	int8	29,800	19ms	GKE europe-west4	Healthy	→
forecast-head-v1	fp16	17,300	41ms	RelativeDB Cloud	Lagging	→

View all 7 checkpoints →

Lower precision serves faster and smaller; int8 trades a little accuracy for roughly half the latency. Placement chips match the Compute page.

💬

Degraded — news_s3 stopped degraded

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸Observability

Observability

Ingestion

Queries

Pipelines

Models

Infrastructure

Total ingest · records/s (24h)

now 5,104/s

2026-07-27 14:00

13:41 · stop

now

Connector	Rate	24h volume	Lag	Health
Kfkalshi_ws	3,398/s	281.1M	0.3s	Healthy	→
BQkalshi_markets	1,238/s	102.7M	4.0s	Healthy	→
Pgorders_pg	214/s	17.8M	0.9s	Healthy	→
S3news_s3	0/s	6.9M	21m	Failed	→
BQmarket_bq	58/s	4.9M	38m	Lagging	→

View all 14 connectors →

💬

The stop is one critical dot on the chart at 13:41 and one critical-edged row — the failure reads on the spine and in the table without recoloring anything else.

Loading skeleton

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸Observability

Observability

Ingestion

Queries

Pipelines

Models

Infrastructure

Total ingest · records/s (24h)

now

Connector	Rate	24h volume	Lag	Health

💬

The spine axis and "now" caret draw immediately; the chart body and rate cells shimmer until metrics stream in.

Tablet · 834px responsive

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸Observability

Observability

Ingestion

Queries

Pipelines

Models

Infra

Total ingest · records/s (24h)

-24h

now

Connector	Rate	Health
Kfkalshi_ws	3,412/s	Healthy
BQkalshi_markets	1,240/s	Healthy
Pgorders_pg	214/s	Healthy
S3news_s3	86/s	Healthy
BQmarket_bq	58/s	Lagging

View all 14 connectors →

💬

Mobile · 390px responsive

acme-supply-chain

Observability

Ingestion

Queries

Pipelines

Models

Total ingest (24h)

now

Kfkalshi_ws3,412/s

Healthy

BQkalshi_markets1,240/s

Healthy

BQmarket_bq58/s

Lagging

💬

⌂Home

⛁Sources

◫Graph

💬Chat

▶Runs

05 · Context graph

Loaded — graph canvas helper panel open

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Context graph

Context graph

1:N

1:1

N:1

PGcustomers12.4K

customer_id PKint

emailvarchar

planvarchar

◈ signup_attimestamptz

+6 more

PGorders48.1K

order_id PKint

customer_id FKint

◈ placed_attimestamptz

total_usdnumeric

+8 more

PGtransactions512K

txn_id PKbigint

order_id FKint

pricenumeric

◈ settled_attimestamptz

+5 more

RDBcustomers_dedup11.9K

entity_id PKint

customer_id FKint

match_scorenumeric

derived view · +2 more

BQmarket_history41.2M

bar_id PKint64

symbolstring

closefloat64

◈ bar_tstimestamp

+5 more

KFkalshi_ticks

9 cols · ◈1 · 128M rows

KFnews_wire

6 cols · ◈1 · streaming

S3news_archive

7 cols · ◈1 · 3.4M rows

⋯ 4 more tables → ⋯ 7 more below ↓

Showing 8 of 63 tables · pan to explore

map

Chat helper✕

Where do the Kalshi ticks join in?

kalshi_ticks joins market_history on symbol — I centered that edge for you. It also feeds one derived edge into news_wire.

Continue in workspace →

💬

No full-canvas skeleton: the graph hydrates node-by-node from cached schema metadata, so nodes appear progressively on an already-rendered canvas instead of behind a shimmer.

Node selected — relationship stats panel state

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Context graph

Context graph

1:N · 0.83

1:N · 0.97

PGcustomers12.4K

customer_id PKint

emailvarchar

◈ signup_attimestamptz

+7 more

PGorders48.1K

order_id PKint

customer_id FKint

◈ placed_attimestamptz

total_usdnumeric

+8 more

PGtransactions512K

txn_id PKbigint

order_id FKint

◈ settled_attimestamptz

+6 more

RDBcustomers_dedup11.9K

entity_id PKint

derived view · +4 more

BQmarket_history41.2M

symbolstring

◈ bar_tstimestamp

+7 more

⋯ 4 more tables →

Showing 5 of 63 tables

PGorders

raw · oildesk_pg · orders

Rows48,102

Last sync2m ago

HealthHealthy

Temporal◈ placed_at

Edge — customers → orders

Selectivity0.83

share of customers with at least one order

Avg fan-out4.7

orders per customer, mean

Orphaned FK0.2%

orders whose customer_id matches no customer

Edge — orders → transactions

selectivity 0.97 · fan-out 10.6
orphaned FK 0.0%

💬

Edge statistics render on hover as the inverted pill on the edge itself; selecting either endpoint pins the full stats into the panel.

Add edge — derived-edge drawer RelQL / process source

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Context graph

Context graph

new edge — draft

PGcustomers12.4K

customer_id PKint

emailvarchar

+8 more

RDBcustomers_dedup11.9K

entity_id PKint

emailvarchar

derived view · +3 more

PGtransactions512K

txn_id PKbigint

billing_emailvarchar

◈ settled_attimestamptz

+6 more

⋯ 5 more tables →

New derived edge ✕

From

To

Source

Query — define the join in RelQL/SQL Process — Entity resolution fuzzy match

edge_dedup_txn.rqlvalid

1
2
3
4
5

-- link deduped customers to card txns SELECT c.entity_id, t.txn_id FROM customers_dedup c JOIN transactions t ON LOWER(c.email) = LOWER(t.billing_email)

💬

The two endpoint nodes stay highlighted while the drawer is open; the dashed blue edge is the draft. Choosing "Process — Entity resolution" swaps the editor for the matcher's column-pair picker.

Ablations — in-context toggles reversible

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Context graph

Context graph

PGcustomers12.4K

customer_id PKint

+9 more

PGorders48.1K

order_id PKint

+11 more

PGtransactions512K

txn_id PKbigint

pricenumeric

feenumeric

◈ settled_attimestamptz

+5 more

BQmarket_history41.2M

◈ bar_tstimestamp

+8 more

S3news_archive

7 cols · 3.4M rows

↺ Unablate

⋯ 4 more tables →

2 ablations active

Context ablations

Toggled-off tables and columns are dropped from every scored context. Reversible any time.

Tables

PGcustomersIn context

PGordersIn context

PGtransactionsIn context

BQmarket_historyIn context

S3news_archiveAblated

View all 63 tables →

Columns — transactions

priceIn context

feeAblated

venueIn context

View all 9 columns →

💬

Ablated nodes stay on the canvas at 45% with a dashed border and an inline "↺ Unablate" affordance — removal from context is never removal from the map.

EXPLAIN ABLATE — running ~30s progress

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Context graph

Context graph

PGcustomers

10 cols · 12.4K rows

PGorders

12 cols · 48.1K rows

PGtransactions

9 cols · 512K rows

BQmarket_history

9 cols · 41.2M rows

S3news_archive

7 cols · 3.4M rows

EXPLAIN ABLATE — quiet_customers_90d

Scoring the query once per candidate ablation, then ranking by how far predictions move.

1
2
3

EXPLAIN ABLATE PREDICT NOT EXISTS(orders.*) FROM customers AS OF :anchor RETURN PROBABILITY

Re-scoring 14 candidates · ~30s 6 of 14 · elapsed 00:12

💬

This is the async region: a determinate progress card (candidates scored + elapsed time), not a skeleton — the canvas underneath stays visible and dimmed.

EXPLAIN ABLATE — results relevance tint + ranked list

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Context graph

Context graph

PGcustomersΔ .019

entity table — always kept

PGordersΔ .028

12 cols · 48.1K rows

PGtransactionsΔ .041

priceΔ .041

venueΔ .006

+7 more

BQmarket_historyΔ .062

largest movement — load-bearing

S3news_archiveΔ .003

near-zero — safe to ablate

⋯ 4 more scored →

Run finished 00:29 · 14 candidates

Ablation ranking

Δ = mean shift in predicted probability when the candidate is dropped from context.

Load-bearing ↑

BQmarket_historyΔ 0.062

PGtransactions.priceΔ 0.041

PGordersΔ 0.028

KFkalshi_ticksΔ 0.017

Safe to ablate ↓

PGtransactions.venueΔ 0.006

KFnews_wireΔ 0.004

S3news_archiveΔ 0.003

PGcustomers.planΔ 0.001

View all 14 candidates →

💬

Relevance tint uses the signal-blue ramp only (never green/amber/red — those stay reserved for health). Darker fill = larger movement = load-bearing.

Path view — traversal highlighted customers → orders → transactions

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Context graph

Context graph

Path · customers → orders → transactions 2 hops · the join route a context walk takes from the entity table

1

2

PGcustomersstart

customer_id PKint

+9 more

PGorders

customer_id FKint

order_id PKint

+10 more

PGtransactionsend

order_id FKint

pricenumeric

+7 more

RDBcustomers_dedup

5 cols · 11.9K rows

BQmarket_history

9 cols · 41.2M rows

KFkalshi_ticks

9 cols · 128M rows

⋯ 5 more tables →

Hop 1: customer_id · Hop 2: order_id

💬

Tablet · 834px responsive

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Context graph

Context graph

PGcustomers12.4K

customer_id PKint

◈ signup_attimestamptz

+8 more

PGorders48.1K

customer_id FKint

◈ placed_attimestamptz

+10 more

PGtransactions512K

order_id FKint

+8 more

RDBcustomers_dedup

5 cols · 11.9K rows

⋯ 4 more → ⋯ 3 more →

4 of 63 tables

💬

Mobile · 390px responsive

Oil desk ▸ Context graph

Context graph

PGcustomers→

10 cols · 12.4K rows · ◈ signup_at

PGorders→

12 cols · 48.1K rows · ◈ placed_at

PGtransactions→

9 cols · 512K rows · ◈ settled_at

BQmarket_history→

9 cols · 41.2M rows · ◈ bar_ts

KFkalshi_ticks→

9 cols · 128M rows · streaming

S3news_archive→

7 cols · 3.4M rows · ◈ published_at

⌂Home

🗄Sources

◫Graph

💬Chat

⇶Runs

💬

06 · Dataset detail

Loaded — Columns tab

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Datasets ▸ orders

orders

PGraw · oildesk_pg · orders

⬒ 48.1K rows Healthy· synced 1m ago

Overview

Columns

Profile

Lineage

Preview

Stats

Source PG oildesk_pg.orders · "In context" controls whether a column is offered to prediction contexts (mirrors the graph's ablations).

Column	Type	Null %	Distinct	Min	Max	Sample	In context
order_id PK	int	0%	48.1K	100001	148102	121483	On
customer_id FK	int	0%	10.3K	1	12403	8821	On
◈ placed_at	timestamptz	0%	48.1K	2024-01-02 00:14	2026-07-28 13:59	2026-07-11 08:42	On
status	varchar	0%	6	cancelled	shipped	fulfilled	On
total_usd	numeric(12,2)	0%	41.9K	0.00	184920.00	184.99	On
discount	numeric(12,2)	61.2%	1.1K	0.00	500.00	15.00	Off
currency	varchar(3)	0%	4	CAD	USD	USD	On
◈ updated_at	timestamptz	0.1%	48.0K	2024-01-02 00:14	2026-07-28 14:02	2026-07-28 09:14	On
channel	varchar	2.4%	5	app	web	web	Off
quantity_bbl	numeric(10,2)	0%	2.8K	1.00	50000.00	250.00	On
unit_price_usd	numeric(8,2)	0%	9.4K	41.20	128.75	78.14	On
region	varchar	0.8%	11	gulf_coast	west_tx	gulf_coast	On

Showing 12 of 38 columns · 2 ablated

💬

Profile tab

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Datasets ▸ orders

orders

PGraw · oildesk_pg · orders

⬒ 48.1K rows Healthy· synced 1m ago

Overview

Columns

Profile

Lineage

Preview

Stats

Temporal coverage — ◈ placed_at profiled 2026-07-28 09:41

2024-01

2024-09

2025-05

2026-01

48.1K rows

cutoff

now

status — distribution

fulfilled

48%

shipped

25%

pending

16%

cancelled

8%

other

3%

total_usd — histogram

0.00184,920.00

discount — null trend

61.2%

null rate, stable ±0.4% over 30 days

💬

Preview tab — mono data grid

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Datasets ▸ orders

orders

PGraw · oildesk_pg · orders

⬒ 48.1K rows Healthy· synced 1m ago

Overview

Columns

Profile

Lineage

Preview

Stats

Preview reads directly from the source. 8 of 38 columns, first 8 of 100 sampled rows, sampled 2m ago.

order_id	customer_id	◈ placed_at	status	total_usd	quantity_bbl	currency	region
148102	8821	2026-07-28 13:59:41+00	pending	19,535.00	250.00	USD	gulf_coast
148101	10477	2026-07-28 13:58:02+00	pending	4,092.00	52.40	USD	west_tx
148100	902	2026-07-28 13:57:44+00	fulfilled	31,200.00	400.00	EUR	rotterdam
148099	7204	2026-07-28 13:55:19+00	shipped	8,820.00	112.90	USD	gulf_coast
148098	5511	2026-07-28 13:54:03+00	fulfilled	120,499.00	1,540.00	USD	cushing
148097	3087	2026-07-28 13:52:51+00	cancelled	5,400.00	70.00	CAD	alberta
148096	6104	2026-07-28 13:51:12+00	fulfilled	24,075.00	308.50	USD	gulf_coast
148095	8821	2026-07-28 13:50:58+00	fulfilled	1,899.00	24.30	USD	gulf_coast

View all 100 sampled rows → Showing 8 of 100 sampled rows · 8 of 38 columns

💬

Create view — view-builder drawer lands in graph as RDB

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Datasets ▸ orders

orders

PGraw · oildesk_pg · orders

⬒ 48.1K rows

Overview

Columns

Profile

Lineage

Preview

Stats

Column	Type	Null %	Sample
order_id PK	int	0%	121483
customer_id FK	int	0%	8821
◈ placed_at	timestamptz	0%	2026-07-11 08:42
total_usd	numeric(12,2)	0%	184.99

Create view over orders ✕

View name

orders_daily_latest.sqlvalid

1
2
3
4
5

-- one row per customer per day; latest order wins SELECT DISTINCT ON (customer_id, placed_at::date) customer_id, order_id, placed_at, total_usd FROM orders ORDER BY customer_id, placed_at::date, placed_at DESC

Preview31,204 rows · 4 columns · ◈ placed_at kept

💬

Entity-resolution style dedup views (e.g. customers_dedup) are built the same way; the drawer's editor accepts any SQL/RelQL SELECT over this source.

Loading skeleton

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Datasets ▸ orders

orders

PGraw · oildesk_pg · orders

Overview

Columns

Profile

Lineage

Preview

Stats

Column	Type	Null %	Distinct	Min	Max	Sample	In context

💬

Identity band renders immediately from metadata (name, provenance chip, primary actions); async pills and the column statistics shimmer.

Edge case long column name + sample value

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Datasets ▸ news_archive

news_archive

S3raw · oildesk-news-bucket · news_archive

⬒ 3.4M rows Lagging· last sync 6h ago

Overview

Columns

Profile

Lineage

Preview

Stats

Column	Type	Null %	Distinct	Min	Max	Sample	In context
normalized_entity_resolution_ticker_symbol_mapping_confidence	float	12.4%	8.9K	0.0021	0.9998	0.9231	On
article_id PK	uuid	0%	3.4M	0001f8a2-04c1-…	ffff03d9-6e22-…	7f9c2b14-88ad-4c11-9b0e-52a…	On
◈ published_at	timestamptz	0%	3.4M	2024-03-01 00:00	2026-07-28 08:03	2026-07-28 08:03:11+00	On
source_url	text	0.2%	1.2M	https://feeds.e…	https://wire.re…	https://wire.reuters.example/markets/c…	Off

Showing 4 of 7 columns · 1 ablated

💬

Sample values truncate at 24 characters with the full value on hover; long column names ellipsize inside a fixed name column. Nothing wraps — row height stays 36px.

Degraded — stale profile

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Datasets ▸ orders

orders

PGraw · oildesk_pg · orders

⬒ 48.1K rows Healthy· synced 1m ago

Overview

Columns

Profile

Lineage

Preview

Stats

Temporal coverage — ◈ placed_at profiled 2026-07-19 06:12

2024-01

2024-11

2025-09

47.3K rows at profile time

now

status — distribution

fulfilled

47%

shipped

26%

pending

17%

cancelled

10%

total_usd — histogram

0.00181,204.00

💬

Tablet · 834px responsive

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Oil desk ▸ Datasets ▸ orders

orders

PGraw · oildesk_pg · orders

Healthy

Overview

Columns

Profile

Preview

Column	Type	Null %	In context
order_id PK	int	0%	On
customer_id FK	int	0%	On
◈ placed_at	timestamptz	0%	On
status	varchar	0%	On
total_usd	numeric(12,2)	0%	On
discount	numeric(12,2)	61.2%	Off
quantity_bbl	numeric(10,2)	0%	On
unit_price_usd	numeric(8,2)	0%	On

Showing 8 of 38 columns

💬

Mobile · 390px responsive

Oil desk ▸ orders

orders

PGoildesk_pg · 48.1K rows

Healthy

Columns

Profile

Preview

order_id PKOn

int · 0% null · 48.1K distinct

customer_id FKOn

int · 0% null · 10.3K distinct

◈ placed_atOn

timestamptz · 0% null · temporal column

total_usdOn

numeric(12,2) · 0% null · max 184,920.00

discountOff

numeric(12,2) · 61.2% null · ablated

⌂Home

🗄Sources

◫Graph

💬Chat

⇶Runs

💬

07 · Relation studio

Loaded — worksheet open

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Relation studio ▸ customer_spend_90d

Relation studio

worksheet · unsaved changes

Worksheets12

customer_spend_90dRDB

oil_featuresRDB

customer_churn_scoreRDB

dedup_customersRDB

news_sentiment_dailyRDB

View all 12 →

customer_spend_90d.relqlValidRelQL

1
2
3
4

-- expected 90-day spend per customer PREDICT SUM(transactions.price) OVER (90 DAYS FOLLOWING) FROM customers WHERE customers.plan = 'basic' AS OF :anchor RETURN EXPECTED VALUE

Anchor time AS OF :anchor = 2026-06-01

2025-09

2025-12

2026-03

2026-06

◂ drag ▸

now

Context is cut at the anchor; the query scores what happens in the 90 days after it.

last run 0.9s · 2026-07-28 09:38

Results Expected spend over the next 90 days, one row per basic-plan customer.

customer_id	plan	expected_spend_90d
1042	basic	241.80
1047	basic	198.32
1063	basic	176.05
1071	basic	154.99
1088	basic	122.47
1094	basic	96.10
1102	basic	88.63
1115	basic	71.24

View all 1,204 rows → 1,204 customers scored · 0.9s

💬

No skeleton on load: worksheets and the last saved result set come from local cache and render synchronously. Only a fresh Run is async (see running state).

Error — unknown table diagnostic typo

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Relation studio ▸ customer_spend_90d

Relation studio

Worksheets12

customer_spend_90dRDB

oil_featuresRDB

customer_churn_scoreRDB

View all 12 →

customer_spend_90d.relql1 error

1
2
3
4

-- expected 90-day spend per customer PREDICT SUM(transactons.price) OVER (90 DAYS FOLLOWING) FROM customers WHERE customers.plan = 'basic' AS OF :anchor RETURN EXPECTED VALUE

fix the error to run

Resultsstale · from last valid run 09:38

customer_id	plan	expected_spend_90d
1042	basic	241.80
1047	basic	198.32
1063	basic	176.05

View all 1,204 rows →

💬

Running — scoring in progress skeleton

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Relation studio ▸ customer_spend_90d

Relation studio

Worksheets12

customer_spend_90dRDB

oil_featuresRDB

customer_churn_scoreRDB

View all 12 →

customer_spend_90d.relqlRunning

1
2
3
4

-- expected 90-day spend per customer PREDICT SUM(transactions.price) OVER (90 DAYS FOLLOWING) FROM customers WHERE customers.plan = 'basic' AS OF :anchor RETURN EXPECTED VALUE

elapsed 00:07 · 412 / 1,204 entities scored

Resultsrows stream in as entities finish scoring

customer_id	plan	expected_spend_90d
1042	basic	241.80
1047	basic	198.32
1063	basic	176.05
x	x	x
x	x	x
x	x	x

💬

Streaming — "Stream to table" config drawer pipelineable tasks

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Relation studio ▸ oil_features

Relation studio

Worksheets12

oil_featuresRDB

customer_spend_90dRDB

customer_churn_scoreRDB

View all 12 →

oil_features.relqlValid

1
2

-- next-day expected oil settle, per market PREDICT SUM(oil_futures.settle) OVER (1 DAYS FOLLOWING) FROM markets AS OF :t

Stream to table

✕

Keep scoring this query as new data arrives and write every result into a live table.

Task type

Binary — probability per rowok

✓ Regression — expected valueselected

Forecast — value per horizonok

Rankingnot pipelineable

Ranking can't stream: a ranked list needs the full candidate set on every run.

Target table

Created in RelativeDB (RDB) on first write.

Refresh

● Continuous Every 15 min Hourly

Continuous scores each entity as its upstream rows land.

💬

Streaming — triggers news → oil example

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Relation studio ▸ oil_features ▸ Triggers

Triggers

Healthy· streaming to oil_features_live

trigger · oil_move_1sigmaArmed

1
2

-- fire when a scored move breaks one sigma WHEN prediction - baseline > 1.0 * stddev_24h

Publish to

Kafka topic Pub/sub Webhook

Topic

One JSON event per firing entity. Consumers see it in under a second.

How this trigger runs — worked example

A headline lands, the pipeline re-scores tomorrow's oil settle, and a >1σ move publishes an alert.

KFnews_wire

headline arrives09:41:06

pipeline re-scores

1

PREDICT SUM(oil_futures.settle) OVER (1 DAYS FOLLOWING) FROM markets AS OF :t

move > 1σ ? yes · +1.09σ

KFoil-alerts

kafka://oil-alerts

27 fires in the last 30 days · last fire 09:41:07

💬

Streaming — trigger fired event payload

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Relation studio ▸ oil_features ▸ Triggers

Triggers

event · oil-alerts · offset 84,201JSON

1
2
3
4
5
6
7
8
9
10
11

{ "trigger": "oil_move_1sigma", "fired_at": "2026-07-28T09:41:07Z", "entity": "markets/CL-2026Q3", "prediction": 84.62, "baseline": 82.10, "stddev_24h": 2.31, "move_sigma": 1.09, "cause_hint": "news_wire headline 09:41:06", "published_to": "kafka://oil-alerts" }

Recent fireseach row is one published event

Fired at	Entity	Move (σ)
09:41:07	markets/CL-2026Q3	+1.09	→
07:12:44	markets/CL-2026Q4	−1.21	→
yesterday 16:03	markets/BZ-2026Q3	+1.55	→

View all 27 fires →

💬

Tablet · 834px responsive

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Relation studio

Relation studio

customer_spend_90d.relqlValid

1
2
3
4

-- expected 90-day spend per customer PREDICT SUM(transactions.price) OVER (90 DAYS FOLLOWING) FROM customers WHERE customers.plan = 'basic' AS OF :anchor RETURN EXPECTED VALUE

Anchor timeAS OF :anchor = 2026-06-01

2025-12

2026-04

now

customer_id	expected_spend_90d
1042	241.80
1047	198.32
1063	176.05
1071	154.99
1088	122.47

View all 1,204 rows →

💬

Mobile · 390px responsive

Riverbend Trading

Relation studio

RDB

customer_spend_90d ▾

customer_spend_90d.relqlValid

1
2
3
4
5

-- expected 90-day spend PREDICT SUM(transactions.price) OVER (90 DAYS FOLLOWING) FROM customers WHERE customers.plan = 'basic' AS OF :anchor RETURN EXPECTED VALUE

Anchor:anchor = 2026-06-01

now

customer 1042241.80

expected spend, next 90 days · basic plan

customer 1047198.32

expected spend, next 90 days · basic plan

customer 1063176.05

expected spend, next 90 days · basic plan

View all 1,204 rows →

💬

⌂Home

🗄Sources

◫Graph

💬Chat

⇶Runs

08 · Lineage

Loaded — oil_features lineage

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Lineage ▸ oil_features

oil_features

RDB View

◈ anchored on markets.trade_date

edges = last run: Healthy Lagging Failed

Upstream 6

This object

Downstream 2

KFkalshi_ticks

84.1M rows◈ tick_ts

Healthyran 2m ago

S3news_archive

2.2M docs◈ published_at

Laggingran 3h ago

PGorders

2.4M rows◈ created_at

Healthyran 5m ago

⋯ 3 more upstream

RDBoil_features

derived view◈ trade_date

412K rowsrebuilt 09:40

Healthy0.9s

Blends market ticks, news, and order flow into one daily feature row per market.

RDBoil_model_scores

streamed tablecontinuous

Healthywrote 12s ago

KFoil-alerts

kafka://oil-alerts

Healthy27 fires / 30d

💬

Same node language as the schema graph: syschip + a 3px top-border tint say where each object physically lives. Edge color is the health of the last run that moved data along it. No skeleton — lineage renders from cached run metadata.

Failure propagation — upstream failed stale downstream

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Lineage ▸ session_features

session_features

RDB View

Upstream 2

This object

Downstream 2

KFclickstream_events

consumer lag ∞◈ event_ts

Failed43m ago

PGcustomers

184K rows◈ created_at

Healthyran 4m ago

RDBsession_features

derived view◈ session_ts

stale · last good 08:58

RDBchurn_scores_live

streamed tablestale

KFchurn-alerts

kafka://churn-alertspaused

Dimmed = not run since the failure; values shown are the last good ones.

💬

Node hover — path highlight

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Lineage ▸ oil_features

oil_features

RDB View

hovering kalshi_ticks — its full path stays lit

Upstream 6

This object

Downstream 2

KFkalshi_ticks

84.1M rows◈ tick_ts

Healthyran 2m ago

S3news_archive

2.2M docs◈ published_at

Laggingran 3h ago

PGorders

2.4M rows◈ created_at

Healthyran 5m ago

⋯ 3 more upstream

RDBoil_features

derived view◈ trade_date

412K rowsrebuilt 09:40

Healthy0.9s

RDBoil_model_scores

streamed tablecontinuous

Healthywrote 12s ago

KFoil-alerts

kafka://oil-alerts

Healthy27 fires / 30d

kalshi_ticks feeds oil_features → oil_model_scores

💬

Empty downstream — nothing reads this view

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Lineage ▸ dedup_customers

dedup_customers

RDB View

Upstream 2

This object

Downstream 0

PGcustomers

184K rows◈ created_at

Healthyran 4m ago

PGtransactions

9.8M rows◈ txn_ts

Healthyran 4m ago

RDBdedup_customers

derived view◈ created_at

Healthyrebuilt 08:12

Nothing reads this view yet

Schedule it or stream it to a table so other objects can build on it.

💬

Tablet · 834px responsive

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Riverbend Trading ▸ Lineage ▸ oil_features

oil_features

RDB

KFkalshi_ticks

Healthy

S3news_archive

Lagging

PGorders

Healthy

⋯ 3 more upstream

RDBoil_features

◈ trade_date

Healthy

RDBoil_model_scores

Healthy

KFoil-alerts

Healthy

💬

Mobile · 390px responsive

Riverbend Trading ▸ Lineage

oil_features

RDB

Upstream · 6

KFkalshi_ticks Healthy

S3news_archive Lagging

PGorders Healthy

⋯ 3 more upstream

↓

This object

RDBoil_features Healthy

◈ trade_date · rebuilt 09:40

↓

Downstream · 2

RDBoil_model_scores Healthy

KFoil-alerts Healthy

💬

⌂Home

🗄Sources

◫Graph

💬Chat

⇶Runs

09 · Skills

Loaded — editing oil-prices.md unsaved changes

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Oil prices▸ Skills▸ oil-prices.md

Skills

markdown notes agents read first

Project files + New

oil-prices.md modified

kalshi-markets.md clean

data-caveats.md new

A skill is just a markdown file in the project repo. Commits go through git like any other change.

oil-prices.md Edit Preview

# Oil prices — read me first Agents read this before any query touches this project's data. ## Units & joins - kalshi_ticks.price is in cents (int). Divide by 100 before comparing to market_history.settle. - settle lags the tape by 10–20s at close — treat same-minute joins as suspect. ## Time - Always AS OF the quote timestamp, never ingest time. - Anchor before the 14:30 ET settlement window when backtesting daily questions. ## Known noise - news_archive headlines before 2019 are OCR'd; expect noise. - orders.region is free text before 2024-03.

Attached to

📁 Oil prices

Used by agents — read before any query touches this project's data.

Mentions

KFkalshi_ticks BQmarket_history S3news_archive PGorders

Last commit

a41f2c9 · tighten AS OF guidance · danielh · 2d ago

Version history →

💬

No skeleton: skills are small markdown files read from the local git checkout in the same round-trip as the page — the editor renders synchronously.

Preview — rendered markdown

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Oil prices▸ Skills▸ oil-prices.md

Skills

markdown notes agents read first

Project files + New

oil-prices.md clean

kalshi-markets.md clean

data-caveats.md new

oil-prices.md Edit Preview

Oil prices — read me first

Agents read this before any query touches this project's data.

Units & joins

kalshi_ticks.price is in cents (int). Divide by 100 before comparing to market_history.settle.
settle lags the tape by 10–20s at close — treat same-minute joins as suspect.

Time

Always AS OF the quote timestamp, never ingest time.
Anchor before the 14:30 ET settlement window when backtesting daily questions.

Known noise

news_archive headlines before 2019 are OCR'd; expect noise.
orders.region is free text before 2024-03.

Attached to

📁 Oil prices

Used by agents — read before any query touches this project's data.

Last commit

a41f2c9 · tighten AS OF guidance · danielh · 2d ago

💬

Version history — git log for oil-prices.md

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Oil prices▸ Skills▸ oil-prices.md▸ History

Version history

Commit	Message	Author	When
a41f2c9	tighten AS OF guidance	danielh	2d ago	→
9f03b21	add OCR noise caveat for news_archive	danielh	5d ago	→
c77d410	cents → dollars conversion note	mruiz	1w ago	→
5be9a02	split units section out of intro	danielh	2w ago	→
e10ff37	initial skill: oil-prices.md	danielh	3w ago	→

View full history (12 commits)

Each save is a git commit on the project repo — pick any commit to view or restore that version of the file.

💬

Empty — no skills yet

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Oil prices▸ Skills

Skills

Add a skill

A markdown note agents read before touching this project's data — units, caveats, join rules, anything you'd tell a new analyst.

💬

Tablet · 834px responsive

R

⌂

⛁

✦

◈

▶

⚙

Oil prices▸ Skills▸ oil-prices.md

Skills

oil-prices.md kalshi-markets.md data-caveats.md

oil-prices.md Edit Preview

# Oil prices — read me first ## Units & joins - kalshi_ticks.price is in cents (int). Divide by 100 before comparing to settle. - settle lags the tape by 10–20s at close. ## Time - Always AS OF the quote timestamp. ## Known noise

📁 Oil prices a41f2c9 · tighten AS OF guidance · 2d ago

Used by agents — read before any query touches this project's data.

💬

Mobile · 390px responsive

Oil prices▸Skills

Skills

oil-prices.md kalshi-markets.md +1

oil-prices.md Edit Preview

# Oil prices — read me ## Units & joins - kalshi_ticks.price is in cents (int); ÷100 vs settle. ## Time - Always AS OF quote time.

Last commit

a41f2c9 · tighten AS OF guidance · 2d ago

⌂Home

🗄Sources

◫Graph

💬Chat

⇶Runs

💬

10 · Data quality

Loaded — DQ command center

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Oil prices▸Data quality

Data quality

profiled 2026-07-28 06:10

Datasets checked

14

every table this project can query

Columns profiled

212

nulls, cardinality, drift

Open issues

16

2 critical5 warning9 info

Relationships checked

9

FK edges in the schema graph

Table statistics

Top 5 by open issue count — how empty, how varied, how much each table moved this week.

Table	Null %	Cardinality	Drift vs 7d	Issues
KFkalshi_ticks	0.2%	1.2M	↑ 0.8σ	5	→
PGorders	3.1%	812K	↑ 2.1σ	4	→
S3news_archive	6.4%	4.8M	→ 0.1σ	3	→
BQmarket_history	0.0%	96K	↓ 0.3σ	2	→
PGcustomers	1.2%	41K	→ 0.0σ	1	→

View all 14 tables

Relationship statistics

Per join edge — how selective it is, how many children per parent, and how many child rows point nowhere.

Edge	Selectivity	Orphaned FK	Join coverage
orders → customers	0.97	4.1% ↑	95.9%
transactions → orders	0.99	0.2%	99.8%
kalshi_ticks → market_history	0.94	0.0%	98.7%
news_archive → market_history	0.71	1.3%	88.4%

View all 9 edges

Label & target quality

market_settles_yes

The outcome the query predicts — is it balanced, present, and free of future data.

Class balance21.8% yes · 78.2% no

Coverage — anchors with an observed outcome96.4%

Leakage checks 3 columns read post-anchor data

market_history.settle · kalshi_ticks.close_price ·
news_archive.next_day_volume

Label distributionsettle price, ¢ · 12 bins

0¢50¢100¢

Connectivity warnings

Tables the context graph cannot reach from the entity at scoring time.

ContextConnectivityWarning:
support_tickets unreachable from customers within 2 hops at anchor 2026-06-01

ContextConnectivityWarning:
news_archive reaches market_history only via string-match edge (weak)

Open in graph →

💬

Spot-check — inspect 50 random rows

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Oil prices▸Data quality▸Spot check

Inspect 50 random rows

PGorders

Eyeball raw rows the profiler picked at random. Anything you flag becomes an open issue on this table.

order_id	customer_id	total_usd	region	created_at	Your call
88214	c_04412	418.20	midwest	2026-07-12 09:41
90177	c_99999	-12.00	MIDWEST!!	2026-07-24 14:07	flagged · looks wrong
87903	c_01288	96.50	gulf	2026-07-08 16:22
91042	c_07731	1,204.75	gulf	2026-07-26 11:03	marked ok
86518	c_02194	58.10	northeast	2026-06-30 08:15
89660	c_05502	233.90	midwest	2026-07-19 19:48
90881	c_03356	742.00	west	2026-07-25 10:12

Showing 7 of 50 sampled rows · 1 flagged · 1 ok View all 50

Flags feed the open-issues list on the Data quality overview — the flagged row above created the "orders row with negative total" info issue.

💬

Issue detail — drawer open orphaned FK spike

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Oil prices▸Data quality

Data quality

Datasets checked

14

Columns profiled

212

Open issues

16

2 critical5 warning

Edge	Selectivity	Orphaned FK	Join coverage
orders → customers	0.97	4.1% ↑	95.9%
transactions → orders	0.99	0.2%	99.8%
kalshi_ticks → market_history	0.94	0.0%	98.7%

Warning ✕

Orphaned FK spike

orders → customers

orders→customers orphaned FK rate rose from 0.8% to 4.1% after the 07-24 backfill. About 33,400 order rows now reference customer ids that don't exist.

First seen2026-07-24 14:02

Affected queries6 use this edge in context

Severitywarning → critical at 5%

Suggested fix

Re-run the customers backfill for customer_id > c_90412, or exclude orphaned orders from context until it lands.

💬

Loading skeleton

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Oil prices▸Data quality

Data quality

Datasets checked

Columns profiled

Open issues

Relationships checked

Table statistics

Relationship statistics

Label & target quality

Connectivity warnings

💬

Profiling stats are computed server-side per run and can take a few seconds on 14 tables — every async region shimmers; page chrome and card headers render immediately.

Tablet · 834px responsive

R

⌂

⛁

✦

◈

▶

⚙

Oil prices▸Data quality

Data quality

Datasets

14

Columns

212

Open issues

16

259

Table statistics — top 3 by issues

Table	Null %	Drift vs 7d	Issues
KFkalshi_ticks	0.2%	↑ 0.8σ	5
PGorders	3.1%	↑ 2.1σ	4
S3news_archive	6.4%	→ 0.1σ	3

View all 14 tables

Label & target quality 3 columns read post-anchor data

Class balance21.8% yes · 78.2% no

Coverage96.4%

💬

Mobile · 390px responsive

Oil prices▸Data quality

Data quality

Datasets

14

Issues

16

25

PGorders 4 issues

null 3.1% · drift ↑ 2.1σ · orphaned FK 4.1%

KFkalshi_ticks 5 issues

null 0.2% · drift ↑ 0.8σ · card 1.2M

Label quality 3 leaky cols

21.8% yes · coverage 96.4%

View all 14 tables →

⌂Home

🗄Sources

◫Graph

💬Chat

⇶Runs

💬

11 · Training

Configure — binary classification

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Kalshi Oil Desk ▸ Training ▸ order_churn_head

Training

order_churn_head · draft

task queryrelqlvalid ✓

1
2

PREDICT NOT EXISTS(orders.*) FROM customers AS OF :anchor RETURN PROBABILITY

Auto-detected task

Binary classification → probability

The query defines the task and its labels — no hand labeling.

Anchors — where supervision comes from

12 anchors · 2025-07 … 2026-06

Labels = what actually happened after each anchor. Monthly × 12.

2025-07

2026-01

2026-06

now

future — no data

Training mode

Zero-shotno training

Run the backbone as-is. Good first look before spending compute.

0 epochs · run now

Fit headselected

Backbone frozen; trains a small task head on your anchors.

epochs 100 · lr 1e-3

minutes · CPU ok

Full fine-tuneneeds GPU/MPS

Updates the whole backbone. Binary + regression tasks only.

epochs 1 · batch 8 · lr 1e-5 · grad clip 1.0

Backbone checkpoint — smaller is faster, slightly less accurate

fp32reference

344MB · 612ms / score

fp16full quality

172MB · 483ms / score

int8−0.2% AUROC

88MB · 341ms / score

int4−1.1% AUROC

64MB · 268ms / score

Latest evaluation — binary classification run #41 · held-out anchors 2026-05, 2026-06

AUROC

0.87

how well it ranks churners above non-churners

Log loss

0.31

penalty for confident wrong answers — lower is better

Calibration — bars near the diagonal mean probabilities are honest

💬

The query itself is the label pipeline: each hollow ring is a past cutoff, and the model is graded on what really happened after it. Anchors render on the shared violet time spine like everywhere else. No skeleton artboard: the config page renders from the saved draft in one round-trip — the only async region is the loss curve, which gets its own streaming state below.

Variant — forecasting value per horizon

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Kalshi Oil Desk ▸ Training ▸ oil_settle_forecast

Training

oil_settle_forecast · draft

task queryrelqlvalid ✓

1

PREDICT SUM(oil_futures.settle) OVER (7 DAYS FOLLOWING HORIZONS 4) FROM markets

Auto-detected task

Forecasting → value per horizon

Four successive 7-day horizons — one value each, out to 28 days past the anchor.

Zero-shotno training

0 epochs · run now

Fit headselected

epochs 100 · lr 1e-3

minutes · CPU ok

Full fine-tuneunavailable

Binary + regression tasks only — forecasting heads stay on the frozen backbone.

Latest evaluation — error per horizon run #17 · MAE in $/bbl — how far the forecast lands from the settle

Error grows with distance from the anchor — expected. The nearest horizon is within $0.41 of the actual settle.

💬

Variant — ranking ranked list

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Kalshi Oil Desk ▸ Training ▸ next_articles_rank

Training

next_articles_rank · draft

task queryrelqlvalid ✓

1

PREDICT ARRAY_AGG(transactions.article_id) OVER (30 DAYS FOLLOWING RANK TOP 12) FROM customers

Auto-detected task

Ranking → ranked list

Top 12 articles each customer is likely to buy in the next 30 days.

Zero-shotno training

0 epochs · run now

Fit headselected

epochs 100 · lr 1e-3

minutes · CPU ok

Full fine-tuneunavailable

Binary + regression tasks only — ranking always uses the frozen backbone.

Latest evaluation — ranking run #9 · held-out anchors 2026-05, 2026-06

NDCG@12

0.61

how close the predicted list is to what customers actually bought

Hit rate@12

0.44

share of customers with ≥1 correct article in their top 12

A random top-12 list scores NDCG 0.07 on this catalog — the head is well above chance.

💬

Full fine-tune is disabled for ranking (and forecasting): the backbone is only updated end-to-end for scalar binary and regression targets.

Training — running streaming loss

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Kalshi Oil Desk ▸ Training ▸ order_churn_head

Training

order_churn_head · run #42 · fit head · fp16

Loss — streaming epoch 62 / 100 · loss 0.342 ↓

Loss falling means the head is still learning — cancel is safe, progress is checkpointed each epoch. 62% · ~74s left · cpu-local

Claude is watching this run· will report

Run config

Modefit head

Anchors12 · monthly

lr1e-3

Checkpointfp16 · 172MB

StatusHealthy

💬

Training — completed

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Kalshi Oil Desk ▸ Training ▸ order_churn_head

Training

order_churn_head · run #42 · finished 2026-07-28 09:52

AUROC

0.89

+0.02 vs run #41 — better ranking of churners

Log loss

0.27

−0.04 vs run #41 — fewer confident misses

Calibration error

0.021

predicted probabilities track reality closely

Scoring latency

483ms

fp16 checkpoint · per entity

What next

Serve this head — score entities on demand from queries and the workspace.

Stream to table — binary tasks are pipelineable: write scores continuously and fire triggers on conditions.

💬

Edge case leakage — query reads post-anchor data

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Kalshi Oil Desk ▸ Training ▸ order_churn_head

Training

order_churn_head · draft

task queryrelqlleakage ✕

1
2
3

PREDICT NOT EXISTS(orders.*) FROM customers WHERE orders.delivered_at > :anchor -- reads after the anchor AS OF :anchor RETURN PROBABILITY

Where the leak sits

:anchor

read here — off-limits

now

Context must stop at the ring. The hatched span is exactly what the label is made of — training on it would score perfectly and predict nothing.

Override — not recommended

Saves the draft only — training stays blocked until the confirm text matches.

💬

Blocked-by-default with a typed second confirm: leakage is the one mistake this product exists to prevent, so the primary action is disabled rather than warned.

Tablet · 834px responsive

R

⌂

🗄

⏱

◫

▤

ƒ

⇶

⚙

Kalshi Oil Desk ▸ Training ▸ order_churn_head

Training

task queryBinary classification → probability

1
2

PREDICT NOT EXISTS(orders.*) FROM customers AS OF :anchor RETURN PROBABILITY

Anchors12 anchors · 2025-07 … 2026-06

now

Labels = what actually happened after each anchor.

Fit headselected

epochs 100 · lr 1e-3 · fp16 172MB · 483ms

minutes · CPU ok

AUROC

0.87

Log loss

0.31

Cal. error

0.021

💬

Mobile · 390px responsive

Training ▸ order_churn_head

Training

Binary → probability

1
2

PREDICT NOT EXISTS(orders.*) FROM customers AS OF :anchor

Anchors · 12 · monthly

now

2025-07 … 2026-06 · showing every 2nd

Fit headselected

epochs 100 · lr 1e-3 · CPU ok

AUROC

0.87

Log loss

0.31

⌂Home

🗄Sources

◫Graph

💬Chat

⇶Runs

💬

12 · Workspace

Loaded — thread with query, chart, table

R

⌂

🗄

⏱

◫

▤

ƒ

💬

⇶

⚙

Kalshi oil spread

2m ago

Backfill + retrain

1h ago

Oil trading triggers

3h ago

Churn look

yesterday

Weekly orders digest

2d ago

View all 14 →

Kalshi Oil Desk ▸ Workspace ▸ Kalshi oil spread

show me the kalshi spread for tomorrow's oil prices, four time horizons, and the expected price at those

Reading kalshi_ticks KF joined through markets S3, series OIL. I split the next day into four horizons and predicted the expected mid at each:

relqlran in 1.2s

1
2
3
4

PREDICT AVG(kalshi_ticks.mid) OVER (6 HOURS FOLLOWING HORIZONS 4) FROM markets WHERE markets.series = 'OIL' RETURN EXPECTED VALUE

Horizon	Spread	Expected
h+6h	0.42	78.10
h+12h	0.55	78.24
h+18h	0.61	78.33
h+24h	0.68	78.42

Spread widens with distance — the market is less sure about tomorrow evening than tomorrow morning. Expected mid drifts up $0.32 across the day.

Pin as live card 📌

This thread is the ONE scrollable surface in the product — every other page is fixed-height with bounded lists. The 💬 helper in the bottom-right of every other page opens this same chat and history; there is no FAB here because chat IS the page.

Task kickoff — backfill then retrain

R

⌂

🗄

⏱

◫

▤

ƒ

💬

⇶

⚙

Kalshi oil spread

2m ago

Backfill + retrain

just now

Oil trading triggers

3h ago

View all 14 →

Kalshi Oil Desk ▸ Workspace ▸ Backfill + retrain

backfill the news archive and retrain the oil model when done

Started both. The backfill pulls the remaining news_archive S3 partitions; the retrain of oil_settle_forecast is queued behind it so it trains on the complete archive. I'll post here as each finishes.

Backfilling news_archive · 34% · 2.1M of 6.2M docs Pause View →

Retrain oil model · queued · starts when backfill completes View →

Task chips are live objects: they keep updating in the thread and also appear on Runs. Chaining ("when done") is expressed as a queued chip, not hidden scheduler state.

Advanced — ablation, trigger, live pinned card

R

⌂

🗄

⏱

◫

▤

ƒ

💬

⇶

⚙

Kalshi oil spread

2m ago

Backfill + retrain

1h ago

Oil trading triggers

3h ago

View all 14 →

Kalshi Oil Desk ▸ Workspace ▸ Oil trading triggers

Pinned · livelive

OIL 1-day expected

78.42▲0.6σ

updated 12s ago

which inputs actually matter for the oil model?

Ran EXPLAIN ABLATE — it re-scores the query once per candidate input and ranks by how far dropping each one moves predictions (took 31s):

news_sentiment is load-bearing: dropping it moves predictions 2.3σ.

RDnews_sentiment

2.3σ

KFkalshi_ticks

1.1σ

BQmarket_history

0.4σ

S3news_archive

0.1σsafe to ablate

alert me when the 1-day expected moves more than 1 sigma

Forecasting tasks are pipelineable, so I can stream scores and fire a trigger on that condition:

Publishing to kafka://oil-alerts when |Δ| > 1σ — confirm?

Side effects (Kafka triggers, ablations) always round-trip through an explicit confirm chip — the assistant proposes, the user commits.

Streaming — assistant mid-answer

R

⌂

🗄

⏱

◫

▤

ƒ

💬

⇶

⚙

Kalshi oil spread

now

Backfill + retrain

1h ago

View all 14 →

Kalshi Oil Desk ▸ Workspace ▸ Kalshi oil spread

show me the kalshi spread for tomorrow's oil prices, four time horizons, and the expected price at those

Reading kalshi_ticks KF for series OIL. Splitting tomorrow into four horizons and scoring the expected mid at each▍

Running PREDICT· scoring 4 horizons · 0.8s

Empty — new thread

R

⌂

🗄

⏱

◫

▤

ƒ

💬

⇶

⚙

New thread

now

Kalshi oil spread

2m ago

Backfill + retrain

1h ago

View all 14 →

Kalshi Oil Desk ▸ Workspace ▸ New thread

Ask about your data

Try: "which customers look likely to churn this month?"

Chart oil settle vs kalshi mid Backfill news_archive What moved the oil model today?

Tablet · 834px responsive

R

⌂

🗄

⏱

◫

▤

ƒ

💬

⇶

⚙

☰ Threads ▸ Kalshi oil spread

show me the kalshi spread for tomorrow's oil prices, four time horizons, and the expected price at those

Four horizons over the next day for series OIL:

spread 0.42 → 0.68 · expected 78.10 → 78.42

Mobile · 390px responsive

☰

Kalshi oil spread

Kalshi Oil Desk

+

show me the kalshi spread for tomorrow's oil prices, four time horizons, and the expected price at those

Four horizons over the next day for OIL:

h+6h 0.42 → 78.10
h+12h 0.55 → 78.24
h+18h 0.61 → 78.33
h+24h 0.68 → 78.42

Pin as live card 📌

alert me if it moves more than 1 sigma

Publishing to kafka://oil-alerts when |Δ| > 1σ — confirm?

⌂Home

🗄Sources

◫Graph

💬Chat

⇶Runs

Chat is the primary mobile surface: the full app is reachable through it, so the thread gets full height and the tabbar's Chat item is active by default.

13 · Runs

Loaded — running rows live

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Runs

Runs

All 412

Jobs 64

Tasks 96

Pipelines 38

Ingestions 214

Name	Type	Status	Where	Started	Duration
finetune_oil_impact	Task	Running	EKS us-east-1	2026-07-28 13:41	21m 12s
kalshi_ws_ingest	Ingestion	Running	RelativeDB Cloud	2026-07-28 14:00	2m 09s
orders_ingest	Ingestion	Healthy	on-prem k8s	2026-07-28 12:00	14m 33s
daily_feature_prep	Pipeline	Failed	GKE europe-west4	2026-07-28 06:00	32m 04s
warehouse_sync	Job	Lagging	on-prem k8s	2026-07-28 05:30	58m 47s

View all 412 runs →

💬

The Where column names the cluster a run executes on with the same placement chips as the Compute page, so a run can be traced to its metal in one glance.

Loading skeleton

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Runs

Runs

All

Jobs

Tasks

Pipelines

Ingestions

Name	Type	Status	Where	Started	Duration
n	t	s	w	d	d
n	t	s	w	d	d
n	t	s	w	d	d
n	t	s	w	d	d
n	t	s	w	d	d

💬

Five skeleton rows at table rhythm; the filters stay interactive while the table loads.

Empty — no runs yet

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Runs

Runs

All

Jobs

Tasks

Pipelines

Ingestions

No runs yet.

Runs appear here once a source syncs or a task executes.

💬

Training-run drawer — loss, lr, AUROC on the spine

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Runs

Runs

All

Jobs

Tasks

Pipelines

Ingestions

Name	Status	Where
finetune_oil_impact	Running	EKS us-east-1
orders_ingest	Healthy	on-prem k8s
daily_feature_prep	Failed	GKE europe-west4

finetune_oil_impact Running

EKS us-east-1 fit_head · epochs 100 · lr 1e-3

Training curves log loss learning rate

plateau · lr cut 1e-3 → 5e-4

ep 0 · 13:41

ep 48 · 13:52

ep 84

now

AUROC (val)

0.874

Log loss

0.312

Learning rate

5e-4

Steps

assemble_context 4m 41s

train_head running · 16m 31s

💬

Loss and learning rate share one chart in the neutral ramp; the x-axis is the violet spine because it maps epochs to wall-clock. AUROC, loss, and lr get plain-label stat readouts under the chart.

Agent monitor — Claude watching the run state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Runs

Runs

All

Jobs

Tasks

Pipelines

Ingestions

Name	Status	Where
finetune_oil_impact	Running	EKS us-east-1
orders_ingest	Healthy	on-prem k8s

finetune_oil_impact Running

✳ Agent monitor · Claude attached watching

Recent agent actions

14:02 observed val loss plateau 3 epochs → reduced lr 1e-3 → 5e-4

14:09 early-stop armed (patience 5 epochs)

14:11 watching val AUROC · no drift

Every agent action is written to the run's audit trail and is reversible — pausing stops interventions, not observation.

Steps

train_head running · 30m 12s

💬

An attached agent is a card, not a hidden daemon: its actions log in mono with timestamps, and the human keeps pause/detach controls plus an audit trail.

Failed-run drawer — log excerpt

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Runs

Runs

All

Jobs

Tasks

Pipelines

Ingestions

Name	Status	Where
finetune_oil_impact	Running	EKS us-east-1
daily_feature_prep	Failed	GKE europe-west4
orders_ingest	Healthy	on-prem k8s

daily_feature_prep Failed

Run timeline

06:00

06:24

06:32 · failed

Steps

extract_orders 14m 12s

join_payments 12m 40s

write_feature_store failed · 5m 12s

06:31:48 INFO writing partition dt=2026-07-27

06:31:55 WARN retry 3/3 on s3://acme-fs/features/

06:32:03 ERROR write_feature_store: AccessDenied

06:32:03 ERROR s3 credentials expired at 06:00:00

06:32:04 INFO run marked failed

💬

Failed steps auto-expand their last log lines; the failed step's bar renders critical on the spine — state color as state, never as decoration.

Tablet · 834px responsive

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸Runs

Runs

All

Jobs

Tasks

Pipelines

Ingestions

Name	Status	Where	Started
finetune_oil_impact	Running	EKS us-east-1	13:41
kalshi_ws_ingest	Running	RelativeDB Cloud	14:00
orders_ingest	Healthy	on-prem k8s	12:00
daily_feature_prep	Failed	GKE europe-west4	06:00
warehouse_sync	Lagging	on-prem k8s	05:30

View all 412 runs →

💬

Mobile · 390px responsive

acme-supply-chain

Runs

All

Tasks

Pipelines

Ingestions

finetune_oil_impact Running

EKS us-east-1 13:41 · 21m 12s

kalshi_ws_ingest Running

RelativeDB Cloud 14:00 · 2m 09s

daily_feature_prep Failed

GKE europe-west4 06:00 · 32m 04s

warehouse_sync Lagging

on-prem k8s 05:30 · 58m 47s

💬

⌂Home

⛁Sources

◫Graph

💬Chat

▶Runs

14 · Compute & deployment

Loaded — clusters

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

▦

⚙

Acme Supply Chain▸Compute

Compute & deployment

Where jobs run, what it costs, and how to move them.

on-prem k8sus-dc-1

Healthy

Open source on your own infra, running the open RT-J reference checkpoints.

Running jobs14

GPU12 / 16

CPU61%

$0/hr (owned hardware) Details →

AWS · EKSus-east-1

Healthy

Elastic GPU pool for training bursts.

Running jobs9

GPU · A1006 / 8

CPU48%

$41.20/hr Details →

GCP · GKEeurope-west4

Healthy

EU-resident serving for European entities.

Running jobs5

GPU · L42 / 4

CPU33%

$18.60/hr Details →

RelativeDB CloudManaged

Healthy

Managed inference on checkpoints we train and serve ourselves — tuned beyond the open RT-J reference weights.

Running jobs11

Capacityautoscales

Serving p5038ms

$0.84 / 1M predictions Details →

💬

Four placements, one vocabulary: the same chips appear on Runs and Observability. Capacity bars use the neutral chart ramp — utilization is a quantity, not a state.

Deploy streaming job — choose flow 1/2

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

▦

⚙

Acme Supply Chain▸Compute

Compute & deployment

Deploy streaming job · step 1 of 2

✕

Pipeline to deploy

oil_impact_streamnews_s3 → oil_impact_scores · fires kafka://oil-alerts reorder_forecastorders_pg → reorder_predictions kalshi_edge_streamkalshi_ws → market_edges

Target cluster

on-prem k8s EKS us-east-1 GKE europe-west4 RelativeDB Cloud

💬

Deploy streaming job — review flow 2/2

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

▦

⚙

Acme Supply Chain▸Compute

Compute & deployment

Review deployment · step 2 of 2

✕

oil_impact_stream → GKE europe-west4

What moves, what it costs

Reads news_s3 (us-east-1) cross-region~2.1 GB/day egress

Estimated egress cost~$0.19/day

Serving checkpoint rtj-oil-ft int8 copy1.8 GB once

Compute on GKE+$2.10/hr

💬

Review names the egress before the button: cross-region reads are the hidden cost of moving compute, so the estimate sits between choose and deploy. Move compute and Offload data reuse this same choose → review shell.

Moving — job mid-migration state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

▦

⚙

Acme Supply Chain▸Compute

Compute & deployment

Moving oil_impact_stream EKS us-east-1→GKE europe-west4

Draining on EKS · warming on GKE · no dropped records — the old placement serves until the new one is caught up 62%

checkpoint copied · consumer offsets synced · replaying 14:01:12 → now

AWS · EKSus-east-1

Healthy

Running jobs9 → 8

GPU · A1006 / 8

GCP · GKEeurope-west4

Healthy

Running jobs5 → 6

GPU · L43 / 4

💬

Loading skeleton

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

▦

⚙

Acme Supply Chain▸Compute

Compute & deployment

on-prem k8sus-dc-1

Checking…

Running jobs14

GPU

CPU

AWS · EKSus-east-1

Checking…

Running jobs9

GPU · A100

CPU

GCP · GKEeurope-west4

Checking…

Running jobs5

GPU · L4

CPU

RelativeDB CloudManaged

Checking…

Running jobs11

Capacity

Serving p50

💬

Cluster identity and job counts come from config and render at once; only the capacity bars stream in from each cluster's metrics endpoint, so only they shimmer. Health shows "Checking…" until the first heartbeat.

Tablet · 834px responsive

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

▦

⚙

Acme Supply Chain▸Compute

Compute & deployment

on-prem k8sus-dc-1

Healthy

14 jobs · GPU 12/16 · CPU 61%$0/hr

AWS · EKSus-east-1

Healthy

9 jobs · GPU 6/8 A100 · CPU 48%$41.20/hr

GCP · GKEeurope-west4

Healthy

5 jobs · GPU 2/4 L4 · CPU 33%$18.60/hr

RelativeDB CloudManaged

Healthy

11 jobs · autoscales · p50 38ms$0.84/1M pred

💬

Mobile · 390px responsive

acme-supply-chain ▸ Compute

Compute

on-prem k8sHealthy

14 jobs · GPU 12/16$0/hr

AWS · EKSHealthy

9 jobs · GPU 6/8$41.20/hr

GCP · GKEHealthy

5 jobs · GPU 2/4$18.60/hr

RelativeDB CloudHealthy

11 jobs · managed$0.84/1M

💬

⌂Home

⛁Sources

◫Graph

💬Chat

▶Runs

15 · Settings

Loaded — General tab

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Settings

Settings

General

API tokens

Integrations

Preferences

Project

Name

Slug

acme-supply-chainCopy

Description

Created

2026-02-11 09:04

Version

2.14.0

Metadata

region = us-west
env = production

Members

Name	Email	Last active
Daniel Henneberger you	daniel@acme.com	now
Anna Lee	anna@acme.com	2h ago
Mark Shah	mark@acme.com	1d ago
Priya Nair	priya@acme.com	3d ago

Danger zone

Deleting a project removes its sources, tasks, and runs. Type the project slug to confirm.

💬

No skeleton for Settings: form values load with the page, so there is no async region to shimmer. Also no time spine — Settings is the one operational page with no temporal reasoning, deliberately.

Permission denied — Viewer read-only degraded

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Settings

Settings

General

API tokens

Integrations

Preferences

Project

Name

Slug

acme-supply-chain

Members

Name	Email	Last active
Anna Lee	anna@acme.com	2h ago
Mark Shah	mark@acme.com	1d ago
Daniel Henneberger	daniel@acme.com	now
Priya Nair you	priya@acme.com	now

💬

Viewers see values but every control disables; the banner says which role unlocks editing. The danger zone hides entirely for non-Admins.

API tokens tab — revealed once + revoke confirm state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Settings

Settings

General

API tokens

Integrations

Preferences

Token created

rql_live_4f2a9c81e7d340bb92c6a1f08d5e7723

Copy it now — it won't be shown again.

API tokens

Name	Token	Scope	Created	Expires
ci-deploy	rql_live_4f2a… Copy	read-write	2026-07-28	2027-07-28
notebooks	rql_live_b81c… Copy	read-only	2026-03-02	Expires in 3d
legacy-etl	rql_live_09dd…	read-write	2025-11-19	2026-11-19	Revoke this token? Anything using it stops working.

💬

The full token appears exactly once after creation. Revoke asks inline — the confirm names the consequence, and the button says what it does.

Invite member — pending row state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Settings

Settings

General

API tokens

Integrations

Preferences

Members

Name	Role	Email	Last active
Invited · 0m ago	Viewer	sofia@acme.com	Resend
Anna Lee		anna@acme.com	2h ago
Mark Shah		mark@acme.com	1d ago
Priya Nair		priya@acme.com	3d ago

💬

On send, the pending row appears with the dot-less status word "Invited · 0m ago" in ink-soft and a Resend text button. An invalid email keeps the form open: "That doesn't look like an email address."

Tablet · 834px — General responsive

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸Settings

Settings

General

API tokens

Integrations

Preferences

Project

Name

Slug

acme-supply-chainCopy

Description

Members

Name	Role	Last active
Daniel Henneberger you	Admin	now
Anna Lee	Admin	2h ago
Mark Shah	Data Engineer	1d ago
Priya Nair	Analyst	3d ago

💬

Mobile · 390px — General responsive

acme-supply-chain

Settings

General

Tokens

Integrations

Project

Name

Slug

acme-supply-chainCopy

Members

Invite

Daniel Henneberger youAdmin

Anna LeeAdmin

Mark ShahData Engineer

Priya NairAnalyst

💬

⌂Home

⛁Sources

◫Graph

💬Chat

▶Runs

16 · Alerts

Loaded — leakage alert prominent

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Alerts

Alerts

All alerts 5

Incidents 1

Notifications 12

Alert	Severity	Source	Triggered	Status
oil_impact_1sigma	Warning	stream	1m ago	Open
payments_final_timeout	Critical	ingestion	2m ago	Open
⧗ churn_model_leakage Leakage	Critical	schema	9m ago	Open
daily_sales_delay	Warning	pipeline	1h ago	Investigating
schema_drift_orders	Info	schema	5h ago	Open

⧗ churn_model reads payments_final after anchor time — a temporal-correctness alert, not infrastructure.

💬

Severity owns color, status owns lifecycle. Unacknowledged Critical rows carry the 2px critical left edge — the only row-level color, and only until acknowledged. The ⧗ glyph + violet badge marks leakage.

Loading skeleton

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Alerts

Alerts

All alerts

Incidents

Notifications

Alert	Severity	Source	Triggered	Status
a	s	s	t	s
a	s	s	t	s
a	s	s	t	s
a	s	s	t	s
a	s	s	t	s
a	s	s	t	s

💬

Six skeleton rows; the severity filter renders immediately so triage can be scoped while rows load.

Empty — all clear state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Alerts

Alerts

All alerts

Incidents

Notifications

All clear

No open alerts. Rules are evaluating on every profile run.

💬

Alert detail drawer — leakage timeline

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Alerts

Alerts

All alerts

Incidents

Notifications

Alert	Severity	Triggered
payments_final_timeout	Critical	2m ago
⧗ churn_model_leakage	Critical	9m ago
daily_sales_delay	Warning	1h ago

churn_model_leakage Critical

Leakage

Open Investigating Closed

Timeline

cutoff

read

now

12:00 anchor

14:03

churn_model reads payments_final after anchor time — orders.settled_at is read 2h before its availability cutoff. Add a backdating rule or exclude the column.

task churn_model dataset payments_final run task_run_88121

Activity

Rule leakage_guard fired · 13:54

Priya set status to Open · 13:56

💬

Leakage drawers render the anchor (hollow caret), the offending column's availability window as violet hatching, and the leaking read as a critical tick inside the hatch.

Non-leakage drawers show a short trigger-time-vs-now spine per spec — same drawer shell, so no separate artboard is drawn for them.

All filtered out state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Alerts

Alerts

All alerts

Incidents

Notifications

No alerts match these filters.

💬

Alert rules — metric · condition · route

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Alerts

Alerts

All alerts

Incidents

Notifications

Rules 6

A rule watches a metric, checks a condition over a window, and routes wherever you work.

Metric	Condition	Window	Route	Status
oil_impact_score	moves > 1σ vs 24h baseline	5m	kafka oil-alerts#trading	Active	→
connector kalshi_ws	lag > 60s	1m	page on-call	Active	→
reorder_prediction	expected stockout < 7 days	1h	email owner#ops-alerts	Active	→
query p95 latency	> 500ms	10m	#eng-perf	Active	→
leakage_guard	any read after anchor	per run	page on-call#ml-platform	Active	→

View all 6 rules →

💬

Rules mix infra (connector lag) with prediction conditions (a score moving 1σ) — the same table grammar covers both, and routes are chips: Kafka topic, Slack channel, email, pager.

Rule builder drawer — condition editor state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Alerts

Alerts

All alerts

Incidents

Notifications

Rules

Metric	Condition
oil_impact_score	moves > 1σ vs 24h baseline
connector kalshi_ws	lag > 60s
reorder_prediction	expected stockout < 7 days

Edit rule Active

Name

Streaming query

oil_impact.rqlregression · streams to a table

1
2
3

PREDICT AVG(prices.wti_return) OVER (1 DAY FOLLOWING) FROM headlines RETURN EXPECTED VALUE

Fire when

prediction moves > 1σ vs 24h baseline

Evaluated every 5m on the streamed prediction · one alert per entity · cooldown 15m.

Routes

+ #trading (Slack) + Add route

💬

The rule watches a streaming RelQL prediction, so the query gets the IDE shell — violet on the window keywords because they are temporal. The trigger itself is config, not SQL: fire when the streamed prediction moves past 1σ. "Test on last 24h" replays it before it goes live.

Fired trigger — news → oil impact state

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸ Alerts▸ oil_impact_1sigma

Alerts

All alerts

Incidents

Notifications

Rules

Alert	Severity	Triggered
oil_impact_1sigma	Warning	14:02
payments_final_timeout	Critical	14:01
daily_sales_delay	Warning	13:02

Predicted oil impact +1.4σ Warning

Triggering headline

"OPEC+ signals surprise output cut at emergency session"

news_wire · 2026-07-28 14:01:52 · entity WTI

Prediction delta

+1.4σ

13:00

14:01:52

now

Published to kafka://oil-alerts · 14:02:07

rule oil_impact_1sigma pipeline oil_impact_stream notified #trading

💬

The fired trigger tells the whole story: the headline that moved the model, the score jump on the violet spine (hollow ring = the headline's anchor), and the mono receipt of where and when it published.

Tablet · 834px responsive

R

⌂

⛁

⛓

✦

◈

◎

❯

▶

◉

⚠

⚙

Acme Supply Chain▸Alerts

Alerts

All alerts

Incidents

Notifications

Rules

Alert	Severity	Triggered
oil_impact_1sigma	Warning	1m ago
payments_final_timeout	Critical	2m ago
⧗ churn_model_leakage Leakage	Critical	9m ago
daily_sales_delay	Warning	1h ago
schema_drift_orders	Info	5h ago

View all alerts →

💬

Mobile · 390px responsive

acme-supply-chain

Alerts

All

Incidents

Rules

oil_impact_1sigma →

Warning 1m ago · Open

payments_final_timeout →

Critical 2m ago · Open

⧗ churn_model_leakage →

CriticalLeakage 9m ago · Open

daily_sales_delay →

Warning 1h ago · Investigating

💬

⌂Home

⛁Sources

◫Graph

💬Chat

▶Runs