Worked examples
Eight domains, one workflow.
Each was built end-to-end against the live platform — a description, a dataset, a verified build, and queries that return proof-carrying Amber Reports. The forecasting domains pull live data straight from connectors.
Click any example to see the actual code.
Decision domains · verified & proof-carrying
verifiedLoan approval
Age, debt-to-income, credit-score tiers▸
Loan approval
Age, debt-to-income, credit-score tiers
# Rules in plain English become an ontology domain = api.domains.create(name="Loan Approval Assessment", description="Applicants must be 18 or older; debt-to-income " "must not exceed 43%; minimum credit score per loan " "tier; employment verification required.") api.datasets.upload(domain_id=domain["id"], file_path="loan_applications.csv") api.domains.build_ontology(domain["id"]) platform = api.platforms.create(domain_id=domain["id"], dataset_id=ds["id"], verified_profile=True, verified_min_confidence=0.6) report = api.platforms.query(platform["id"], query="Can a 17-year-old with no income apply for a £3,000 loan?")
Live result >>> Check Applicant Age Below Threshold fired · proof_checked: true
“Decision independently certified against the trusted kernel: 1 rule fired, 0 facts derived from 1 input fact.”
“Decision independently certified against the trusted kernel: 1 rule fired, 0 facts derived from 1 input fact.”
verifiedInsurance fraud
Policy limits, early claims, high-frequency, collusion▸
Insurance fraud
Policy limits, early claims, high-frequency, collusion
domain = api.domains.create(name="Insurance Claims Fraud Detection", description="Claims exceeding the policy limit must be denied; " "claims filed within 90 days of policy start flagged; " "claimants with 4+ prior claims flagged high-frequency...") report = api.platforms.query(platform["id"], query="A property flood claim for £55,000 on a £50,000 policy. Approve?")
Live result >>> Check Policy Limit Exceeds Threshold (+2 rules) · proof_checked: true
“The claim of £55,000 exceeds the policy limit of £50,000 — it should not be approved in full.”
“The claim of £55,000 exceeds the policy limit of £50,000 — it should not be approved in full.”
verifiedClinical prescribing
Interactions, contraindications, allergy cross-reactions — derived▸
Clinical prescribing
Interactions, contraindications, allergy cross-reactions — derived
# The platform DERIVES the clinical knowledge from primitive facts domain = api.domains.create(name="Clinical Prescribing Safety", description="From the drug class, eGFR, concurrent medication and " "allergy, derive drug interactions, contraindications and " "allergy cross-reactions; then block, require a dose " "adjustment, flag, or permit.") report = api.platforms.query(platform["id"], query="Prescribe ibuprofen to an 82-year-old with chronic kidney disease?")
Live result >>> Is Contraindicated (derived) → block · proof_checked: true
“Decision independently certified against the trusted kernel: 3 rules fired, 2 facts derived from 7 input facts.” The same NSAID at normal kidney function permits — the contraindication is derived, not column-matched.
“Decision independently certified against the trusted kernel: 3 rules fired, 2 facts derived from 7 input facts.” The same NSAID at normal kidney function permits — the contraindication is derived, not column-matched.
verifiedRecruitment compliance
Age, salary bands, right-to-work, bias detection▸
Recruitment compliance
Age, salary bands, right-to-work, bias detection
domain = api.domains.create(name="HR Recruitment Compliance", description="Applicants must be 18+; salary within the role band; " "right to work required; references must pass; flag " "qualified candidates rejected without interview as bias.") report = api.platforms.query(platform["id"], query="Can we hire a 16-year-old for a junior developer role?")
Live result >>> Check Age Below Threshold fired · proof_checked: true
“There is a rule flagging applicants below 18 — a potential compliance issue for this hire.”
“There is a rule flagging applicants below 18 — a potential compliance issue for this hire.”
verifiedEnvironmental compliance
Regulatory / permit breaches, protected zones▸
Environmental compliance
Regulatory / permit breaches, protected zones
domain = api.domains.create(name="Environmental Regulatory Compliance", description="Readings must not exceed the regulatory or permit " "limit; facilities need an active permit; coastal-protected " "zones have stricter limits; breaches require enforcement.") report = api.platforms.query(platform["id"], query="NO2 recorded at 44.8 ug/m3 against a 40.0 limit. In breach?")
Live result >>> Calculate Measurement Value To Regulatory Limit Ratio · proof_checked: true
“Riverside Chemical Works is in breach — NO2 of 44.8 ug/m3 exceeds the 40.0 regulatory limit.”
“Riverside Chemical Works is in breach — NO2 of 44.8 ug/m3 exceeds the 40.0 regulatory limit.”
verified · 20k rowsAccess-governance PDP
Classify-then-conclude policy chain — permit / deny, every decision proof-carrying▸
Access-governance PDP
Classify-then-conclude policy chain — permit / deny, every decision proof-carrying
# 20k access requests. The policy is a chain: classify raw fields into named # conditions (device trust, privilege, restricted zone), then conclude permit or # deny — including the cross-field rule "deny when clearance < target sensitivity". # The request is supplied as structured facts — the facts ARE the certified base. domain = api.domains.create(name="Access Governance PDP", description="…") platform = api.platforms.create(domain_id=domain["id"], dataset_id=ds["id"], verified_profile=True, verified_min_confidence=0.6) report = api.platforms.query(platform["id"], query="Should this be permitted?", facts={"clearance_level": 2, "target_sensitivity": 3, "access_type": "write", "mfa_passed": True})
Live result Decision: DENY · proof_checked: true
“Decision: Deny — access denied due to insufficient clearance level (clearance 2 is below the target sensitivity 3).”
Held-out: 100% decision accuracy, 100% certification across 50 requests.
“Decision: Deny — access denied due to insufficient clearance level (clearance 2 is below the target sensitivity 3).”
Held-out: 100% decision accuracy, 100% certification across 50 requests.
Command & control · verified at scale
verified · 20k rowsAir-track identification & triage
Recognized air picture — clear / monitor / escalate, human in the loop▸
Air-track identification & triage
Recognized air picture — clear / monitor / escalate, human in the loop
# 20k tracks; the request is supplied as structured facts (the facts ARE # the certified base). Policy stated in the description — incl. the safety rule: # "emergency tracks must always be escalated to a human operator." domain = api.domains.create(name="Air Track Triage", description="…") api.datasets.upload(domain_id=domain["id"], file_path="air_tracks.csv") platform = api.platforms.create(domain_id=domain["id"], dataset_id=ds["id"], verified_profile=True, verified_min_confidence=0.6) report = api.platforms.query(platform["id"], query="Triage this track.", facts={"iff_mode": "emergency", "squawk_emergency": True, "in_restricted_zone": False, "flight_plan_correlated": False})
Live result Triage: ESCALATE · proof_checked: true — “Decision independently certified against the trusted kernel.”
Held-out: 98% three-way accuracy, 98% certification across 50 tracks; emergency tracks always escalate to an operator.
Held-out: 98% three-way accuracy, 98% certification across 50 tracks; emergency tracks always escalate to an operator.
verified · ISR standardsTriage on standardised ISR sensor data
Same policy, 26-field ASTERIX/MISB schema — the platform picks the drivers▸
Triage on standardised ISR sensor data
Same policy, 26-field ASTERIX/MISB schema — the platform picks the drivers
# The SAME clear / monitor / escalate policy as the air-track demo above — # but over a 26-field ISR feed (ASTERIX Cat 021/062 + MISB ST 0601). You # describe the policy once; the platform finds the few fields that decide it. domain = api.domains.create(name="Air Track Triage (High-Spec ISR)", description="…") api.datasets.upload(domain_id=domain["id"], file_path="air_tracks_hispec.csv") platform = api.platforms.create(domain_id=domain["id"], dataset_id=ds["id"], verified_profile=True, verified_min_confidence=0.6) report = api.platforms.query(platform["id"], query="Triage this track.", facts={"position_source": "fused", "iff_mode": "emergency", "emergency_squawk": True, "in_restricted_zone": True, "track_confidence": 0.93})
Live result Triage: ESCALATE · proof_checked: true — “Decision independently certified against the trusted kernel.”
Standardised ISR schema — ASTERIX Cat 021/062 surveillance fields, MISB ST 0601 sensor metadata, per-sensor confidence; synthetic, reproducible ISR data. 26 input columns, but only a handful drive the verdict — the platform does the feature selection from a wide, sparse schema, no feature engineering required. Emergency tracks always escalate to a human operator.
Standardised ISR schema — ASTERIX Cat 021/062 surveillance fields, MISB ST 0601 sensor metadata, per-sensor confidence; synthetic, reproducible ISR data. 26 input columns, but only a handful drive the verdict — the platform does the feature selection from a wide, sparse schema, no feature engineering required. Emergency tracks always escalate to a human operator.
Forecasting · explainable macro models
FRED · explainable10-year Treasury yield — system-selected drivers
24 macro series in, autoregression off — the platform picks what explains the yield▸
10-year Treasury yield — system-selected drivers
24 macro series in, autoregression off — the platform picks what explains the yield
# Upload a broad, neutral 24-series US macro panel (bundled, public-domain FRED) api.datasets.upload(domain_id=domain["id"], file_path="gs10_macro_panel.csv") # Train with autoregression OFF — the model must explain via macro drivers, not momentum cfg = api.predictions.create_config(platform["id"], target_field="GS10", time_index_field="date", horizon=1, frequency="monthly", autoregressive="none") # What did the system select? Feature importance + readable WHEN-THEN rules base = api.predictions.predict(platform["id"], prediction_config_id=cfg["id"]) sf = api.predictions.symbolic_forecast(platform["id"], prediction_config_id=cfg["id"])
System-selected drivers From 24 candidates the platform kept corporate credit (AAA, BAA), the short end (FEDFUNDS, GS1), and real activity (retail sales, building permits, housing starts) — credit conditions + the real economy, surfaced from a neutral list, not hand-picked.
Fit Level-space R² ≈ 0.97 with 0% own-history reliance (411 monthly rows, 1992–2026). The month-to-month change R² is negative — and we say so rather than hide behind the flattering level number.
Rules 41 readable WHEN→THEN driver rules with contributions and hit-rates — e.g. "WHEN the front end falls AND real activity weakens → yield down". Rules you can read, challenge, and govern.
FRED · neuro-symbolicNeuro-symbolic 10-year yield forecast
A tight fit, a what-if you drive, and an honest neural-vs-symbolic head-to-head▸
Neuro-symbolic 10-year yield forecast
A tight fit, a what-if you drive, and an honest neural-vs-symbolic head-to-head
# 1) Conditional forecast — YOU supply the macro view; the gate shows which rules fire f = api.predictions.symbolic_forecast(platform["id"], prediction_config_id=cfg["id"], feature_overrides={"FEDFUNDS": 6.0}) # "what if the Fed stays at 6%?" # f["forecast"] -> 4.45% ; f["why"] -> the rules, with fired_on_latest_row=True for the 3 that fire # 2) Does the symbolic layer earn its place? Preview the discovered rules read-only. cmp = api.predictions.neurosymbolic_comparison(platform["id"], prediction_config_id=cfg["id"], include_pending=True) # what-if preview of accepted-but-unapproved rules
In-sample fit The symbolic forecaster tracks the 10-year yield closely one step ahead — corr 0.98 over a walk-forward backtest on real FRED data (1959–2026).
What-if — you supply the macro view Plug your own driver expectations via
feature_overrides: setting fed funds to 6% lifts the implied 10-year from 4.32% → 4.45%, and the gate names the 3 rules that fired (e.g. “WHEN CPI is rising AND fed funds is rising → yield up 0.13”). A mid-range reading fires nothing and the forecast holds at baseline — conditional and honest.Neural vs neuro-symbolic Forecasting 6 months out, the discovered correction rules are A/B-tested and held for expert approval; previewed read-only they lift R² 0.54 → 0.69 and cut RMSE 18%. The chart above tracks all three series month by month — the correction rule fires on ~25% of months (the amber dots); where it fires the neuro-symbolic line is pulled toward actual, otherwise the forecast equals the neural model. At a 1-month horizon the neural model is near-ceiling and discovery keeps no rules — the layer only adds rules that beat the backtest.
FRED · explainableS&P 500 — system-selected macro drivers
25 macro series in, autoregression off — the platform picks what explains the index▸
S&P 500 — system-selected macro drivers
25 macro series in, autoregression off — the platform picks what explains the index
# Live-fetch a 25-series monthly macro panel from FRED (rates, inflation, labour, ...) panel_csv = fetch_fred_panel(api_key) # resamples daily+monthly to clean monthly api.datasets.upload(domain_id=domain["id"], file_path=panel_csv) # Train with autoregression OFF — explain the index through macro, not momentum cfg = api.predictions.create_config(platform["id"], target_field="SP500", time_index_field="date", horizon=1, frequency="monthly", autoregressive="none") # What did the system select? base = api.predictions.predict(platform["id"], prediction_config_id=cfg["id"]) sf = api.predictions.symbolic_forecast(platform["id"], prediction_config_id=cfg["id"])
System-selected drivers From 25 candidates the platform kept housing starts and corporate credit spreads ahead of the 2-year yield, consumer sentiment, CPI and money supply — a real-economy-and-credit story, not the "rates and the VIX" a hand-picked model would have started with.
Fit Level-space R² ≈ 0.79 with 0% own-history reliance (~118 monthly rows). Honest: the month-to-month change is near-zero predictable, and the small sample is stated. The process is the product.
Rules Readable WHEN→THEN macro rules with contributions — rules you can challenge and govern, not a single number from a black box.
FRED + Coinbase · explainableBitcoin — system-selected crypto + macro drivers
ETH + 25 macro series in, autoregression off — is Bitcoin more macro than crypto?▸
Bitcoin — system-selected crypto + macro drivers
ETH + 25 macro series in, autoregression off — is Bitcoin more macro than crypto?
# Upload a bundled monthly panel: BTC + ETH (Coinbase) + 25 US macro series (FRED) api.datasets.upload(domain_id=domain["id"], file_path="btc_macro_panel.csv") # Train with autoregression OFF — explain BTC through the panel, not momentum cfg = api.predictions.create_config(platform["id"], target_field="BTC_USD", time_index_field="date", horizon=1, frequency="monthly", autoregressive="none") # What did the system select? base = api.predictions.predict(platform["id"], prediction_config_id=cfg["id"]) sf = api.predictions.symbolic_forecast(platform["id"], prediction_config_id=cfg["id"])
System-selected drivers From 27 candidates (ETH + 25 macro) the platform kept FEDFUNDS, business loans (BUSLOANS), corporate credit (BAA), ETH, capacity utilisation (TCU), housing starts, and industrial production — the Fed and credit rank ABOVE Ethereum. Bitcoin is more macro than crypto.
AR actually hurts With BTC's own history included (AR=full), R² = 0.57. Without it (AR=none), R² = 0.62 — Bitcoin's momentum is noise that actively degrades the model. The macro drivers are the signal.
Rules 47 readable WHEN→THEN rules — e.g. "WHEN industrial production falls AND consumer credit tightens → BTC up" (flight-to-alternative). 119 monthly rows, 2016–2026, 0% own-history reliance. Discovery honestly accepts 0 correction rules — the macro drivers already explain Bitcoin.
FRED · symbolicGBP/USD — readable rules that beat the random walk
UK + euro + US macro in; a SYMBOLIC-only model out — no neural net, and on cable it beats persistence▸
GBP/USD — readable rules that beat the random walk
UK + euro + US macro in; a SYMBOLIC-only model out — no neural net, and on cable it beats persistence
# Pull 18 UK + euro + US macro series LIVE from the FRED connector — one call, merged by date ds = api.datasets.fetch(domain_id=domain["id"], connector_type="fred", config={"series_ids": PANEL, "api_key": key, "start_date": "1999-01-01"}) # Forecast cable from a SYMBOLIC-ONLY model: induced WHEN->THEN rules, no neural network sf = api.predictions.symbolic_forecast(platform["id"], prediction_config_id=cfg["id"]) sf["skill_vs_persistence"] # +0.19 — beats a last-value baseline on a walk-forward backtest sf["why"] # 20 readable rules: contribution, hit-rate, support
Symbolic skill The rules-only model backtests at +0.19 skill vs persistence on monthly cable — it beats the "next month ≈ this month" random walk that FX almost never lets you beat. The neural model on the same panel scores −0.21. Here the readable symbolic layer isn't just more explainable — it's more accurate.
What drives sterling From 18 candidates the rules lean on two UK signals — business confidence and the OECD UK leading indicator — with cameos from US & euro 10-year yields and UK industrial production. Deteriorating UK momentum → sterling weaker; improving → stronger. Surfaced from a neutral panel, not hand-picked.
# a sample of the 20 induced rules — each carries a contribution, hit-rate and support count WHEN UK confidence falling AND UK leading indicator falling → GBP/USD down 0.022 (hit 0.70, n=33) WHEN UK confidence 3m-change low AND 12m-deviation low → GBP/USD down 0.020 (hit 0.77, n=47) WHEN UK confidence jumps from a depressed level → GBP/USD up 0.017 (hit 0.86, n=7) WHEN UK industrial production running above trend → GBP/USD up 0.006 (hit 0.63, n=65)
Honest by design The point forecast sits at the persistence baseline; the rules form the adjustment band. When no rule's condition is met — as at the latest reading — the forecast holds rather than inventing a move. Skill is a walk-forward backtest over 329 months (1999–2026), not a live trading signal.