The research that finds the edge now runs itself.
HtAG Analytics is known for proprietary market metrics — the Relative Composite Score™, Growth Spillover Effect and Growth Pattern Deviation among 150+ metrics across 7,000+ suburbs. Every one of them began as correlation research: events tested against data, metric against metric, until an edge survived. This loop industrialises that R&D — a live daily orchestrator moves external signals through a fixed evidence ladder, while the data team spends its hours on what only people can build: new metrics, new APIs, new features and products.
The edge is proprietary. Now the pipeline that finds it is, too.
HtAG’s metric families — RCS™, GSP, GPD — are not in the standard feeds, and its market-selection methodology carries a documented record of 135 recommendations without a loss. Each of those metrics began the same way: substantial research hunting correlations between events and data, between metric and metric, until something survived testing.
The work this used to take
Scanning twelve signal categories — supply, credit, rents, migration, infrastructure, climate, policy, labour — then structuring, de-duplicating and validating what survived. A standing research function, run out of the data team’s calendar between everything else it builds.
The ledger the loop now carries
365 runs a year: 52 weekly signal scans, 26 QA passes, 12 validation rounds against live HtAG data, 4 synthesis labs — 94 research stages on cadence, none of them dependent on anyone’s spare hours, and a heartbeat on every other day.
Where the hours went
The data team builds instead — new metrics, new endpoints on the developer platform, new features and products. The first product seed to come out of the loop, an internal validation-queue app, is already in the store awaiting operator approval.
For the businessSome consequences resist a single number — this one compounds instead. Every validated finding is a candidate metric, skill or product, and the pipeline that used to run on stolen hours now never misses a week.
A learning system that refuses to guess.
Built inside Morpheus for HtAG Analytics. Most vendors would not publish a workflow like this until it had results. The absence of results is the part worth showing: it is what a discipline looks like before it pays off.
| Question | Answer |
|---|---|
| What starts it | A daily schedule at 21:00 UTC — 0 21 * * *. It wakes whether or not there is work to do. |
| What it decides first | Which cadence is due today: weekly, fortnightly, monthly, quarterly, or none. |
| What it produces | Structured hypotheses with an expected direction, the metrics that would test them, a risk caveat and a validation window. |
| What it holds | One shared evidence store — the canonical state carried between runs, and the reason a Sunday run knows what the previous month concluded. |
| What it may never do | Publish a claim, produce client-facing material, launch a product, or overwrite the underlying skills. All four sit behind an operator gate. |
| In the store now | Twelve open hypotheses, moving through four evidence tiers. |
| Proof status | Customer zero — HtAG Analytics, running to its cadences on the live instance. |
Four rhythms, each bounded by what it may not conclude.
The router checks the date and runs only what is due, in dependency order. Every stage is bounded by what it is forbidden from concluding.
Weekly Watchtower
- Scans external market, product and research signals
- Turns each into a structured hypothesis
- Records source, quality and expected direction
Hypothesis QA
- Deduplicates the register
- Rejects anything that cannot be tested
- Maps each survivor to exact metrics and a direction
Validation engine
- Tests prepared hypotheses against live market data
- Records the query evidence for each test
- Moves evidence tier up, or leaves it where it is
Research and product lab
- Synthesises only repeated or backtested evidence
- Looks for research edge and product opportunities
- Scores candidates against published weights
And on every other day
A heartbeat. It verifies the evidence store is intact, writes a run note saying no cadence was due, and invents nothing. A scheduled agent with nothing to do should produce a record of having done nothing — not filler.
Why the state is shared
All four cadences read and write one evidence store. That file is the operating memory: it is what makes the monthly stage able to test a hypothesis the weekly stage formed three weeks earlier, and what stops the quarterly stage synthesising from work nobody validated.
Four tiers. Only the last one may be spoken aloud.
Every hypothesis carries a tier, and the tier decides what may be done with it. Nothing is promoted by age, confidence or repetition of the same source.
Observed
An external signal has been seen and structured. It is a hypothesis, not a finding.
Internal only. No claim of any kind.Supported once
A single validation, or coherence with the internal data baseline.
Internal only. Still not a position.Repeated
Validated across more than one refresh or market. The signal has survived time.
Operator-reviewed. Still not public.Backtested
Repeatedly or consistently validated. The only tier eligible to become something a client sees.
Eligible to be proposed — and only proposed.Twelve hypotheses, under test.
Read from the live evidence store: what is under test, what it covers, and the discipline every hypothesis must pass before it may be used.
What the twelve cover
Construction supply, lending serviceability, rental vacancy, planning policy, climate insurance cost, migration, infrastructure corridors, regional imbalance, planning reform, auction liquidity, labour conditions and build capacity.
Why half are blocked
Six were formed from search extracts rather than the primary report. The workflow records that as a reason not to proceed, instead of treating a secondary summary as evidence.
What each one carries
An expected direction, the exact metrics that would confirm or refute it, an explicit risk caveat, and a window inside which it must be tested or reconsidered.
A system that had reached a conclusion in its first week would be a system with no gate. Twelve open questions and no answers is the correct state for a loop that has run for days rather than quarters, and saying so is cheaper than explaining it later.
Absent evidence is recorded as absent.
The most common way an automated research system misleads its owner is by treating a missing check as a passed one. This loop is built so that cannot happen quietly.
The rule
Monthly validation must record a live data access log for each test, or mark itself blocked or failed. There is no implied validation and no default pass.
What it looks like in the record
The most recent run states plainly that no live market data was accessed and no data tools were used — because no cadence stage was due that day. The absence is written down rather than left to inference.
Access was tested separately
A live smoke test returned real results from the market-data tools for a real locality, and carries an explicit caveat: testing access once does not excuse any scheduled run from logging its own.
Four things this workflow may never decide.
The gate is not a review step at the end. It is a standing constraint the workflow carries into every stage, including the ones that run while nobody is watching.
No public claim
Nothing it concludes reaches the outside world without a person approving it.
No client-facing output
It cannot produce material that goes to a client, at any tier.
No product launch
It may score and propose a product idea. It may not start one.
No skill overwrite
It cannot change the underlying analysis logic it runs on.
What it is allowed to propose
One product idea currently sits in the store, scored against six published weights covering client pain, data advantage, repeatability, distribution fit, evidence defensibility and build complexity. The score and the weights are both recorded. The idea remains a proposal.
Decisions waiting on a person
Four are open in the record, including approving the live data connector for scheduled validation and ratifying the Tier 3-only rule for public claims. The workflow lists what it needs a human to decide rather than proceeding on assumption.
The tiering discipline itself. The loop’s durable output is not a market view — it is a standing rule about what evidence is required before anyone in the business is allowed to say something. That rule now applies to work the loop never touched.
What self-learning should mean → The other build: intake to draft → All eight domains this runs beside →
What people ask about this build.
Answered from the evidence store, including where the answer is unflattering.
What does fail closed mean here?
Monthly validation must record a live data access log or mark itself blocked or failed. There is no path where an absent data connection is treated as a silent pass, and no implied validation.
Can this workflow publish a market claim?
No. Public claims, client-facing claims, product launches and changes to the underlying skills all sit behind an operator approval gate. Only Tier 3 evidence is even eligible to be proposed.
Is the schedule proven, or just switched on?
Switched on and running to its cadences — every run leaving files that can be inspected.