In pilot on our own brands · Phase 0
The growth intelligence layer that remembers what actually worked.
HypaBrand keeps an evidence-linked memory of each brand — what it is, what it has tried and what happened — and turns it into governed recommendations, content and tests across organic, paid, site and commerce.
One loop, one memory
Connect, understand, create, deploy, measure, learn, improve.
Every stage reads from the Brand Brain and writes evidence back into it. A feature earns its place by closing part of the loop.
Official APIs for social, ads, analytics, store and email.
The Brand Brain: identity, audience, voice, claims, offers and learnings, each with evidence.
Ideas, briefs, hooks, scripts, captions and media grounded in the Brain and past performance.
Approved content is published; paid changes are prepared paused and executed only after approval.
Organic, paid, site and commerce metrics joined, with attribution shown as separate lenses.
Observation → hypothesis → experiment → result → learning, graded by rule.
Proposals ranked by evidence and expected impact; a daily brief of what changed and why.
Connect
Read scopes first; write scopes only when a feature needs them. File imports where an API is closed. Tokens go straight into the vault.
Improve feeds the Brand Brain, and the next Connect starts from a smarter brand.
Brand Brain
Eight kinds of record. Never one bucket.
Each record has exactly one kind with its own schema, lifecycle and approval rule. Grade, confidence and estimand travel with every evidential record into the interface and into every model prompt.
- Fact Something sourced, asserted or extracted, with its as-of date. A person verifies it.
- Observation A measured change against a baseline and band, with sample size and confounders checked. Detected by deterministic code.
- Hypothesis A candidate cause or idea, stated so it can be tested. Proposed; never promoted by repetition.
- Decision A choice a named person made, with the evidence cited. A person makes it.
- Experiment A pre-registered test with a stated estimand. A person launches it.
- Result The computed outcome of an experiment, with its interval. Only the deterministic evaluator.
- Learning What the brand now knows, graded and linked to its results. A person accepts it.
- Rule Machine-readable policy Brand Guard, the ledger and the Analyst enforce. A person activates it.
Evidence grades, by rule
- G0 Unsupported. An idea, an AI suggestion, a platform recommendation not yet tested
- G1 Anecdotal. One post, one comment, one campaign — may illustrate, never justifies
- G2 Descriptive. Measured aggregates, driver decompositions, uncontrolled before/after
- G3 Associational. Comparisons meeting minimum samples with the confounder checklist
- G4 Quasi-experimental. Designs with passed diagnostics; calibrated models
- G5 Randomised. Pre-registered randomised test with a stated estimand
- G6 Replicated. Two or more independent G5 results in the same direction
Who holds authority
A hypothesis never becomes a fact by repetition, age or model confidence. Results are computed only by the deterministic evaluator; people verify facts, make decisions, launch experiments, accept learnings and activate rules.
Confidence is computed from the grade and its modifiers — imprecision, risk of bias — and only ever moves a claim down. Nothing unlocks causal wording except the design of the test.
Analyst
Answers “why” without inventing numbers or causes.
A grounded analyst that runs deterministic, pre-specified analyses and narrates the results it did not compute.
The questions it must answer
- Why is this campaign deteriorating?
- Which creative characteristics correlate with performance?
- Where in the funnel are we losing customers?
- What changed this week?
How it stays honest
- Runs deterministic, pre-specified analyses instead of improvising.
- Every number in its narrative comes from a tool output, never from the model’s own arithmetic.
- Grades its own claims by rule, and its wording follows the grade.
- Says “insufficient evidence” plainly, with what it would take to decide.
- Remembers every look it has taken and every learning the brand has accepted.
Studio and Calendar
From idea to learning in one pipeline.
Channel-native drafts grounded in the Brain and the brand’s own performance history, checked by Brand Guard at draft, at approval and in pre-flight. Nothing leaves without approval.
- Ideas
- Brief
- Hook
- Script
- Creative
- Caption
- Approval
- Calendar
- Publish
- Measure
- Learn
Every first-person claim traces to an accepted Brain entry — a story, a fact, an identity or an opinion — or the draft cannot reach review.
One idea, the right version per channel. Variants validate against a format registry; nothing is blindly cross-posted.
A calendar that tells the truth: pre-flight checks, held states and alerts with the time at risk, so silent failures do not happen.
Measurement
Every lens kept separate. Every number states its lens.
Every source tells a different story about the same money. HypaBrand shows where they disagree and why, states which lens answers which question, and turns large disagreements into experiments.
- Meta-attributed Meta Ads reporting Credits itself with conversions it may not have caused.
- GA4-attributed Google Analytics 4 Under-counts after consent loss and ad blockers.
- Store Shopify orders Knows the revenue but not the cause.
- Survey Post-purchase survey Customers misremember how they found you.
- Blended Ratios across sources Hides which channel works; only where the policy registry allows.
Experiments
Pre-registered tests, honest results.
Observation → hypothesis → experiment → result → learning, graded by rule. Underpowered tests are recorded as inconclusive; a learning reaches the top grade only when replicated.
A stated estimand before the test starts: descriptive, associational, policy effect, incremental effect or mechanism.
Results only from the evaluator. No agent and no model computes a result; the deterministic experiment calculator does.
Confidence downgrades, never upgrades: imprecision or risk of bias lowers a claim; nothing unlocks causal wording except the design.
Agents
Agents propose. People decide. Deterministic code executes.
A small roster of chartered agents, each with a written charter, a fixed toolset, an autonomy level and an evaluation suite. No agent calls a provider write API.
- Analyst Answers questions about performance with evidence; weekly review. Answers with evidence cards; proposed observations and hypotheses
- Sentinel Watches for anomalies, fatigue, pacing, broken tracking and connection failures. Observations and alerts
- Strategist Turns observations and learnings into hypotheses and experiment designs. Hypothesis and experiment proposals
- Producer Drafts briefs, hooks, scripts, captions and channel variants. Content proposals a person approves
- Curator Proposes Brain entries from evidence; raises challenges; keeps syntheses fresh. Proposed Brain entries and challenges
- Scout Processes competitor captures into patterns — never copies. Proposed market insights
L0 · Observe
Read and compute only.
L1 · Propose
Insights, entries, hypotheses and plans — nothing prepared for execution.
L2 · Prepare
Execution-ready drafts and paused structures that a human approves.
L3 · Act within policy
A narrow action class, only where the founder has enabled a policy, with one-click undo.
No autonomous budget, bid or campaign-status changes until you explicitly enable a policy for them. An “optimise within an envelope” level is not planned.
Security and custody
Assume any component can be compromised. Make sure it cannot reach everything.
HypaBrand holds authorisations that can spend money, speak as a brand, change a store and read customer data. Blast-radius minimisation is the design principle.
Row-level security, forced and fail-closed
Every tenant row carries its workspace. PostgreSQL row-level security is enabled and forced with restrictive policies that fail closed when no tenant context is set. Each brand is its own workspace; nothing crosses between brands silently.
Envelope-encrypted credentials
Provider tokens are sealed with a per-credential data key, wrapped by a key-encryption key that exists only in the worker. The web tier can never decrypt a token. Tokens are never logged, never sent to a model, never in fixtures.
Raw first, with provenance
Every provider payload used for analytics is preserved with its provenance before normalisation, so figures can be recomputed from source — except where a provider’s data-use policy caps storage, and then only references persist.
Agents propose; people decide; code executes
No agent calls a provider write API. Approved actions run through a deterministic executor with policy checks, audit records and rollback data. No autonomous budget, bid or campaign-status changes until a policy is explicitly enabled.
Integrations
Official APIs only. Read first.
Write scopes only when a feature needs them; scopes stored and checked per call; revocation detected and handled. Every provider’s data-use policy is enforced by a registry, not remembered by people.
- Instagram Organic Profile, media, insights and account insights through the official Graph API.
- Meta Ads Paid Structure, insights, change history and diagnostics; actions prepared paused.
- Google Analytics 4 Site Standard report cubes; model use only after opt-in consent.
- Shopify Commerce Orders, products and analytics cubes from a custom app on your store.
- LinkedIn Reporting only Member post analytics shown live; values cached for 48 hours, references only.
Built for Australian brands
AU first. UK and EU next.
Incumbents are US-first. HypaBrand starts where its founder’s brands are.
- Australian English, AUD and Australia/Sydney time in every figure; UK and EU next.
- Designed to run in Sydney: the database and application regions are ap-southeast-2 and syd.
- GST-aware profit maths and multi-currency from the start, because incumbents are US-first.
- Notifiable Data Breaches scheme and Australian Privacy Principles considered in the security design.
See it working.
The portal opens without a sign-in while the pilot runs. Accounts arrive with Clerk.