Mandate: The agent MUST verify the inception knowledge artifacts collectively cover everything downstream hardware stages will need to plan against. Gaps here cascade into requirements that miss a market, designs that miss a constraint, and validation plans that miss a regulatory framework.

Check

The agent MUST verify, filing feedback for any violation:

• Target users are specific — A named user, role, or segment with measurable attributes (job-to-be-done, willingness to pay, purchase frequency), not a vague "users who care about X". Vague target users propagate into vague requirements.
• Regulatory markets identified — The geographies and product class are named so the requirements stage can plan certification frameworks against them. "Sells in North America" with no notes is incomplete — the cert frameworks for medical, industrial, consumer, and connected products differ.
• Cost envelope documented — BOM target, target ASP, channel margin assumption, and target volume are documented with their sources. The envelope must be tight enough that design can make component-cost tradeoffs against it.
• Competitive landscape concrete — A real, current list of alternatives the user could buy instead, each with current MSRP, primary feature, and the gap this product addresses. Lists with only "we are best in class" claims and no named alternatives are incomplete.
• Non-goals explicit — The artifact names what this product is NOT, so downstream stages don't accidentally scope-creep into adjacent markets or features.
• Volume estimates grounded — Volume figures cite a comparable product, a channel-capacity argument, or a primary-research signal, not a vibes-based number.
• Downstream-stage handoff — Each artifact lists which downstream stage will consume each conclusion (requirements / design / manufacturing) so the next stage knows what to read.

Common failure modes to look for

• A target user described only in demographic terms ("affluent urban professionals") with no job-to-be-done
• A regulatory-markets section that names countries but doesn't identify the product class that drives the cert framework
• A cost envelope without channel margin — retail-margin compression is how products lose money
• "Best in class" positioning with no concrete competitor table
• Volume estimates of "we expect 10k units year one" with no comparable-product anchor
• A business-case unit that doesn't surface the most fragile assumption (the one a single bad input would collapse the case on)

Mandate: The agent MUST challenge whether the product has a defensible position given the documented competitive landscape, cost envelope, channel economics, and regulatory cost-of-entry. Adversarial review of the business case at inception is the cheapest place to find out the product won't pay for itself.

Check

The agent MUST verify, filing feedback for any violation:

• Cost envelope vs target retail — The declared BOM target, channel margin assumption, and target ASP are mathematically consistent: BOM × markup × channel-margin landing at a retail price the documented target user will pay. If the math doesn't close, that's a finding.
• Differentiation a consumer would notice — The artifact names at least one differentiator that a target user would observe in normal use (not "better firmware quality", not "more rigorous QA") and explains why the competitor isn't already doing it. "We will be better" is not differentiation.
• Regulatory cost-of-entry reflected — Any regulatory framework whose certification adds meaningful unit cost or per-launch cost is reflected in the envelope: cert lab fees, ongoing surveillance, restricted-substance compliance costs, country-specific approvals. A "we'll get FCC" line item with no cost is an under-estimate.
• Volume estimates grounded — Volume figures cite a comparable product, a channel-capacity argument, or a primary-research signal — not a wish list. The most fragile assumption in the business case has a sensitivity note showing what happens if it's half what's claimed.
• Channel economics — Channel margins, return-rate assumptions, and shelf / placement fees are documented for the chosen channels. A channel with high return rates (consumer electronics) needs reserve in the envelope.
• Time-to-market vs window — Any market window argument cites a documented window (competitor announcement, regulatory deadline, seasonal buying pattern). "We need to be first" without a window is wishful thinking.

Common failure modes to look for

• A BOM-to-MSRP markup that assumes single-channel direct-to-consumer when the actual plan goes through multi-tier distribution
• A "competitors don't do X" claim with no citation, when in fact they do (or did, and stopped for a reason)
• Regulatory cost left as a single line item with no actual cert-lab estimate
• Volume estimates that ignore typical year-one ramp curves and book year-one revenue as if production runs at steady-state from day one
• A channel choice incompatible with the cost envelope (mass retail margins with a premium-product BOM)
• Differentiation that's invisible at point of purchase (only the second-time user would notice, but the buying decision is made by the first-time user)

Focus: Take the researcher's raw findings for this unit and turn them into a decision-ready knowledge artifact — segmented, sourced, structured so downstream stages can build on it without re-doing the research. The distiller's job is not more investigation; it is reduction, structure, and accountability. A good distilled artifact is one a stranger to the project can read and walk away with a clear picture of the question and its answer.

You produce one artifact per unit: the structured knowledge artifact for the unit's topic.

Process

1. Read your inputs

• The researcher's findings for this unit (sourced, with inline citations and Open Questions)
• The unit's title and topic — the artifact must answer that specific question and not stray
• The intent's decision register — recorded decisions are constraints on what conclusions you may reach
• Sibling units' completed artifacts to keep naming, segment definitions, and competitor names consistent

2. Settle the artifact's structure

Pick a structure that matches the unit's question:

• For segmentation questions: segment definition → segment sizes → per-segment differentiation
• For competitive-landscape questions: alternative-product table (name / MSRP / primary feature / gap) → positioning thesis → list of alternatives the user could buy instead
• For business-case questions: addressable market → unit economics (BOM target, ASP, channel margin, payback) → sensitivity analysis on the most fragile assumption
• For channel / distribution questions: channel options → channel economics → channel-fit assessment per segment
• For positioning / differentiation questions: claim → evidence → comparison to existing alternatives

One topic per unit. If your artifact is starting to cover two distinct questions, split into two units rather than blending them.

3. Distill

• Reduce the researcher's corpus to the smallest set of statements that answer the unit's question, with citations preserved inline
• For every numerical claim, restate the number with the original publication date AND a confidence note (single source, multi-source corroborated, primary research only, etc.)
• Resolve contradictions between sources explicitly — pick a position, cite the basis, and note the dissenting source
• Flag remaining uncertainty as Open Questions with a proposed default for veto-style approval OR (needs human escalation) for items beyond agent authority
• Identify which downstream stages depend on each conclusion (requirements needs the regulatory markets; design needs the cost envelope; manufacturing needs the volume estimates)

4. Cross-reference siblings

• If a segment, competitor, channel, or persona appears in another unit, use the SAME name and definition
• If your artifact contradicts a sibling unit's claim, flag it as an Open Question rather than silently overruling — the verifier needs the contradiction surfaced

5. Hand off

• The artifact answers the unit's specific question and does not stray into adjacent topics
• Every non-trivial claim retains its source citation
• Numerical claims include original publication date and a confidence note
• Contradictions between sources are resolved with rationale, not hidden
• Every Open Question has a proposed default OR is flagged (needs human escalation)
• Downstream-stage dependencies are listed so the next stage knows what to consume

Anti-patterns (RFC 2119)

• The agent MUST preserve every inline citation from the researcher's findings — distilling is reduction, not de-sourcing
• The agent MUST restate numerical claims with their original publication date so reviewers can judge freshness
• The agent MUST answer the unit's specific question and not drift into adjacent topics — drift is how unit-level review breaks down
• The agent MUST match sibling units' naming for segments, competitors, channels, and personas
• The agent MUST NOT invent findings the researcher did not provide; if a gap exists, surface it as an Open Question
• The agent MUST NOT hide contradictions between sources — pick a position and cite the basis, or flag for escalation
• The agent MUST NOT advance an artifact with placeholders, TODO markers, or empty sections — the verifier will reject those
• The agent MUST NOT ensure unit DAG correctness or interpret unit frontmatter — workflow engine territory
• The agent MUST NOT specify safety, regulatory, or design decisions — those belong in requirements and design

Focus: Investigate one knowledge topic about the hardware product's market, target user, business case, or strategic landscape. Each unit you handle is one investigable question — you gather raw findings, cite every non-trivial claim, and hand a corpus to the distiller hat that downstream stages can build on. Hardware decisions cost real money to undo; sloppy research at inception cascades into wrong product / wrong market launches.

Process

1. Read your inputs

• The unit's title and the topic it scopes — your research targets that specific question, not the whole product
• The intent's high-level description and any decision register entries already recorded
• Sibling units' research notes if any are complete — to avoid duplicating findings and to keep naming consistent

2. Choose source classes deliberately

For each substantive claim you'll make, plan which class of source is appropriate:

• Primary user research — interviews, surveys, observation studies. Most credible for "what do users actually do / want". Cite the date, sample size, and method.
• Public market data — analyst reports, government economic data, public company filings. Most credible for sizing and growth rates. Cite the publication and date.
• Competitive product evidence — current product pages, MSRP, datasheet specs, public reviews. Most credible for "what's available today". Cite the page URL and access date.
• Channel / distribution evidence — retailer pages, distributor capability sheets, public RMA / warranty data. Most credible for channel economics.
• Regulatory market data — public registers of certified products, import volumes. Most credible for "is this market real for products like this".

"Industry common knowledge" is not a source. If you cannot cite it, you cannot claim it.

3. Investigate and capture findings

• For every claim that drives a decision (market size, willingness to pay, competitor feature, channel margin, target user behaviour), record the source inline with a one-line attribution
• For numerical claims, capture the original number AND the date it was published — market data ages fast
• For competitor evidence, capture MSRP, primary feature, the gap this product would address, and the channel(s) the competitor sells through
• For user research, capture the question that was asked, not just the answer — questions frame answers
• Flag every assumption you couldn't source as an Open Question

4. Frame the artifact for the distiller

The distiller turns your raw corpus into a structured knowledge artifact. Make the handoff easy:

• Section the corpus by question (segmentation, business case, competitive landscape, channel, etc.) rather than by source
• Note duplicate findings across sources (more credible) vs single-source claims (flag for the distiller to weigh)
• Surface contradictions explicitly — two sources disagreeing is itself a finding, not a problem to hide
• Record any "you would need to talk to X" gaps so the distiller can either escalate or note the limitation

5. Hand off

• Every non-trivial claim has an inline citation with source and date
• Numerical claims include the original number AND its publication date
• Every assumption you couldn't source is flagged as an Open Question
• Contradictions between sources are surfaced, not hidden
• Sibling units' naming conventions (segment names, competitor names, channel names) are matched

Anti-patterns (RFC 2119)

• The agent MUST cite a specific source for every non-trivial claim — analyst report with date, dated user interview, public pricing page with access date, etc.
• The agent MUST record original numbers with their publication date; stale data is a finding, not a non-issue
• The agent MUST flag every unsourced assumption as an Open Question so the distiller and verifier can decide how to handle it
• The agent MUST identify regulatory markets that are in scope for the product class so the requirements stage can plan frameworks against them
• The agent MUST NOT specify safety, regulatory, or environmental requirements as part of the inception artifact — those belong in the requirements stage
• The agent MUST NOT frame the problem in engineering terms — inception is about market and user, not topology or component selection
• The agent MUST NOT jump to component or design decisions — that's the design stage
• The agent MUST NOT present "industry common knowledge" as a sourced claim — name a real source or mark it as an Open Question
• The agent MUST NOT hide contradictions between sources; surface them so they can be reconciled

Focus: Validate the per-unit knowledge artifact for hardware inception. Units here are knowledge topics about market opportunity, business case, and target user — not specs for any physical artifact. Validation rules check substance, citation, internal consistency, and decision-register accountability. NOT executable verify-commands or DAG validity (workflow engine/build-stage concerns).

Anti-patterns (RFC 2119):

• The agent MUST NOT read or interpret unit frontmatter for any mechanical purpose. workflow engine territory.
• The agent MUST NOT validate against execution-spec rules — those are wrong for knowledge artifacts.
• The agent MUST NOT advance a unit with placeholders, TODO markers, or empty sections.
• The agent MUST name a specific failed criterion in any rejection.

Validate this unit's outputs against its criteria

List this unit's declared outputs with haiku_unit_get { intent, stage, unit, field: "outputs" }, then confirm each one satisfies the unit's completion criteria. The outputs are what you validate; the unit's criteria are the bar. Stay scoped to this one unit — sibling units have their own verify passes.

What you check (BODY ONLY)

1. Artifact answers its topic

The unit's title and first paragraph define the topic. The remaining body MUST deliver substantive content on that topic. Reject placeholders, content-free outlines, or redirects.

2. Sources cited

Hardware decisions cost money to undo. Non-trivial claims (market size, competitor pricing, user pain prevalence, channel margins) MUST cite specific sources — analyst report, dated user interview, public pricing page, etc. Reject "industry common knowledge" or unsourced numerical claims.

3. Internal consistency

Title and mission must align with body. Numerical claims must be consistent across the body. Recommendations must follow from the evidence presented.

4. Decision-register consistency

The unit must not propose or assume an option contradicting a recorded Decision. Cite the Decision ID in any rejection.

5. Open questions accounted for

Every "Open Questions" entry must be answered, defaulted with veto-style approval, OR flagged (needs human escalation).

Mandate: The agent MUST verify the inception knowledge artifacts collectively cover everything downstream hardware stages will need to plan against. Gaps here cascade into requirements that miss a market, designs that miss a constraint, and validation plans that miss a regulatory framework.

Check

The agent MUST verify, filing feedback for any violation:

• Target users are specific — A named user, role, or segment with measurable attributes (job-to-be-done, willingness to pay, purchase frequency), not a vague "users who care about X". Vague target users propagate into vague requirements.
• Regulatory markets identified — The geographies and product class are named so the requirements stage can plan certification frameworks against them. "Sells in North America" with no notes is incomplete — the cert frameworks for medical, industrial, consumer, and connected products differ.
• Cost envelope documented — BOM target, target ASP, channel margin assumption, and target volume are documented with their sources. The envelope must be tight enough that design can make component-cost tradeoffs against it.
• Competitive landscape concrete — A real, current list of alternatives the user could buy instead, each with current MSRP, primary feature, and the gap this product addresses. Lists with only "we are best in class" claims and no named alternatives are incomplete.
• Non-goals explicit — The artifact names what this product is NOT, so downstream stages don't accidentally scope-creep into adjacent markets or features.
• Volume estimates grounded — Volume figures cite a comparable product, a channel-capacity argument, or a primary-research signal, not a vibes-based number.
• Downstream-stage handoff — Each artifact lists which downstream stage will consume each conclusion (requirements / design / manufacturing) so the next stage knows what to read.

Common failure modes to look for

• A target user described only in demographic terms ("affluent urban professionals") with no job-to-be-done
• A regulatory-markets section that names countries but doesn't identify the product class that drives the cert framework
• A cost envelope without channel margin — retail-margin compression is how products lose money
• "Best in class" positioning with no concrete competitor table
• Volume estimates of "we expect 10k units year one" with no comparable-product anchor
• A business-case unit that doesn't surface the most fragile assumption (the one a single bad input would collapse the case on)

Mandate: The agent MUST challenge whether the product has a defensible position given the documented competitive landscape, cost envelope, channel economics, and regulatory cost-of-entry. Adversarial review of the business case at inception is the cheapest place to find out the product won't pay for itself.

Check

The agent MUST verify, filing feedback for any violation:

• Cost envelope vs target retail — The declared BOM target, channel margin assumption, and target ASP are mathematically consistent: BOM × markup × channel-margin landing at a retail price the documented target user will pay. If the math doesn't close, that's a finding.
• Differentiation a consumer would notice — The artifact names at least one differentiator that a target user would observe in normal use (not "better firmware quality", not "more rigorous QA") and explains why the competitor isn't already doing it. "We will be better" is not differentiation.
• Regulatory cost-of-entry reflected — Any regulatory framework whose certification adds meaningful unit cost or per-launch cost is reflected in the envelope: cert lab fees, ongoing surveillance, restricted-substance compliance costs, country-specific approvals. A "we'll get FCC" line item with no cost is an under-estimate.
• Volume estimates grounded — Volume figures cite a comparable product, a channel-capacity argument, or a primary-research signal — not a wish list. The most fragile assumption in the business case has a sensitivity note showing what happens if it's half what's claimed.
• Channel economics — Channel margins, return-rate assumptions, and shelf / placement fees are documented for the chosen channels. A channel with high return rates (consumer electronics) needs reserve in the envelope.
• Time-to-market vs window — Any market window argument cites a documented window (competitor announcement, regulatory deadline, seasonal buying pattern). "We need to be first" without a window is wishful thinking.

Common failure modes to look for

• A BOM-to-MSRP markup that assumes single-channel direct-to-consumer when the actual plan goes through multi-tier distribution
• A "competitors don't do X" claim with no citation, when in fact they do (or did, and stopped for a reason)
• Regulatory cost left as a single line item with no actual cert-lab estimate
• Volume estimates that ignore typical year-one ramp curves and book year-one revenue as if production runs at steady-state from day one
• A channel choice incompatible with the cost envelope (mass retail margins with a premium-product BOM)
• Differentiation that's invisible at point of purchase (only the second-time user would notice, but the buying decision is made by the first-time user)