LeanOS-Technologies Insight Extraction
Process governor-provided expert sources into structured, tier-labeled insights. Every extracted claim carries a tier label and maps to a specific hypothesis. Behavioral vs hypothetical evidence distinction maintained throughout.
Procedure
Step 1: Ingest Source [S]
Read: source URL, file, or transcript (governor-provided).
- If URL: WebFetch to retrieve content.
- If file: Read file content.
- If transcript: Read transcript.
Identify structure (chapters, sections, speakers). Capture metadata: date, author, platform, publication context.
Produce: raw content + metadata.
Gate: source_ingested: bool -- content retrieved, metadata captured (date, author).
- Pass: Step 2.
- Fail: If URL returns error, report to governor. If content is gated/paywalled, report: "Cannot access -- governor must provide content directly."
Step 2: Extract Claims and Frameworks [K-grounded]
Grounded in: raw content from Step 1.
Identify and extract:
| Category | What to Look For | Example |
|---|---|---|
| Frameworks | Mental models, decision structures | "The 4 properties of a good problem" |
| Principles | Universal rules, guidelines | "Never price below 10x the cost of the alternative" |
| Tactics | Specific actions, playbooks | "Use 5-second tests to validate landing page messaging" |
| Data points | Benchmarks, metrics, statistics | "Average PLG conversion rate is 3-5%" |
| Warnings | Anti-patterns, failure modes | "Teams that skip problem validation fail 3x more often" |
For each: state the claim, cite the specific quote or passage (with location/timestamp if applicable), note the context.
Produce: extracted claims list.
Gate: claims_extracted: bool -- at least 3 claims extracted, each with specific citation from source.
- Pass: Step 3.
- Fail: If source is too thin for 3 claims, extract what exists and note: "Source produced limited actionable claims."
Step 3: Tier-Label Each Claim [R]
For each claim, apply the operational test: "What new data would I need to see to change my mind about this?"
| Answer | Tier | Example |
|---|---|---|
| "None -- this is definitional or from cited data" | T1 | "SaaS gross margins are typically 70-85%" (from benchmark report) |
| "Validate the reasoning chain" | T2 | "Teams with this problem spend 40% of time on workarounds" (expert estimate) |
| "Customer/market data I don't have" | T3 | "Customers will pay 10x for this solution" (expert opinion about WTP) |
Also distinguish:
- Behavioral evidence: "We observed X happen" -- stronger
- Hypothetical evidence: "X should work because Y" -- weaker
Produce: tier-labeled claims.
Gate: claims_labeled: bool -- every claim has tier label and behavioral/hypothetical classification.
- Pass: Step 4.
- Fail: Claims without clear tier are likely T2 or T3. Default to T2 for reasoning, T3 for predictions.
Step 4: Map to Hypotheses [R]
For each claim, determine which hypothesis it supports or challenges:
| Mapping Target | Evidence About |
|---|---|
| Problem hypothesis | Pain existence, frequency, severity, alternatives |
| Segment hypothesis | Who has the problem, observable characteristics |
| Unit Economics hypothesis | Pricing, costs, conversion rates, benchmarks |
| Solution Design | Features, growth architecture, MVP approach, positioning, differentiation, jobs-to-be-done |
| General | Domain knowledge, not hypothesis-specific |
Explicitly search for claims that contradict current register. Challenging evidence is more valuable than confirming evidence.
Produce: hypothesis-mapped claims.
Gate: claims_mapped: bool -- every claim mapped to at least one hypothesis or marked "General."
- Pass: Step 5.
- Fail: If all claims map to "General," the source may not be relevant to current strategy work. Report to governor.
Step 5: Write Outputs [S]
For hypothesis-specific insights: format as EvidenceItems:
- [{TYPE: WEB_RESEARCH or OBSERVATION}] [{T1|T2|T3}] {date} -- {source}: {detail}
- Note which hypothesis they should be added to
For general insights: write to knowledge/ as standalone file if source is broadly relevant.
The agent integrates evidence items into the register on the next BUILD/CHALLENGE pass.
Produce: evidence items for register integration + optional knowledge file.
Gate: outputs_written: bool -- evidence items formatted for register integration, each with hypothesis target.
- Pass: Done.
- Fail: Ensure every output has the required format fields: type, tier, date, source, detail.
Quality Criteria
- Every extracted claim cites specific passage from source (not paraphrased without reference)
- Tier labels applied to every claim with justification
- Behavioral vs hypothetical distinction maintained (quotes about what happened vs what should happen)
- Claims mapped to specific hypotheses (not all dumped as "general")
- Source metadata complete (date, author, URL if applicable)
Failure Modes
| Mode | Signal | Recovery |
|---|---|---|
| Uncritical extraction | All claims extracted as T1 or all labeled as important | Apply tier test rigorously. Expert opinion about what customers want is still T3. Expert frameworks are T2. Only cited data is T1 |
| Confirmation bias in extraction | Only insights supporting existing hypotheses extracted; challenging insights omitted | Explicitly search for claims that contradict current register. Challenging evidence is more valuable than confirming evidence |
| Quote without context | Extracted quote is stripped of qualifying language that changes its meaning | Include surrounding context. "This works" may have been preceded by "In enterprise markets above $50K ACV" -- the qualifier matters |
Boundaries
In scope: Source ingestion (URL, file, transcript), claim extraction (frameworks, principles, tactics, data points, warnings), tier labeling, behavioral/hypothetical classification, hypothesis mapping, evidence item formatting.
Out of scope: Autonomous web research (strategist does this directly), competitive analysis (stg-analyzing-competition), market sizing (stg-sizing-markets), hypothesis construction (skills handle their own domains).