How PDE Turns Real Evidence Into Structured Problems
PDE starts with real human evidence — not ideas. We collect, structure, group and validate recurring problem signals before any commercial opportunity is considered.
Core Pipeline
Every conclusion in PDE follows the same chain. No step may be skipped. No step may be replaced by AI speculation.
Forbidden shortcut: Evidence → AI → Startup Idea. This path is explicitly banned by the PDE Constitution Addendum (§1): "AI is prohibited from creating opportunities out of thin air."
What Counts as Evidence?
Evidence is what a real person published about their own experience. It is never an AI summary, never a search snippet, never a company landing page.
Allowed Source Types
- Public forum discussions
- Public support threads
- GitHub issues
- Reddit discussions
- Public Q&A
- Community posts
- Public complaints with context
- Public reviews with context
Refused as Sole Evidence
- Company landing pages
- AI-generated articles
- SEO articles
- Search snippets
- Generic blog summaries
- AI-generated text
Evidence ≠ AI summary. Every piece of evidence must trace back to a real source URL, a real source date, and a bounded raw quote from the person who experienced it. (PDE Constitution Article 3, 5)
Evidence Hierarchy
PDE does not currently publish formal evidence tiers (A–E or similar). All evidence is evaluated by its source type, provenance, and the structured fields a human reader can extract from it. What PDE does enforce is a commercial signal priority:
No model may substitute a weaker signal for a stronger one. (Addendum §5)
Traceability Matters
Every conclusion in PDE must be traceable back to its source. If a signal cannot be resolved to the raw evidence it came from, it is an assertion — and PDE does not make assertions. (Article 4)
Derived interpretation cannot replace source evidence. A signal carries the extractor's identity and version (Article 6). Re-interpreting evidence with a new extractor must not touch the original evidence or overwrite the earlier interpretation.
What Is a Problem Signal?
A signal is a structured expression of pain, failure, or unmet need — extracted from real evidence by a named method with a version number.
A Signal IS
- A structured interpretation of raw evidence
- Traceable to a specific evidence record
- Carrying extractor identity + version
- Expressing actor, context, pain, or failure
- Able to carry confidence (of the extraction, not the opportunity)
A Signal IS NOT
- A startup idea
- A keyword or search term
- AI guesswork
- A generic emotional expression
- A score, rank, or rating
Signal Schema
Each signal carries these structured fields when the evidence supports them:
actorWho experienced thisroleTheir role or contextcontextWhat was happeningjob_to_be_doneWhat they were trying to accomplishpain_or_failureWhat went wrongcurrent_workaroundHow they copedmoney_signalReal money at stakeswitching_signalLeaving existing solutionsfrequency_hintHow often it recursconfidenceExtractor's confidence (not opportunity score)A null field is an honest "the evidence does not state this" — not a zero, not a guess. (Article 12)
When Is a Signal Valid?
A signal's status is one of three values: SIGNAL_VALID, SIGNAL_INVALID, or INSUFFICIENT_EVIDENCE. There is no score.
| Criterion | Requirement |
|---|---|
| Source authenticity | Source URL must be a real, accessible page of an allowed type |
| Actor clarity | A real person must be identifiable in the evidence |
| Context clarity | The situation must be specific enough to evaluate |
| Pain / failure | A concrete friction, failure, or unmet need must be present |
| Traceability | The signal must link to its raw evidence record (Article 4) |
| Content hash | The stored document must hash to what was collected |
| Human verification | A human must confirm the extraction matches the source |
| Independence | Independence is judged, not assumed (Article 10) |
No universal score exists on a signal. Confidence is a property of the extraction, not of the commercial opportunity. A signal that scores "high" is not authorised to skip any downstream gate. (Addendum §2: HIGH SCORE ≠ BUILD)
Raw Evidence vs Derived Fields
This distinction is central to PDE's credibility.
Raw Evidence
- What someone actually published
- Never AI-rewritten, never summarised, never translated
- Protected by database constraints, not good intentions
- If the source changes, a new recapture is recorded — the original row stays
Derived Fields
- Structured interpretations extracted from raw evidence
- Can be recomputed, replaced, or versioned
- Carry extractor identity and version
- Must never overwrite or touch the raw evidence
AI may assist extraction, but cannot override raw evidence. Derived text lives in derived tables. Raw text is what was actually published. (Article 5, 6)
How Signals Become Clusters
A single signal is not a market problem. Clustering finds patterns that repeat across sources, users, or contexts.
A cluster groups related signals that share a common theme — a recurring job, failure mode, or unmet need. PDE's clustering is a human judgement declared in reviewable code, not an algorithm. The grouping carries its evidence basis and a disposition (selected, considered, or rejected), each with stated reasons.
No fixed thresholds are published. PDE does not claim "3 signals = cluster" or "5 users = valid." A cluster's validity depends on whether its members independently share a job + failure + outcome — not on a count. (Article 7, 9)
A single-narrator grouping with no shared failure is not a cluster — it is a watch candidate. The identity rule is strict: members must have something to share.
The Atomic Problem
A PDE Problem is not a keyword and not a product. It is a structured identity:
A valid Problem is not "Zapier is too expensive." It must be structured to: who, in what context, wants to accomplish what, is blocked by what. The same product is not the same problem. The same keyword is not the same problem. (Article 7)
Problem strength is computed from the evidence relations that support it — never accumulated by adding points. It can always be recomputed from evidence and always explained. (Article 8)
When Is a Problem Supported?
PDE considers a problem supported when multiple independent signals converge on the same actor + context + blocking step, each traceable to real evidence.
| Dimension | What PDE Checks |
|---|---|
| Actor clarity | Can a real person be identified? |
| Context clarity | Is the situation specific and verifiable? |
| Blocking step | Is the obstacle concrete, not vague? |
| Repeated evidence | Do multiple independent sources describe the same friction? |
| Source traceability | Can every signal be resolved to its raw evidence? |
| Contradiction handling | Are conflicting evidence sources preserved, not silently dropped? |
Problem existence ≠ Commercial opportunity. A problem can be real and well-documented without being commercially attractive. PDE records the problem regardless. Commercial significance is a separate gate.
Commercial Proof
Pain alone is not commercial proof. A repeated problem can be real without being commercially attractive. PDE separates commercial signal types and evaluates each independently.
Payment Signal
Real money already moving through the problem: plan prices, prepaid balances, actual upgrades, disputed charges.
Strongest kindSwitching Signal
Evidence that people are actively leaving existing solutions for alternatives.
StrongWorkaround Signal
Documented attempts to cope with the problem when nothing works well enough.
ModerateDemand Signal
Explicit requests, inquiries, or active search for a better alternative.
SupportingPayment > Usage > Signup > Click > AI Judgment. No weaker signal may be substituted for a stronger one. Commercial proof establishes that real money moves through the problem today. It does NOT establish that any new product earns revenue. (Addendum §5)
When May an Opportunity Appear?
Opportunity is always the last step in the chain. It may only appear after every preceding gate has been evaluated.
Forbidden
Evidence → AI → Startup Idea Required
Evidence → Signal → Cluster → Problem → Commercial Proof → Opportunity Before an opportunity may proceed, five mandatory gates must be evaluated:
| Gate | Question |
|---|---|
| PAIN_REAL | Is the pain documented in first-person evidence? |
| MONEY_REAL | Is real money at stake (not AI speculation)? |
| FREQUENCY | Does the problem recur (daily, weekly, monthly)? |
| SMALL_MVP | Can a ≤7-day test artifact be described? |
| REACHABLE_USERS | Is there a known channel to the first 20 operators? |
If evidence is insufficient, the opportunity must not be generated. BUILD remains forbidden by name regardless of how many gates are satisfied. (Addendum §2, §3)
The Role of AI
AI is hired for measured roles. Its role is bounded, explicit, and never authoritative over evidence.
AI CAN
- Extract structured fields from bounded evidence
- Classify evidence by source type
- Group similar signals as candidates
- Summarize bounded evidence sets
- Identify candidate relations between signals
- Help recompute derived fields
- Assist with MVP construction after validation
AI CANNOT
- Invent source evidence
- Fabricate pain or demand
- Create fake commercial proof
- Turn weak evidence into certainty
- Generate unsupported opportunities
- Override raw evidence with derived text
- Decide BUILD based on scores alone
- Autonomously discover or mass-crawl
Market decides. AI assists. A model's rejection is a durable finding stored in the database so a future session cannot re-adopt it. (Article 15)
What Happens When Evidence Is Insufficient?
PDE allows and expects these outputs:
The evidence does not state this. Not a zero, not a guess.
Some evidence exists, but not enough to conclude.
The gate has not been satisfied. It may be in the future.
The signal cannot be validated from available material.
Unknown is a valid research result. A null field is an honest "the evidence does not state this" — not a zero, not a guess. UNCERTAIN is a visible state, not a deletion. (Article 11, 12)
Contradiction Handling
When different sources conflict, PDE does not silently keep only the evidence that supports a preferred conclusion.
Independence is a three-valued judgement: INDEPENDENT / DEPENDENT / UNCERTAIN, with recorded reasons. Only INDEPENDENT counts toward a repeat. (Article 10)
METHODOLOGY_GAP: PDE's current Constitution does not formalize a contradiction resolution procedure beyond Article 10 (independence judgement) and Article 11 (nothing dropped silently). Conflicting evidence is preserved with its disagreement recorded. A formal contradiction handling procedure may be added in the future as an explicit, dated decision.
Current Example: From Evidence to Problem
The only problem PDE has selected so far illustrates the full chain:
3 independent sources on Zapier's own support community (2023–2026), each from a different author, each naming a pain and a money signal.
No one-off top-up exists · Opaque consumption (48 billed vs 14 counted) · Fail-closed pause despite paid upgrade
Metering design blocks every path to cost predictability on per-task automation platforms
Actor: Small operator on a metered automation platform
Desired result: Keep automations running predictably within a known budget
Blocking step: No one-off top-up, invisible consumption, fail-closed pause
Kind: PAYMENT — real dollar amounts, actual mid-cycle upgrade, prepaid balances in dispute.
Caveat: This establishes that real money moves through the problem. It does NOT establish that any new product earns revenue.
What PDE Does Not Claim
A Final Research Principle
Do not build a company and then look for customers. Find real evidence of a real problem first, then decide whether it deserves a solution.
— PDE Constitution Addendum, §10