Three sheets, never merged.

PHOR keeps what a document says, what a model read into it and what you decided on separate layers, so a startup trend can be traced back to its source plate.

Open PHOR Turn to the evidence plate

Drag sideways or use the arrow keys to fan the plates. Press 1, 2 or 3 to lift a sheet. Pressed together, they still keep a gap.

The model reads every plate. It never redraws the one underneath.

How to read these plates

Source: The document as retrieved, with its page and hash.
Machine reading: What a model read into it. Kept on its own sheet, pointing at the passage.
Human decision: The red sheet is yours.

A fluent answer paints over the source.

When a summary replaces the document, nothing underneath can be checked. Can you lift the sheet?

Search result≠Evidence
Model answer≠Source
Company claim≠Independent validation
Mention≠Affiliation
Repository≠Company

Same project, two sources

One project, two numbers: the OpenMAIC v1.0.0 release says 20 built-in skills. The current README says 24.

20built-in skillsv1.0.0 release notes2026-08-27

24built-in skillsCurrent README2026-09-26

Both are right. The release note is frozen at v1.0.0; the README describes the repository today.

Summary, source dropped24 built-in skills

A summary that just says “24” has lost the version. PHOR keeps each claim tied to its source and date.

Seven plates, in registration.

Each stage is a sheet laid over the last, and every one can be lifted off.

1
Signal
A weak lead, such as a rising repository. Not evidence yet.
2
Question
Written down: what would have to be true?
3
Source
Retrieved and hashed, with its URL, page and role.
4
Claim
One narrow statement, tied to the exact passage.
5
Evidence
A claim the source supports, with a stable ID.
6
Thesis
Propositions you can test, each with a status.
7
Monitor
Searches aimed at gaps and catalysts.

Decision

Made by a person, outside the pipeline.

GitHub is the discovery feed today. Other signal sources are intended, not built.

The page, and its limit.

A record keeps its source sheet and says exactly where the reading stops.

Example case: Instava

E009Cadvex annual report

“Investments — Instava Inc USA”

Source: Annual Report 2018–19
Page: Physical page 98
Role: Historical financial report
URL: https://www.cadvex.example/investors/annual-report-2018-19.pdf illustrative
Document date: 2019-08-30 illustrative
Retrieved: 2026-09-26
SHA-256: 00697bc7e5793d06f41dd6c4848a6a62f89fa647ae7a634a03ae35c7dbcc4972

Establishes

historical reported investment entry

Does not establish

current ownership or current investment access

The entry is real history. It says nothing about who owns what today.

The URL and document date on this card are illustrative.

Evidence packet P01Analysis runs against a frozen evidence packet with its own hash. New evidence never rewrites an old conclusion.952344713d54bf276b0f3012c5edd1586b7cfa937993047eb29073f9f5c24a67

Machine readingPHOR is moving toward typed automated review for obvious matches. Overstatement, ambiguity and layout risk go to a person.

Sheets accumulate. Chats don’t.

A conversation scrolls away. A case keeps every layer: searched, read, accepted, rejected, unknown.

Chat transcript

Prompt
Answer
Forgotten

The reasoning lives in prose you can't cite.

Research state

E001Indian legal entity: incorporation and registration details
- SUPPORTST01Instava is an identifiable company with an Indian legal entitySUPPORTED
- G01Find customer-owned evidence
- C01Independent customer documents production AI-agent deploymentWATCHING
E002Company-authored AI-agent positioning
- QUALIFIEST02Customers deploy Instava AI-agent functionalityUNTESTED

Evidence, theses, gaps and catalysts, each with an ID you can follow.

E001 supports T01. G01 names what T02 still lacks. C01 watches for it.

Propositions on their own sheet.

A person sets each status. A model’s output is a separate sheet and never changes it.

T02PropositionUNTESTED

Customers deploy Instava AI-agent functionality

Support: No independent deployment evidence yet
Gap: G01 Find customer-owned evidence
Evidence: QUALIFIESE002Company-authored AI-agent positioning
QUALIFIESE003Company-published customer and testimonial material
QUALIFIESE009Cadvex annual report, Instava Inc USA entry
SUPPORTSnone yet

Model output

INVESTWATCHPASS

Evidence packet P01

Human decision

PHOR can draft INVEST, WATCH or PASS. That is model output, stored with its evidence packet. Your decision is a separate record.

Watching the margins.

PHOR looks for new sources that bear on a proposition, a gap or a catalyst, and sets them aside to be inspected.

A catalyst is a note in the margin, not a prediction.

C01CatalystWATCHING

Watching for

Independent customer documents production AI-agent deployment

Direction: STRENGTHENS
Proposition: T02 Customers deploy Instava AI-agent functionality

Each result is sorted: new source, known source, duplicate lead or irrelevant.

NEW_SOURCE
KNOWN_SOURCE
DUPLICATE_LEAD
IRRELEVANT

Monitoring finds possible new evidence. It never rewrites a thesis or marks a catalyst observed.

How the plates stack.

Lead → inspectA search hit points at a source. Open the source.
Claim → verifyCheck that the source says this, and no more.
Evidence → citeAnalysis cites evidence IDs, not untraceable prose.
Thesis → testA proposition is only as strong as its evidence.
Catalyst → watchName what would change the case, then watch.
Decision → humanA model may recommend. A person decides.

Send us the question.

PHOR is a small beta for people who do primary research, alone or in a small team. Leave your email and what you research. We'll write back.