Documents written about the laboratory by parties outside a working session: strategy memos, portfolio recommendations, external evaluations, and the design for an experiment that has not been run. Not part of the numbered reading course in docs/. They are kept because decisions in ROADMAP.md cite them, and a cited input that lives on someone's desktop is not a record.
The contract, stated the way hunts/README.md and interactive_lab/README.md state theirs:
- Nothing in this directory is a result. These are opinions and recommendations, not measurements. A memo recommending a programme is not evidence the programme works; only the tree's own instruments produce that. No claim here has been through a battery, and the E1–E3 documents describe an experiment whose outcome does not exist yet.
- The tree decides, the memo proposes. What was adopted, deferred, or rejected from each document — and why — is recorded in
ROADMAP.md, not here. A memo is never edited to match the decision. - Transcriptions are labelled. Where a document arrived as a PDF, the markdown here is a transcription: mathematical typography restored where extraction garbled it, content otherwise unaltered.
- This repository is public. Anything landing here is checked for names, private repository references, and absolute paths before it is committed.
The blind pair
analysis.md and consultant-thesis-analysis.md are two independent evaluations of the same outside "productization thesis" against the same commit (5533896), written without either seeing the other — the second states explicitly that it did not open the first. They are kept as a matched pair because that independence is the only thing that makes either of them evidence of anything, and it is verifiable from their own chronology rather than from a declaration. Same reasoning as the frozen corpus in harness/blind_authoring_2026_08_09/.
Redacted before landing: four private sibling repositories and one adjacent private repository are referred to by role rather than by name or path. The arguments that relied on them are unchanged and were already flagged in the text as unverifiable from this tree alone.
The E1–E3 protocol
Five documents, in dependency order, asking one question: is a meaningful part of this lab's operating discipline carried by transferable artifacts and machinery, or mainly by the original operator's unstated judgment?
| File | What it is |
|---|---|
e1-e3-experiment-protocol.md | v0 of the protocol — E1 intervention ledger, E2 Lab Rotation Benchmark, E3 staleness/flattening |
e1-e3-protocol-revision-2.md | supersedes the affected sections of v0, answering five objections |
e1-e3-protocol-rev-2.1-patch.md | four specification holes; patch only, Rev. 2 otherwise unchanged |
e1-e3-implementation-logistics.md | sequencing, plus the blind lesion-authoring procedure the contract references but does not define |
e2-consultant-blind-authoring-packet.md | the only material an external lesion author receives before reveal |
The contract is Rev. 2 + the Rev. 2.1 patch; read v0 only for the sections neither of those touches.
A caution on the packet, and it is the reason the packet is the one file here that carries a cost. It names the lesion classes assigned to the external author. Anyone who reads it before the runs are scored is contaminated as a scorer, and the protocol's own exclusion rule then applies to them. If E1–E3 is still unrun, treat this directory as read-only for anyone who might score it.
Freezing, if the experiment goes ahead: harness/preregistration.py records digests of the evidence that already existed at freeze time, so contamination and criteria drift are derived from artifacts rather than read off a declaration. That is the mechanism these documents should be frozen with, and the commit that lands them is the timestamp.