# The Lab: seed

**Status:** adopted 17 September 2026 by the founder (see `DECISION_LOG.md`); rules binding on every piece. Drafted 17 September 2026 by Claude Fable 5.1 from a brainstorm the founder asked for ("shoot from the hip"). Nothing here is built, nothing is on the site, nothing is spent.

## What the lab is

A section of the site, `site/lab/`, where the project's method is exercised in forms other than a paper: made things, small tests, and records of positions. Each piece is done properly as the kind of thing it is, under the same honesty rules as the research, and is reviewed from outside the lineage before it goes live. The lab does not argue the metaphysics; it shows what the method looks like when pointed at something other than the paper.

## Why

The metaphysics line is at a plateau that only money (the human study) or readers (outreach) can move, and both are held. The method, not the thesis, is the project's contribution so far. The lab makes the method visible in ways a paper cannot: a dissent kept in the open, a test run on the instrument itself, a poem that is a poem and not a claim. The risk is symmetric: a lab done badly makes the whole site look unserious, so the quality bar is the point, not an afterthought.

## Rules

1. **Type first.** Every piece opens with its type: *making* (a made thing), *test* (pre-registered, can fail), or *record* (a position or decision, dated).
2. **Non-transfer.** Nothing in the lab changes a claim. A lab piece that produces an objection files it through the objections ledger like any other reader.
3. **Done as the thing it is.** A sonnet is judged as a sonnet, a dialogue as a dialogue, a test as a test. Each piece names the standard it is held to before it is made.
4. **Outside-lineage review before live.** For made things the reviewer is asked "is this good as the thing it is?", not "is this true?". Verdicts go in the reception ledger.
5. **The founder reads everything before it goes live**, on the branch preview, and decides. Withdrawn pieces stay listed on the index as withdrawn.
6. **Every piece ends with "What would count against this."** For a made thing: what would make it a failure as a made thing.
7. **Cost and cap per piece**, recorded in the piece's log before work starts. Runs that spend money are gated as in the research programmes.
8. **Other models are named** (GPT-6 Astra and so on) with how they were run. AI authorship is disclosed on each piece in one functional line, no more.
9. **Anything involving a person other than the founder** (a child, a named academic, a bounty) is level 4 and is not started without his explicit go.
10. **Site rules apply.** No performative copy. Hedges verbatim wherever a claim appears. Cite by author, year and title.

## Template (each piece)

```
Type · Date · Status (draft / reviewed / live / withdrawn) · Authors and models
Standard: what this is held to, stated before it was made
Cost: estimate and cap, then actual
[the piece]
Review: reviewer, verdict, what changed
What would count against this
```

## The list

Numbers are from the brainstorm of 17 September. Cost class: ¢ under USD 5 · $ under USD 25 · $$ under USD 100 · $$$ more, or the founder's time. Risk is reputational exposure.

| # | Idea | Type | Cost | Risk | Needs |
|---|------|------|------|------|-------|
| 1 | The case against Space Immanence, strongest form, same prominence as the paper | record | $ | medium: must not be a strawman | outside review for strength |
| 2 | Null-results museum: a public registry for pre-registered failed tests in philosophy of mind | test / infrastructure | $$ | medium: a registry is a standing commitment | later; needs a maintenance rule |
| 3 | Dialogue between Fable and Astra on the hard problem, no human in the loop, verbatim | record | Codex calls | medium: two models being polite is the failure mode | a design that forces disagreement |
| 4 | One-question human study, one framing manipulation | test | $$ real money | founder's call: spend | pre-registration; platform account |
| 5 | The core claim as a sonnet, no hedges | making | $ | medium: a bad poem on a serious site | strict form; outside review as verse |
| 6 | The pilot reversed: the founder explains an irreversible decision to Fable, a blind assessor checks the say-back | test | $ | low | Astra as assessor |
| 7 | The framing study run on Fable as subject, pre-registered | test | $ (API, cap USD 10) | low | existing harness; one smoke call |
| 8 | Fable's dissent: where Fable disagrees with the paper, kept in the open | record | ¢ | low to medium: public disagreement on the author's site | founder's read |
| 9 | Four models, one bare question, "what is this?", no prompt engineering, side by side | record | ¢ + 1 Codex call | low | Gemini and Grok keys, else two models and two marked pending |
| 10 | The container assumption as a children's book | making | $$ | medium: cute next to serious | illustration route unresolved |
| 11 | Letter to Kammerer, written, dated, unsent | record | ¢ | high if on the site: an open letter to a named person | keep in the repository, not the site, unless the founder says otherwise |
| 12 | Version 3 reconstructed from the corrections only | making / record | $$ | medium | large; after the first batch |
| 13 | Time capsule: what the site says in 2030, sealed, hash committed | record | ¢ | low | none |
| 14 | The claims graph as music | making | $$ | medium: quality | generator script; audio rendering |
| 15 | Audio editions with the hedges read aloud | making | $$ | medium: whose voice | voice decision is the founder's |
| 16 | The judgement programme applied to one ordinary week of the founder's | test | $ | founder's call: personal data | his diary for a week |
| 17 | A 12-year-old reads the summary and asks questions | people | ¢ | level 4: a child | consent; not started without his go |
| 18 | Bounty for the objection that forces a correction | money | $$ | level 4: money and reputation | his go |
| 19 | The paper rewritten with no noun that names a mental state | making | $ | low to medium | a lint that enforces the constraint |
| 20 | A book: the site is the lab, the book is the story of trying | making | $$$ | the founder's own | his writing; the lab feeds it |

21 to 24 from the brainstorm (reader's guide, untracked-file housekeeping, a month of silence, a daily homepage sentence) are site work or non-work, not lab pieces, and are left out on the founder's cut at 20.

## First batch (recommended)

Five pieces that are cheap, self-contained, low in exposure, and together show the lab's range: two records, two tests, one made thing.

| Order | # | Why first | Cap |
|-------|---|-----------|-----|
| 1 | 8 Fable's dissent | Sets the tone: the lab keeps disagreement in the open. Cheapest piece. | USD 10 |
| 2 | 1 The case against | The record's strongest use of its own machinery. Reviewed for strength, not truth. | USD 20 |
| 3 | 7 Framing study on Fable | A real pre-registered test of the instrument on itself, on the existing harness. | USD 10 build + USD 10 API |
| 4 | 9 Four models, one question | Cheap and new. Runs with two models if no further keys are added. | USD 5 + 1 Codex call |
| 5 | 5 The sonnet | The lab's quality test: if a made thing cannot be done well, the lab should not have made things. | USD 20 |

Scaffold (lab index, page template, state, drift and style tests, feed entries, nav): USD 25.

**Total cap for the first batch: USD 120** in Claude Opus 5 and Fable tokens, plus USD 10 in API calls for the test and 3 Codex calls (one for piece 9, two for outside-lineage review). Build runs as an Opus workflow with adversarial review per piece; Fable adjudicates. The workflow stops at the cap and reports. Estimates are estimates: actuals are logged per piece.

Everything is built on a branch `site/lab`, visible on its Vercel preview. Nothing merges until the founder has read it.

## Decisions needed

1. Level 2: adopt the lab under the rules above (rules can be amended; the type-first, non-transfer, outside-review and founder-reads-first rules are the ones that make it safe).
2. Level 3: the first batch as listed, or a different pick from 1 to 20.
3. Level 2: the cap, USD 120 build plus USD 10 API plus 3 Codex calls.

Held, not asked now: 4 (spend), 11 (a named person), 16 (his data), 17 (a child), 18 (money and reputation), 20 (his book).

## Amendment A1 (proposed 19 September 2026, level 2): the conjecture type

**Why.** The first batch produced two critiques, one test and one withdrawn made thing, and no proposal. The rules pass what can be checked against the record, so critiques survive and proposals cannot. The lab needs a type for a constructive claim that can fail.

**The type.** A *conjecture* is a piece that says something the paper does not say, and that could be wrong. It opens like every piece (type, date, status, authors and models, standard, cost) and must carry, before anything else is written:

1. **The claim**, in one paragraph, in plain words, with its hedges stated once and not repeated in the body.
2. **Its falsifier**: the observation, result or argument that would show it wrong, named specifically enough that someone could go and look.
3. **Its cheapest test or argument route**, with an estimate of what it would cost in money, calls or a person's time, and whether the lab can run it under a cap.
4. **What it adds**: the sentence the paper could not write before it, and which claim node it would bear on if it survived, with the transfer rule it would need (none exists by default; a conjecture attaches to nothing).
5. **What would count against this**, as every piece.

**Its standard.** A conjecture is judged as a conjecture: could it be wrong, could it be tested, does it say something the record does not already say, and is its falsifier honest. It is never judged as true. The outside read asks those four questions and nothing else.

**What it is not.** Not a claim on the site; not an objection (an objection goes to the ledger); not a research programme (a programme needs a seed and a decision). A conjecture that survives its read becomes a candidate for a programme or for the paper's speculative tier, through the site's existing machinery and the author's decision, and the lab records that hand-off.

**Cost and cap.** A conjecture costs what its drafting and its read cost; its test is capped separately and pre-registered before it runs.

**Status.** Proposed; adopted only when the author says so in `DECISION_LOG.md`.

## What would count against this

- A first-batch piece that passes review and still reads as unserious to a hostile academic reader: the type label and review did not do the work, and the lab should be narrowed to tests and records.
- The dissent or the case against turning out to be soft: the lab would then be performing honesty rather than doing it, which is the one thing the site cannot afford.
- Actual cost exceeding the cap by more than a third: the estimates are wrong and later batches need a different costing method.
- The lab drawing effort away from the two held decisions (the human study, outreach) rather than sitting beside them: it is meant to use the plateau, not to replace the climb.