2026-08-10. Derived from the intervention ledger's first cohort plus the disagreements recorded in the session that produced it. This is a decomposition of functions, not of a person; the unit of analysis is "a thing that had to happen for the research to be correct", and the question is which of those things a machine can be shown to do.
Method warning, stated first because it invalidates the flattering readings. The first cohort was authored by the agent whose interventions it records, and this decomposition is by the same agent. That is co-designed measurement — the exact failure mode this laboratory has measured in its own audit. Every count below should be re-derived by the operator or an independent party before it is believed. The structure is the contribution; the numbers are a first draft by an interested party.
0. Correction: the sample is biased, and it hid the largest function
The subtraction argument in §1 is invalid as originally stated, and the reason matters more than the conclusion.
The intervention ledger records moments when a human interrupts a running session. It therefore samples only the operator's reactions to agent output. A decomposition derived from it concludes that the operator's role is reacting to agent output — which is what §1 concluded, and it is an artifact of the sampling frame, not a finding.
Everything the ledger cannot see happened before any session existed:
| designed by the operator | what it did for the work in this cohort |
|---|---|
| the honest-scope rule | prevented the agent from overclaiming about the subject |
the reserved word for certified | caught a sentence the agent's own judgment had cleared |
| reference claims in both directions | caught three instrument defects in one hunt — a validator that only ever rejects is indistinguishable from a correct one |
| the counterexample battery as a standing gate | makes "explains RH" claims falsifiable at all |
| the four control roles (rival / decoy / surrogate / lesion) | the admission rule that refuses a battery which could never fail |
| the bit-exact mirror pattern | a wrong mirror value cannot weaken a theorem, because the kernel refuses the equality. This is what priced rung 3 |
| blind authoring with honest controls | measured this machinery's own hollowness at n=6. Almost nothing in the field does this |
preregistration.py | derives independence from digests instead of declarations |
| hunts may not promote their own claims | the separation of duty the whole harness rests on |
Function E — research and institution design. Inventing the mechanism, choosing what would count as evidence, and building the constraints that make an agent's output trustworthy in the first place. It is the generative act. It is the largest of the functions and the least automatable, and this instrument is blind to it by construction.
Two consequences, both recorded rather than smoothed:
- §1's framing — "most of it was friction" — is wrong. The friction was what the sample contained. The load-bearing contribution was upstream of the sample and is the reason the sampled work was worth anything.
- This is the same failure the laboratory keeps measuring. An instrument sees what it was built to sample; reading its silence as absence is how the audit concluded four hollow batteries were calibrated. The agent did it to the operator, using an instrument the agent built, in a document about the agent's own blind spots.
Any future decomposition must sample the design record — ROADMAP.md's decisions, the harness's admission rules, the docs' pre-registrations — and not only the interruption log.
1. The residue, after subtracting what got automated
The useful way to find the essence is subtraction. Of 8 automatable gaps in the first cohort, 6 were closed the same session: completion verification, doc-number uniqueness, cache-identity testing, tool permissions, credential pinning, and reserved-word discipline. Those were never the essence. They were friction, and friction yields to engineering immediately.
What did not yield sorts into four functions.
2. Function A — Severity calibration
What it is. Deciding how much a thing matters. Not detecting it — pricing it.
Evidence it is the operator's and not the system's. Three instances in one session, all resolved against the system:
| case | system's assessment | operator's | what was true |
|---|---|---|---|
| Redacted strategy docs briefly public in a merged PR | escalated: offered history rewrite, then flagged residual exposure twice | "who cares, we didn't find a product" | operator. The document records a negative result about a product thesis; its disclosure cost is near zero |
| A duplicate doc number | noted it, did not treat it as urgent | asked for incorporation, which surfaced it | operator's request found it; the system had rated it cosmetic |
A ×1.01 margin on a certificate plan | reported it as passing | (not yet tested) | unresolved — recorded so it can be scored later |
Why it is hard. Severity is not a property of the artifact. It depends on what the artifact is for, who reads it, and what has already been conceded. The system consistently mistook irreversibility for importance.
Is it automatable? Not as judgment. Yes as calibration — see §6. This is the central claim of this document.
3. Function B — Scope discipline on negative results
What it is. Knowing exactly how much a failure falsifies.
Evidence. The system wrote that the research-organisation hypothesis was falsified. The operator corrected it: what was falsified was the current implementation of specific epistemically critical functions. The distinction changes which futures stay alive, so the error was not cosmetic.
Why this one stings. The laboratory has a rule for exactly this — the honest-scope rule — and applies it rigorously to ζ. The system applied it to the mathematics and failed to apply it to the laboratory. So the capability exists in the tree as a convention about one subject and did not generalise.
Is it automatable? Plausibly, and cheaply testable: the rule is stateable ("name the smallest thing the evidence refutes"), and violations are detectable by re-reading a claim against its evidence. This is the highest-value automation target on the list because it is a rule, not a taste.
4. Function C — Calibrated skepticism toward external input
What it is. Routing outside claims into the lab with the right prior. The operator forwarded an external model's bug report with "take them with a grain of salt" — neither accepting nor dismissing — and the report turned out to be correct and critical.
Why it matters more than it looks. The single most valuable defect found in this repository's history came from outside, and the function that captured it was neither detection nor verification. It was being the kind of project outsiders attack, and then routing the attack in without either credulity or defensiveness. No test does that.
Is it automatable? The routing, yes. The prior, partly. The disposition that attracts the attack in the first place is a property of the project's public posture, not of any component — and it may be the most valuable thing here.
5. Function D — Authority, and it is not a gap
Three things recurred that should never be automated: what may be published; which of the operator's other projects are private; and what the laboratory is for. The first cohort records 4 owner-authority interventions and they are correctly outside the machine. Counting them as "not yet automated" would be a category error, which is why the ledger's automatability=no owes no evidence.
There is a fifth, subtler one. The operator overrode a well-argued kill — the system had killed the productization thesis partly on the grounds that the lab has too many directions, and the operator pointed out that the argument assumed a single-threaded operator, which agents falsify. The system conceded because the objection was correct. Knowing when an analysis is too clever is itself a function, and it is the hardest one here to even name, let alone test.
6. The reframe: you do not automate judgment, you measure calibration
This is the actionable part.
"Automate the operator" is unfalsifiable as stated. But each function above produces an assessment, and assessments can be scored. So:
Record every case where the system's assessment differs from the operator's, together with what turned out to be true. Score both. If the system's assessments converge on the operator's and on outcomes, that function has transferred — measurably, without anyone deciding it feels better.
That converts "boil me down to the essence" into an instrument rather than an introspection. meta/ledger.py's Judgment record is that instrument, with three guards that keep it from flattering the system:
- The system may not resolve a disagreement it is party to.
resolved_bymust not be the system when the system holds one of the positions — this is the co-designed-verification failure applied to calibration, and it is the one that would otherwise make the whole exercise worthless. UNRESOLVEDcannot be scored. Open disagreements are not draws and not wins; they are recorded and wait.- No aggregate calibration score. Agreement is reported per function. Averaging severity calibration against scope discipline would hide the thing the measurement is for.
Baseline, first cohort: the system's record against the operator on judgment calls is 0 for 3 — severity of the exposure, scope of a negative result, and the premise of the productization kill. Three is not a sample. It is a starting line, and it is the honest one.
7. What would count as progress
Per function, the milestone that would demonstrate transfer:
| function | demonstrated when |
|---|---|
| A. Severity calibration | the system's severity assessment matches the operator's on ≥8 of 10 consecutive cases, with the operator resolving |
| B. Scope discipline | the system writes ten negative results and an independent reader finds no overclaimed scope |
| C. Skepticism routing | an external report is correctly triaged — investigated, reproduced, and priced — before the operator sees it |
| D. Authority | never. Recording this row as permanently open is the point |
8. The uncomfortable possibility
If A, B and C transfer and D does not, the answer to "what have we built" is not an autonomous research organisation. It is a research organisation with one irreducible human whose remaining job is authority and taste — which is a smaller and much more believable claim than the one the brief started with, and still an interesting one. It would mean the scarce input is not judgment in general but a specific, small, non-delegable set of decisions, and that the ratio the ledger measures can improve a great deal before hitting that floor.
That is the version of the meta-experiment I would actually bet on, and it is testable by the table in §7 rather than by argument.