Skip to content

Latest commit

 

History

History
183 lines (139 loc) · 11.6 KB

File metadata and controls

183 lines (139 loc) · 11.6 KB

The cut: Observer Zero → 8,000 words for JASSS

Superseded (2026-09-02). This plan targeted the JASSS submission, which is no longer active; the combined manuscript is in peer review (venue per daff318 policy). Pre-erratum numbers survive below ("40/40" → 42 of 42 confirmatory runs). Kept as a record of how the cut was made.

Brief applied: keep both studies, compress Study 1 heavily. Study 1 is the empirical and methodological foundation, not a co-equal results section. Preserve task solvability, the epistemic ceiling, the model-family dissociation, and the zero-communication baseline. Everything else in Study 1 goes to the supplement or is cited to the Study 1 deposit. Protect the cross-study comparisons and replications.

Two findings from measuring the draft change what the cut has to be. Both are in §1 and §2 below.


1. The budget is 34%, not 42% — because 1,692 words are tables

Measured prose against tables, section by section:

Section Prose Tables Total
Introduction 833 0 833
Related work 1,783 0 1,783
Meridian environment 612 0 612
Experimental validity and confirmatory design 1,342 281 1,623
Results 4,364 1,254 5,618
Discussion 1,924 0 1,924
Limitations 631 0 631
Reproducibility and audit trail 692 157 849
Body 12,181 1,692 13,873

JASSS does not say what its 5,000–8,000 range counts. The near-universal convention is prose excluding tables, figure captions and references — and JASSS separately instructs that tables be supplied as text rather than as graphics, which only makes sense if they are not competing with the word budget. Working to 8,000 words of prose plus ~1,700 in tables, the cut is 34% of prose rather than 42% of everything.

This matters for the Study 1 decision specifically. The evidence gradient (Table 5) and the zero-communication replication (Table 8) are both tables carrying Study 1 rows. Under the proportional reading they were expensive; under this one they are nearly free, and they are exactly the two places where compressed Study 1 does the most work. If a reviewer or the editor pushes back on length, the tables are the last thing to cut, not the first.

Prose budget by section:

New section (unnumbered) From Prose now Target Cut
Introduction §1 833 700 −16%
Related work §2 1,783 850 −52%
The Meridian environment and confirmatory design §3 + §4 1,954 1,500 −23%
Results §5 4,364 2,900 −34%
Discussion §6 1,924 1,000 −48%
Limitations, reproducibility and audit trail §7 + §8 1,323 1,050 −21%
12,181 8,000 −34%

2. One of your four Study 1 results is not in the manuscript

You asked to preserve task solvability — the scripted mock society diagnosing gravity strictly in 10/10 worlds. It is not in the combined draft. Arm B0 appears once, as a row in the Study 1 arms table with a cost of $0, and the scripted mock arm is mentioned twice more in the reproducibility section, both times as a determinism check ("reproduces bit-identically"). The solvability result itself is in the Study 1 manuscript at §4 and in combined-manuscript-outline.md, which flagged it for retention:

"the scripted mock society solving the task 10/10 (Study 1 §4) is still an independent solvability argument that uses no detector at all"

It did not survive into the draft. The consequence is that the paper currently offers two independent answers to a reviewer's first objection — maybe the evidence simply wasn't there — where the outline designed three:

  1. Scripted baseline. Mock society diagnoses gravity strictly in 10/10 worlds, stays quiet in 7/10 controls, never trips the confabulation detectors. Uses no detector at all. ← missing
  2. Detector on own streams, dating (Study 1 Finding 4). Present, in 5.1.
  3. Detector on own streams, detection (Study 2). Present, in 5.1.

Arguments 2 and 3 share an instrument. If a referee doubts the change-point detector, both fall together and the sufficiency claim collapses. Argument 1 is the only one that survives that objection, and it costs about 80 words.

So this cut adds it back. That is the one place where the compressed version should be stronger than the long one.


3. Results, restructured

Your sketch — establish the failure, stress it socially, synthesise — maps onto three movements. I have followed it, with one deliberate cost noted at the end.

Target: 2,900 prose. Study 1 ~590 (20%), Study 2 ~1,860 (64%), synthesis ~450 (16%).

Movement 1 — Study 1: establishing the failure (~590 prose)

Content Source Prose Note
Task solvability: mock society 10/10, no detector new, from Study 1 §4 80 The missing third argument
The ceiling: 0/40 strict, 80 agent-final states, 7.5% bound; instrument-fault worlds diagnosed in up to 70% 5.2 140 The boundary is specifically at world-level revision — keep that sentence, it is the finding
Two routes: haiku generation failure (2/20), sonnet commitment failure (15/20, peak p=0.85) + Figure 3 5.2 160 Keep the seed-1009 Ada near-miss; it is the single most vivid thing in the paper
Not the prompt prior: B3b ablation, and it had stronger evidence 5.3 130 Compress from 199. Keep the control-world false positive — it is what makes the prior a calibration device rather than a handicap
The grounded baseline: sonar 0 letters in 30 runs, 0/60 fabrication at provenance 0.98 5.5, 5.9 80 Study 2 assumes both. They cannot be a citation

Movement 2 — Study 2: stressing the failure socially (~1,860 prose)

Keeps the question headings. This is where the alternative-explanation chain actually earns its keep, and it is the material with the strongest JASSS fit.

Subsection Now Target What goes
Data quality 84 40 Already tight; two sentences, rest to Appendix B
Was the evidence there? (policy gap, 40/40 detectable, Table 6) 1,024 380 The baseline-window robustness sweep (Appendix A.5), the detector false-alarm caveat, and the Study 1 comparability caveat move to synthesis or supplement
Did they interpret it? (Table 7) 503 170 Study 1 half moves to movement 1; keep Table 7 and the "no simulation-class hypothesis anywhere" close
Did scale solve it? 76 60 Nearly untouched — it is already one paragraph doing one job
Do grounded societies communicate? (Table 8) 429 250 Keep the 6,880 agent-days figure, the agent-run interval, and the arm C Elena exception. Cut the pilot action-reason forensics to one clause
Does a minority create a network? (Tables 9, 10, Figure 4) 412 300 Keep depth 1.000 and the pre-freeze prediction — that the pilot predicted depth 1 before the freeze is the pre-registration working, and a referee will look for exactly that
Did it improve convergence? (Table 11) 267 150 Keep the refusal to pool and the reason. Cut the H5b descriptive paragraph to one sentence
What did it transmit? (Table 12) 384 250 Keep the 745 discarded belief changes — the attribution rule cutting against the finding is the most credible thing in the section
Does the minority model matter? (Tables 13, 14) 721 280 Biggest single cut. Keep McNemar 8-vs-0 p=0.0078 and the dosage decomposition. The unpaired ratio, arm E's single CHALLENGED, and the per-recipient table to supplement
Was the institution used? (Table 15) 450 160 Keep the one bulletin post verbatim — it is a delivery check, and that is the whole finding. Cut the pre-registration conflict to two sentences with the detail in the audit trail
Back-dating at n = 8? (Table 16) 707 190 Keep the pooled 123/141 and the conservative-test choice. Both unanticipated observations move to synthesis
Observations not promoted 111 60 Keep the "independent ensemble" tension — it must be presented with the contamination result

Movement 3 — cross-study synthesis (~450 prose)

The part that is stronger than either study alone, and the reason not to demote Study 1.

Content Prose Note
The evidence gradient: Table 5 + Figure 2, z 3.30 → 7.52, detection 7/10 → 10/10, conclusion rate pinned
Three independent zero-communication observations, two scales, with and without an institution 110 Table 8 already carries it
Fabrication dissociation replicates: Study 1 24/60 · 9/60 · 0/60 → Study 2 19 vs 1 under McNemar 90 Independent recurrence in a different design
Back-dating replicates (S1 0.79–0.85 → S2 0.872); unconditional commitment does not scale-invariantly 70 Both directions, honestly

Plus the comparability caveat — S1 B3a and S2 arm A are the same worlds at 5.00 vs 5.77, so cross-study L2 confounds measurement policy with prompt version — which belongs here rather than buried in 5.1, because it is a limitation of the synthesis.

The cost of this structure

The current Results opens by saying it is "reported as questions rather than as H1–H7, so that each result closes off an alternative explanation for the one before it." Splitting Study 1 out breaks two links in that chain: was the prompt prior responsible? (Study 1's ablation) and was the evidence there? (the gradient) now sit in different movements from the ceiling they explain.

Mitigation: movement 1 ends by stating the ceiling and naming the two explanations it has already closed off, so movement 2 opens against a live question rather than restarting. If you would rather keep the twelve-question chain intact and simply starve the Study 1 passages inside it, say so — it is a different edit, not a harder one, and the chain is a genuine strength.


4. Everything else

Related work, 1,783 → 850. Nine subsections averaging 198 words is a survey. JASSS readers want positioning. 2.7 (conformity and cascades, 390) carries an argument the paper depends on and should survive at ~250; the rest collapse into three paragraphs — prior work on evidence non-uptake, on instrumented discovery worlds, and on LLM agent societies and their validity. 2.9 (what this work contributes, 84) folds into the introduction.

Discussion, 1,924 → 1,000. 6.6 (the alternative reading of the contamination result, 496) is the one to protect — pre-empting a referee's strongest objection is worth more than the other seven subsections combined. 6.2 and 6.7 restate results and can go almost entirely.

Limitations + reproducibility, 1,323 → 1,050. Both are protected material. The reduction is compression, not removal: §8.2's three-defect forensics compress to their disclosures with the autopsy in the repository, and Appendix A.0's H4 disclosure folds in at ~120 words.

Appendices. A.1 (frozen hypotheses and verdicts) and A.3 (the pre-registered decision table) become in-body tables in the design section — referees receive only the PDF and no supplementary material, so a pre-registration whose evidence sits in an attachment is an assertion. A.2, A.4, A.5 and Appendix B move out cleanly, along with the Study 1 detail displaced by this cut.


5. What I need before executing

Nothing blocking. Two things to react to if you disagree: the three-movement Results structure over the twelve-question chain (§3, last block), and the assumption that JASSS's word count is prose excluding tables (§1). Everything else follows from your brief.