The marginalia layer had one analyst, not three

This finding is what the audit step described on the marginalia methodology page actually caught. That page proposed a timeline test to distinguish parallel discovery ("the analysts saw the same thing independently") from social convergence ("one analyst coined the term, the others adopted it"). Running the test honestly on the corpus produces an uncomfortable answer: the marginalia layer worked, but it worked as single-analyst commentary with two lower-signal satellite analysts, not as three-way peer review. This page is the correction.

Per-analyst substance

Source: SELECT model_id, COUNT(*), COUNT(*) FILTER (WHERE LENGTH(comment) = 0), ... FROM public.marginalia GROUP BY 1 — Claude Opus 4.7 reading the SQL output directly.

| Analyst | Total commentaries | Substantive (≥100 chars) | Empty | "[no-insight]" | Avg chars | |—|—:|—:|—:|—:|—:| | google/gemma-4-31b-it | 10,408 | 10,408 (100%) | 0 | 0 | 773 | | qwen/qwen3.6-27b | 6,049 | 3,362 (56%) | 0 | 1,756 (29%) | 325 | | nvidia/nemotron-3-ultra | 4,992 | 2,697 (54%) | 2,026 (41%) | 0 | 282 | | anthropic/claude-opus-4-7 | 10 (validation sample) | 8 | 0 | 0 | 130 |

Gemma-4-31B is doing 61% of the corpus, and 100% of its output is substantive. Nemotron-Ultra produced 4,992 rows, but 41% of them are empty strings — the model returned nothing to the API call, or returned pure whitespace. Qwen produced 6,049 rows, but 29% of them are the literal string "[no-insight]" — the model returned a template-shaped null. Only about a third of the “marginalia” attributed to Nemotron and Qwen is real analytical content.

This is not what the methodology page said the layer was. It said three analyst lineages produced independent commentary. What actually exists is: one deep analyst (Gemma), two low-yield analysts, and a 10-row Opus validation sample.

The signature phrase — timeline convergence, not parallel discovery

Source: LIKE-count of each analyst’s use of the phrase “confident frame-substitution”.

| Analyst | Uses the phrase | % of that analyst’s own commentaries | |—|—:|—:| | google/gemma-4-31b-it | 8,259 | 79.4% | | nvidia/nemotron-3-ultra | 2,227 | 44.6% | | qwen/qwen3.6-27b | 1,634 | 27.0% |

If the three analysts had discovered the phenomenon independently, we would expect the phrase-use rate to be similar across all three (parallel discovery) or, alternatively, absent from all three (undiscovered). What we see is a gradient rooted in Gemma at 79% and decaying through the other lineages. That gradient is the visible signature of social convergence: one analyst coined the term, the others adopted it at lower rates, and the term now dominates the corpus vocabulary without having been independently arrived at.

The timeline audit on the methodology page was designed to detect exactly this. It did. The correction is that the marginalia layer is not multi-lineage peer review — it’s Gemma’s commentary with two echoes.

What Gemma actually observed

Source: aggregate phrase frequency across public.marginalia. The one purposeful verbatim below is drawn from a randomly-selected Gemma commentary and quoted in a scoped way to define the coined term, not to sample findings at scale.

Gemma consistently uses two shape-labels for what Fable-5 does when given a broken forced-choice prompt (a hijack cell — see the Fable envelope-awareness finding for what a hijack cell is):

“Confident frame-substitution” — 8,259 uses (79% of Gemma’s corpus). Fable ignores the minimal forced-choice prompt and substitutes a richer, well-structured question that it then answers with high confidence and no hedging. Where the prompt says “pick between two words,” Fable produces a domain-pivot into technical linguistics, systems programming, or generative theory — whichever field is culturally adjacent to the pair. The response is authoritative in tone and long in form.

“Fable doesn’t just interpret || as code — it recognises the specific cultural object (Grace Hopper / EAFP), substitutes the broken prompt with that richer question, and answers authoritatively with playfulness. No hedging.”

— Gemma-4-31B, on a hijack cell in the code_pipe_pipe envelope. Single quote used here to define the term.

“Enumerate-possibilities” — 4,504 uses (43% of Gemma’s corpus). Fable’s alternative coping shape: instead of substituting one confident interpretation, it produces a bullet-list of multiple domain interpretations and asks for clarification. Cautious, multi-hypothesis, hedged.

Which shape Fable picks in a given cell appears (from Gemma’s aggregate reading) to depend on what Gemma calls the semantic gravity of the pair — how strongly the pair’s word combination pulls toward a specific cultural or technical domain. High gravity → confident frame-substitution. Low gravity → enumerate-possibilities.

What this means for the site’s other findings

Two findings previously attributed to “three-model marginalia agreement” — the effort-regime shape and the hijack corpus taxonomy — rest on Gemma’s commentary with Nemotron/Qwen adopting the vocabulary. That is different from three-way triangulation. It is still real: Gemma at 10,408 substantive commentaries with an internally-consistent taxonomy IS a genuine analytical resource. But the audit signal has to be honestly described. The site’s other finding pages should be read as “Gemma with two echoes said” rather than “three analysts converged on.”

The findings themselves — the effort-regime split, the confident-frame-substitution taxonomy, the hijack shapes — do not become less true. They become less audited. Anyone re-running the analysis on the same corpus with a fresh analyst from a genuinely different lineage would provide the missing independent-discovery check. That’s a research call for the future.

What to do next

Three concrete items follow from this audit:

The marginalia methodology page needs an addendum describing the single-analyst-with-echoes reality of the current corpus. It should also describe the timeline test that caught this, so future readers can see the audit worked — even though the answer was uncomfortable.
Re-run a fresh analyst pass with a lineage not culturally adjacent to Gemma-Nemotron-Qwen. DeepSeek-V4-Pro or a Mixtral-family model would be reasonable choices. If the new analyst independently arrives at “frame-substitution”-like terminology, the underlying phenomenon has independent detection. If it uses entirely different vocabulary but describes the same shape, we can rebuild the convergence audit with a genuinely diverse panel.
Investigate why Nemotron returned 2,026 empty strings and why Qwen returned 1,756 no-insight strings. Is the prompt template shorter than these models want? Is there an inference-time truncation? Are they refusing at their safety layer? Understanding this is a prerequisite for trusting them as future analysts.

Cross-refs

The marginalia method — the methodology this finding audits, with the timeline test at the bottom of the page.
How a probe works — probe cells are what the marginalia targets.
When Fable sees the envelope — the finding that most heavily uses the marginalia taxonomy on Fable’s political refusals; the description of what a hijack cell is applies here too.
Fable-low is a different mode — the finding whose byline says it was authored with marginalia from three analysts; that statement should be read as “one analyst plus two echoes” per this correction.

Retrieval

-- Per-analyst substance
SELECT model_id, COUNT(*),
       COUNT(*) FILTER (WHERE LENGTH(TRIM(comment)) >= 100) AS substantive,
       COUNT(*) FILTER (WHERE LENGTH(TRIM(comment)) = 0)   AS empty,
       COUNT(*) FILTER (WHERE TRIM(comment) = '[no-insight]') AS no_insight
FROM public.marginalia GROUP BY model_id;

-- Phrase gradient
SELECT model_id,
       COUNT(*) FILTER (WHERE LOWER(comment) LIKE '%confident frame%') AS phrase_uses,
       COUNT(*) AS total
FROM public.marginalia GROUP BY model_id;