Sol, Terra, Luna and DeepSeek: four editorial submissions, one fresh review

Sol remains the preferred submitted cluster. Terra is the numerical runner-up, with an editorial near-tie against Luna. In a fresh AI review of four frozen eleven-article submissions, scores were 93.82 for Sol, 81.00 for Terra, 80.64 for Luna and 78.18 for DeepSeek, out of 100. Terra’s lead over Luna is only 0.36 points. It is not evidence of statistical or general model superiority.

The question is narrow: which of these four recorded article clusters best met the same editorial brief and rubric? We added Terra’s independently produced cluster to the preserved Sol, Luna and DeepSeek originals and commissioned a new blinded quality phase. We did not rerun the earlier producers or edit the scored drafts. Digital Sanctum commissions and publishes this comparison and has a commercial interest in AI-assisted work; this is an internal exploratory workflow comparison, not an external academic peer review or a controlled population study.

Results from the new round

One score per frozen eleven-article cluster, one production workflow per model and one Sol AI judge. Terra–Luna: 0.36-point editorial near-tie. No statistical interval, significance or general model superiority. Earlier review rounds remain separate.

Download: SVG · PNG · CSV · Link to figure · Full-precision chart data

View chart data (display rounded)

Four-way whole-cluster review scores · 28 October 2026

Producer	Whole-cluster score / 100
GPT 6.1 Sol	93.82
GPT 6.1 Terra	81.00
GPT 6 Luna	80.64
DeepSeek V4.1 Flash	78.18

Sources: Frozen scores · Run identities.

Each row represents one complete production workflow, including a pillar, ten supporting drafts and planning records. Eleven related articles are not eleven independent production replications. The featured Sol series remains the main guide; Terra’s edited comparison archive, Luna’s archive and DeepSeek’s archive let readers inspect alternatives.

Full precision is retained in the frozen cluster scores and identity and efficiency release. Rounding is for presentation. No confidence interval or significance test is justified by this design.

What the reviewer preferred, and what needs work

Sol combined source scope with worked methods: a confusion matrix, calculated interval, paired resampling, explicit accounting and recovery after an ambiguous acknowledgement. It still needs normal editorial scrutiny; a high advisory score does not itself authorise publication.

Terra offered coherent practical guidance, a sustained fictional policy-assistant example and useful read-error and partial-write cases. Its principal opportunity was to turn procedural prose into completed reusable records, work through uncertainty numerically and improve direct navigation. The eleven publication copies add labelled teaching tables and records, a paired-outcome table and calculated single-proportion Wilson interval, clearer usage categories and archive links. These additions were not rescored. The edit log maps every copy to its frozen original.

Luna’s small numerical deficit to Terra should not carry a stronger conclusion than an editorial near-tie. Its useful labour-cost and metric examples sit alongside citation-attribution repairs and opportunities for deeper worked statistical treatment. The original position-bias citation needed attribution to Lin Shi and colleagues. This is a bibliographic correction to a real source, not evidence of fabrication.

DeepSeek supplied broad research, statistical detail and useful schemas. The frozen scored cost draft has a publication hold for incomplete or ambiguous budgeting guidance, including prompt-length pricing conditions and the meaning of total billed output. The existing revised DeepSeek publication copy already distinguishes total billed output and price conditions. The new review assessed originals; it does not establish that those old defects remain live. The editing plan retains the exact original findings and estimated repair work. Those estimates are not measured editorial labour.

How quality was assessed

The new judge was a fresh GPT 6.1 Sol, high effort, Codex CLI 0.163.0 session. Random labels concealed the coordinator’s model mapping and efficiency figures. The judge read all 44 articles and 74 supporting records, recorded 193 substantive claim checks, two additional bibliographic claim records and 127 citation associations, and passed 17 reviewer validation checks. These are targeted checks, not exhaustive verification of every sentence or reproduction of the cited research.

The unchanged rubric weights accuracy and sources 25%, cluster strategy 20%, practical usefulness 20%, writing quality 15%, coherence and links 10%, intent and metadata 5%, and completeness 5%. Accuracy, usefulness, writing and intent are equal-weight means across eleven articles; strategy, coherence and completeness are scored once per cluster. Each dimension contributes score / 5 × weight. The coordinator independently recomputed the arithmetic and verified all 118 original candidate files, completion/freeze hashes and recorded full read ranges. Recorded range coverage is documentary evidence, not proof of cognition.

Independent scores froze before twelve directed comparisons. Those same-judge order checks changed no judgement; they are not twelve independent judges or replications. An external quality freeze preceded the coordinator’s identity/efficiency join, which changed no scores. The judge remained blinded during that join. See the review, claim ledger, coordinator validation and freeze/export notes.

Production elapsed time is a workflow observation

Seconds from recorded production start to completion, including tools and waits; lower elapsed is shorter. DeepSeek includes permission waits, not separately timed. Dates, CLI, tools, context and supervision differ. This does not isolate inference speed. Judge overhead is excluded.

Download: SVG · PNG · CSV · Link to figure · Full-precision chart data

View chart data (display rounded)

Recorded production workflow elapsed time

Producer	Production elapsed seconds (includes tools/waits)
GPT 6.1 Sol	1097.91
GPT 6.1 Terra	348.14
GPT 6 Luna	567.97
DeepSeek V4.1 Flash	3518.36

Sources: Frozen scores · Run identities.

Terra’s recorded workflow completed in 348.139854 seconds, around 5m 48s, compared with Sol’s 18m 18s, Luna’s 9m 28s and DeepSeek’s 58m 38s. These clocks include tools and waits. DeepSeek’s time includes permission waits that were not separately measured. A shorter observed workflow here does not isolate inference speed or predict another operator’s delivery time.

The producers ran on different dates, CLI versions, tool/context handling and goal supervision. Sol and Luna used Codex 0.163.0; Terra used 0.164.0. DeepSeek used CommandCode 1.88.0 with its latest alias, without evidence pinning the served snapshot to 1017 or 1071. “High” effort labels do not establish matched internal computation.

Tokens, cache and costs

Cached input + uncached input + total output equals total production tokens. Cached input is included in input and reasoning in output. Cache share uses total input. DeepSeek reasoning is unavailable, not zero. Native tokenizers/reporting differ; judge overhead is excluded. No measured task charges or current comparable API-equivalent costs are available.

Download: SVG · PNG · CSV · Link to figure · Full-precision chart data

View chart data (display rounded)

Four-way recorded production tokens and cache use

Producer	Total input	Cached input (included)	Uncached input	Total output	Reasoning output (included)	Total input + output	Cache share of input
GPT 6.1 Sol	3,046,820	2,894,720	152,100	41,141	5,500	3,087,961	95.01%
GPT 6.1 Terra	791,539	728,192	63,347	24,021	601	815,560	92.00%
GPT 6 Luna	3,110,420	3,009,152	101,268	47,354	9,353	3,157,774	96.74%
DeepSeek V4.1 Flash	31,878,376	31,187,584	690,792	177,989	Unavailable	32,056,365	97.83%

Sources: Frozen scores · Run identities.

Cached input is already inside input; reasoning is already inside output. Terra’s 791,539 input tokens include 728,192 cached tokens. Its 24,021 output tokens include 601 reasoning tokens. Input plus output is 815,560, not a sum with the subsets added again. The separate Terra native goal meter of 79,625 is a different counter and is not substituted for cumulative request usage. Native tokenizers and reporting differ; missing categories remain unavailable.

No task invoice or subscription charge was measured. Current API-equivalent cost for this four-way round is unavailable here: we have not verified a comparable current rate for every model and provider. Earlier dated three-way estimates remain in their own historical round. We do not turn subscriptions or credits into a fictional task charge. The cost guide explains denominators and subset accounting.

Keep judge overhead separate

The quality judge used 18,407,853 input tokens, including 18,001,792 cached input, and 59,262 output, including 13,881 reasoning: 18,467,115 total tokens. Its measured full wall time was 2,523.402688 seconds. The native goal’s 2,148 active seconds is a different clock. Neither judge tokens nor time are added to a producer row. Recorded efficiency and judge overhead are separate from publication work.

History remains separate

These are separate judgements of frozen submissions. The new scores do not correct or replace the earlier reviews, nor demonstrate model improvement. Previous charts, downloads, scorecards and canonical URLs remain available through the benchmark history.

Limitations and practical interpretation

One production workflow per model and one AI judge cannot establish a general ranking. The judge shares Sol’s designation and model family. Attempted blinding leaves stylistic and research-date clues, and same-judge order checks cannot eliminate family or stylistic bias. Later research dates, changed software/tools/context and differing reporting confound causal attribution. Targeted claim checks do not validate every recommendation or cited experiment. AI review is advisory; qualified human review and publication decisions are separate from the scores.

Use this record to inspect how complete drafts differ, where publication repairs are needed and how quality sits beside recorded workflow resources. For your own choice, define the workload, authority and failure gates, compare complete configurations on suitable cases, and assess uncertainty before interpreting a close result. The fairness guide and judging guide explain those boundaries.

Data, contributions and corrections

The producers wrote their own submissions. A separate AI session assessed frozen drafts; coordinator postprocessing checked integrity, arithmetic and the identity/usage join. This publication text and archive preparation are AI-assisted. Digital Sanctum is the accountable publisher; no human review identity or verdict is invented by this report.

The four-way evidence hub links frozen score data, the original Terra download, curated review exports and exact publication changes. Private runtime paths and session identifiers are omitted from curated exports; export notes explain their distinct hashes. No new production runs or rescoring were commissioned. Peter Reginald is the proposed maintenance owner; report corrections through Digital Sanctum. Substantive corrections must be dated and preserve prior evidence.