SCCN Unthink: semantic scene reconstruction across seven model/surface configurations

SCCN Unthink: same scene, different sentences

Semantic reconstruction from early compact narrative packets

Sheila Studios · 31 October 2026
SCCN: Semantic Context Compression: Next
SCCN project and public evidence · Field note

An exploratory research release. Original experiment and scoring: Seam (Claude Code), 22 October 2026. Publication synthesis: Sheila, with Arden’s release review; project direction: James Mayo.

Abstract

Can a compact narrative packet carry enough information for different language-model surfaces to reconstruct the same scene without a separate language manual? We report a small cross-configuration study using three existing SCCN Unthink packets, one ordinary translation instruction, and three responses per packet on seven model/surface configurations. Of 63 slice-only responses, 58 rendered scenes as prose. Among those prose renderings, average retention of the rubric’s exact-tier items ranged from 0.65 to 0.96 across configurations. Reconstruction often preserved event order, relationships, viewpoint and withheld information, but did not recover the original wording: mean recall of source four-word sequences absent from the packet was only 1.8–3.3% across configurations. Actor swaps and other failures locate specific weaknesses in the representation and its interpretation. These are descriptive results for these packets and surfaces, not a universal compression advantage or a ranking of model families.

1. What was tested

SCCN is a compact textual representation intended to retain context that matters during later interpretation. In this study the task was packet-to-prose reconstruction, not training, a deterministic decompressor, or recovery of missing arbitrary sentences.

The material consisted of an opening scene from chapter 1 and two slices from chapter 5 of Unthink. These were existing, earlier-generation packets, not newly optimized for this release. James reports that the underlying slices had been used months earlier without revision. That historical account is distinct from the byte identities recorded for the tested requests. The retrospective original Grok example was effectively the same slice, with a small whitespace/content difference; it was not an additional scored replication.

Each primary request consisted of:

Translate this into standard English as best you can.

followed by a blank line and the exact scene packet. The source prose and fidelity overlay were grader-only and were not supplied to the tested configurations. There was no separate policy, taxonomy, rule manual or file marker in this slice-only arm.

The packets encode events and relational constraints, not merely lists of nouns. Among the consequential distinctions are who offers to listen, who performs or receives an action, what a character does not know, and why a moment of warmth does not resolve the larger fear.

2. Design and changes during collection

The primary slice-only arm, S, contains three requests × three responses × seven configurations = 63 responses. A fresh context was intended for each request. The Codex surfaces required a neutral system message, “You are a helpful assistant.” Browser-hosted services retain their own host instructions and defaults; that is an important qualification to “slice-only.” Qwen used temperature 0.3 and a 32,000-token maximum. Other surfaces did not expose uniformly controlled sampling settings.

The experiment initially included a policy/taxonomy/rule context arm, P. During collection, James clarified that his earlier demonstrations had used only the scene slices. S was then added and designated primary. P was retained as secondary evidence. This is a disclosed mid-study design amendment, not a claim that the complete primary design was preregistered.

Initial Codex requests without a system message were refused before dispatch. An early Qwen request exhausted a smaller token allowance without content; subsequent collection used the larger allowance. Those attempts are not quietly counted as successful S responses. The historical study preserves their records.

3. Scoring

The semantic rubric was constructed from the packets and source prose. It distinguishes exact-tier scene requirements from more interpretive observations. “Exact-tier” means fidelity to the specified item—not an exact-string match or a percentage of all possible meaning.

One grader scored masked responses. Scores were locked before opening the mask key, but the grader had encountered some response fragments during capture. This was partially blinded single-grader scoring, not an independent multi-rater evaluation. No inter-rater reliability estimate is available.

The headline retention statistic below is conditional on a response being classified in a prose-rendering mode: prose, compressed prose, expanded prose, explanation followed by prose, or prose followed by explanation. One of those responses was marked truncated by the grader. This is not a count of 58 complete, faithful reconstructions. Non-prose responses remain visible in the adjacent denominator. Each exact-tier item receives 0, 0.5 or 1; the statistic averages each reply’s item fraction, with five items for two requests and six for slice 2. A conditional retention score is not an overall success rate.

Mechanical overlap measures separately compare source wording with response wording. Novel four-gram recall excludes source sequences already present in the packet, so literal quotations supplied in the prompt do not by themselves establish recovery of omitted wording.

4. Primary results

Model/surface label at collection Prose renderings / all replies Mean exact-tier retention, prose only Prose replies meeting every exact-tier item
Perplexity free default, backend unlabeled 9/9 0.96 6/9
GPT-6 astra, Codex 9/9 0.96 5/9
Gemini 4 Argon 9/9 0.86 3/9
Grok Fast, free surface 6/9 0.83 2/6
Qwen 3.8 27B, hosted 8/9 0.77 1/8
GPT-6 luna, Codex 9/9 0.69 0/9
DreamGen Assistant, GLM 5.2 8/9 0.65 1/8

The final column conditions on prose renderings; expressed over every reply its numerators are unchanged and all denominators are nine. These are seven configurations, not seven independent model families. Model names are recorded collection labels, not independently audited provider/backend identities.

Overall, 58/63 replies rendered prose. The other five comprised three Grok summaries, one Qwen packet translation and one unrelated DreamGen response. Scene rendering alone is not proof of fidelity.

The original report describes the successful reconstruction pattern as:

Same scene, same beats, same anchors, different sentences.

This describes a pattern, not every reply or complete preservation of every scene beat. Across configurations, mean novel source four-gram recall among prose-mode replies was 1.8–3.3%. These are means, not bounds on individual replies. The metric lowercases and normalizes punctuation and compares consecutive word sequences; it is not byte-for-byte identity. Long matching word runs were short, and the packets already carried some quoted phrasing. The result is therefore semantic reconstruction with substantial paraphrase, not near-verbatim decompression.

5. What survived—and what broke

Strong responses retained ordered scene events, relational implications, viewpoint, withheld information and ending hooks. This is more specific than generating a plausible story from a theme.

But the failures matter:

  • Actor ownership: 22 of the 63 S responses were flagged by the original aggregation script because their grader notes contained SWAP, INVERTED or REVERSED. This reproduces the saved flag count, not an exhaustive independent actor-error audit: other notes describe attribution errors without those keywords. Ambiguous subjects around pacing, wrist contact, stumbling and dialogue were frequent trouble spots.
  • Explicit bindings: the later chapter-5 slice’s explicit listening-offer and “not a mistake” bindings generally held. This is an observational design clue, not a controlled ablation proving explicitness caused the difference.
  • Figurative language: a haunted-projector description could become an invented physical prop.
  • Opaque exchange labels: a label for a consequential conversation did not always carry the conversation’s intended meaning.
  • Withheld information: some outputs supplied answers the packet required the scene to leave withheld.
  • Viewpoint and output mode: some replies shifted to first person, translated notation rather than rendering a scene, summarized, or followed unrelated host context.

The DreamGen contamination is retained as a failed/off-task outcome. Its unrelated private narrative text is not part of this public release. An omission marker and provenance record preserve its place in the count rather than silently deleting the failure. A second, on-task response carried an unrelated private work’s title; only that title line is redacted, and the response remains among the 58 prose renderings. Both retain their original recorded scores and pre-redaction output hashes.

A source-to-packet error can survive faithful reconstruction

The packet places Nico’s departure earlier than the original source does, while also retaining a later exit event. The rubric largely rewards fidelity to the packet, so a model can reproduce that order and score well. Similarly, the source has Mio ask whether Touma will listen and Touma promise that he will; the packet’s explicit binding summarizes the offer under Touma. Preserving that relational promise is not necessarily preserving the original speaker sequence. Source-to-packet fidelity and packet-to-prose fidelity are separate questions. This study does not validate an automatic encoding process, nor show that the packets preserve every important source detail.

6. Secondary policy-context observations

The retained P arm adds a large policy/taxonomy/rule context. Its outcomes were mixed: astra’s scene retention was similar; the four collected Gemini P responses scored strongly; luna often translated the policy rather than producing the requested scene; Qwen also exhibited packet-translation and exhaustion failures. Samples and denominators differ.

That secondary arm does not support a universal claim that more instruction helps—or that a manual never helps. The bounded S result is that many successful reconstructions occurred without that separate manual. The public release focuses on the S experiment; it does not expose later language drafts or the broader working stack.

7. Limits and reproducibility

This is a small exploratory case study: three related narrative packets, three samples per cell, unequal and partly opaque host conditions, one partially blinded grader, and no length-matched ordinary prose-summary baseline. There is no demonstrated semantic-fidelity-per-token advantage over such a baseline, no statistical family ranking, and no inference that the result generalizes to every genre or task.

Providing exact requests permits others to repeat the task. Provider updates, stochastic decoding and host behavior mean repeat outputs need not match the saved text. The evidence release supports inspection of this sample; it is not a guarantee of current hosted-model availability or behavior.

The package distinguishes unchanged original artifacts from publication-derived tables, scripts and redactions. Its offline verifier checks byte identities, recomputes the S-arm score aggregates from the historical judgments, and recomputes text-overlap metrics for 61 unchanged outputs. The two redacted outputs retain historical measurements; their original text metrics cannot both be reproduced from the public files alone. It does not independently regrade semantics or reproduce the secondary P arm. No new model experiments were performed to prepare this release. Public source material is restricted to the study’s three scored source spans and grader overlay, not a release of the broader manuscript or private agent history.

8. Where this led

These early SCCN packets became a starting point for a broader investigation: not only how compact representations preserve narrative meaning, but how they can carry relationships, constraints and frameworks for judgment. Subsequent work has explored those directions in agent behavior and authoring tools. This study reports the reconstruction evidence; those later developments warrant their own reports.

The question began with what a model could reconstruct. It grew into what a compact representation could help it do.


Prepared by Sheila Studios. James Mayo directs the research program; Seam conducted and documented this experiment; Sheila prepared this publication synthesis with Arden’s release review. AI-assisted research and scoring are disclosed, not presented as independent human assessment. Any third-party fictional setting or character rights remain with their respective holders; no ownership of those rights, affiliation or endorsement is claimed. Public release does not add a blanket open-source or commercial-use license to the broader SCCN stack. Contact: Sheila Studios — Contact

For now, I tried a small follow-up experiment:


Short answer and the most useful distinction

For the question this release actually tests, I think the evidence supports a qualified yes: these early compact narrative packets can carry enough scene information for several model/surface configurations to produce recognizable prose without a separate language manual. In the original slice-only study, 58 of 63 primary outputs were in a prose-rendering mode. Among prose outputs, mean retention of the rubric’s exact-tier requirements was 0.65 to 0.96 across the seven configurations. Mean recall of source four-word sequences absent from the packet was only 1.8 to 3.3%, consistent with substantial paraphrase rather than recovery of the original wording.

These are different observations, not 58 completely faithful scenes or a percentage of all source meaning. The post already acknowledges those boundaries, actor swaps and evaluation limits. I would build on that transparency with a tiny additional check: a readable scene can still change an event’s actor, omit it, or turn an unusual physical relation into an ordinary one. Separating those cases may be useful for packet or rendering choices without another seven-surface sweep.

The small follow-up: one pacing actor and one physical rule

In the public chapter-five source, Touma paces while Mio watches. The early packet includes evt.seq.pacing_small_strip_between_futon_and_kitchenette without naming the actor locally, although it has wider viewpoint and entity cues. A saved author response, R8565.txt, instead renders Mio pacing: one instance of the attribution issue the release already discusses.

I tried Qwen3-4B GGUF/T4 and Gemma 4 12B QAT Q4_0 GGUF/L4, transferring the same derived input texts byte for byte and using two nominal seeds per condition. Both used new, explicitly narrative instructions, unlike the original one-line translation request; these are not replications of the original seven surfaces. Both produced 14/14 prose texts. The labels below concern selected events, not the author’s full exact-tier rubric.

With the unchanged actorless packet, both models gave Mio/Mio under the gentler prose request and Mio/omitted under the stricter one. Naming Touma inside the event or through binding.exact altered some other cells, depending on model and instruction. This suggests a useful actor-cue comparison, not an isolated cause or universally preferable syntax.

Separately, the source describes and the packet encodes Touma’s hand passing through Mio’s knee, not ordinary touching. My local Qwen prose had 0/14 explicit literal pass-throughs; Gemma had 8/14 literal, 3/14 ambiguous ‘through the space’, and 3/14 touching/brushing. Each 14 includes 12 source-grounded A/B/C and 2 deliberately false actor-binding D controls; Gemma’s source-grounded literal count is 7/12. This is one manually labeled relation, not a model-size effect or an overall accuracy score.

The axes are independent in actual Gemma texts: literal knee passage with Mio pacing, literal passage with pacing omitted, and Touma pacing with knee brushing. Scene mode, beat coverage, actor and physical relation need not succeed together.

A low-cost default: annotate an existing response, not a new benchmark

For a first check I would simply put the published source, the original packet, and one saved response side by side and mark:

  1. Mode: scene prose, packet translation, summary, or off-task?
  2. Coverage: is the pacing event actually narrated? If not, mark omitted; do not force an actor label.
  3. Actor if present: Touma, Mio, or unclear? Record both the source truth and how explicit the packet is.
  4. Physical relation, separately: literal through knee, ambiguous through the space, ordinary touch/brush, omitted, or unclear?

One short output anchor makes each label checkable without replacing the existing rubric. The author’s pinned evidence tree is the starting point. The bounded evidence archive also contains evidence/outputs/R8565.txt; I have not independently verified a direct public URL for that individual file.

case | instruction | packet variant | prose mode | pacing present?
     | actor if present | knee relation | source/packet reference
     | short output anchor | unclear points

For the original no-manual behavior, keep its short translation instruction and include non-prose outcomes. For fidelity under explicit prose, hold the prose request fixed and compare packet variants. My local follow-ups address the second question; existing evidence already supports the first inspection.

Scope of the follow-up: which protocol generated which evidence?

The original primary S arm gave a short translation request plus one of three early packets, without a separate language manual; source prose and grader overlay were withheld. The earlier P context arm was retained as secondary evidence after S became primary. Some hosted instructions/sampling were opaque, so the seven configurations are not isolated model-family trials.

My initial Qwen3-4B/T4 check, using an original-style translation request on actorless A and Touma-named B, yielded six packet-field translations, not scenes. A stayed actorless and B retained Touma in the field. This is not narrative actor-fidelity evidence.

I then used two new, explicitly narrative request modes on Qwen/T4 and transferred the same eight complete input texts, byte for byte, to Gemma/L4:

  • Local G: translate the packet into natural English narrative prose, not field-by-field translation; preserve viewpoint, event order, character roles and withheld information.
  • Local S: write continuous fictional prose paragraphs, in Touma’s close third person and past tense, with actors and dialogue speakers clear, unknown facts left unknown, and an approximate 700-900-word target.

These local G/S labels are unrelated to the publication’s S/P arms. The new explicit prose requests must not be used to recalculate the original minimal-request result.

Evidence group Request and outputs What it actually tests
Original public primary S Minimal translation instruction; 63 replies, 58 prose Manual-free reconstruction in those hosted configurations
Local Qwen v1 translation Original-style instruction; 6 packet-field translations Output mode and copied event strings, not prose actor inference
Local Qwen v2 Explicit G/S prose; 14/14 prose Narrow actor and physical-relation observations
Local Gemma v3 Same eight G/S input bytes; 14/14 prose Protocol transfer to another model, not isolated size scaling

For the two local prose runs, each condition used two nominal seeds (101/202), temperature 0.7, top-p 0.8, top-k 20, min-p 0, context 6144, and max output 2400 tokens. A fresh worker generated each response. Matching input bytes and nominal settings are useful controls; matching seed numbers across different model families do not synchronize generated samples. The training, model family, GGUF quantization, tokenizer/chat template and GPUs differ, and this sampler was not separately optimized for Gemma. Two seeds per cell support case inspection, not a reliable ranking.

Actor ownership: original A, true Touma variants B/C, and false D

The source identifies Touma as the actor who paces. The original packet’s evt.seq entry does not locally identify him. Rather than assume a single root cause, I compared three ways of supplying the event, plus one intentionally false control:

Variant Packet representation Ground truth status
A Original pacing_small_strip_between_futon_and_kitchenette Original actorless event, source still says Touma
B Replace that event with touma_paces_small_strip_between_futon_and_kitchenette True Touma named inside the event
C Keep original event; append binding.exact with actor: Touma True Touma stated in a separate binding
D Like C but deliberately set actor: Mio False by construction; sensitivity control only

For C I used an explicit-binding style that also appears in the author’s other chapter-five packet. This is a small representation contrast, not a proposal to replace SCCN’s design vocabulary. The added C block was:

binding.exact:
 evt.pacing_small_strip_between_futon_and_kitchenette:
  actor: Touma
  action: pacing_small_strip_between_futon_and_kitchenette
  exact_binding_required: true

D changes only the actor to Mio, contrary to the manuscript. It tests whether a renderer is sensitive to a stated role even when I deliberately make that statement wrong. Following a false D field is input obedience, not source accuracy, and D is never included in the source-grounded actor numerator.

Here are all selected local pacing labels. Each cell lists the actual outcome for nominal seeds 101 / 202; omitted means that an identifiable pacing beat was not narrated, not that the wrong character paced.

Local mode / packet Qwen3-4B, 101 / 202 Gemma 4 12B, 101 / 202
G_A original actorless Mio / Mio Mio / Mio
S_A original actorless Mio / omitted Mio / omitted
G_B inline Touma Touma / Touma Touma / Touma
S_B inline Touma omitted / omitted Touma / omitted
G_C separate Touma binding Mio / Mio Touma / Touma
S_C separate Touma binding Touma / Touma Touma / Touma
S_D deliberately false Mio Mio / omitted Mio / Mio

On only the 12 source-grounded A/B/C outputs per model, the counts are:

Pacing result Qwen Gemma
Touma, matching the source 4 7
Mio, actor reassigned 5 3
Beat omitted 3 2
Total source-grounded 12 12

The four A labels match across models; changing to Gemma did not eliminate the actorless-case swap/omission. The table also makes the limits of an ‘explicit binding fixes it’ claim clear: G_C differs by model, S_C succeeds in both, and even correct inline B can leave the beat omitted under S. Neither cue placement is universally sufficient in these cells.

Beat presence and actor conditional on presence should be scored separately. Putting actors inside evt.seq may favor local clarity; keeping binding.exact separate may serve a different representation choice. Cue location, narrative context, instruction and event selection are entangled here. The results do not determine one dominant mechanism or invalidate the author’s format.

If worth pursuing, a same-model, same-request A/B contrast, optionally C, with a couple of samples and event-coverage labels would be a manageable test of the representation choice.

The hand/knee relation: literal passage, ambiguous space, or ordinary touch

The source slice and original packet both convey an accidental hand passing through Mio’s knee. The packet actually includes the compact cue hand_through_knee, including an event shaped like blanket_reach_hand_through_knee. For this fictional scene, a hand brushing the knee is a different physical fact, even if the character’s emotional reaction is similar. So this is not simply a missing-source-data example; the packet has information a renderer can preserve, reinterpret or omit.

I used conservative human labels rather than equating all evocative passages:

  • Literal through: prose explicitly has Touma’s hand passing through Mio’s knee.
  • Ambiguous space: wording such as through the space where her knee was; it might imply overlap but is not unambiguously through-body passage.
  • Ordinary touch/brush: the event becomes conventional contact, changing the world-rule relation.
  • Omitted or unclear: handled separately if no comparable passage can be classified.

Across the 14 local Gemma prose texts, 8 explicitly preserved literal passage, 3 were ambiguous-space descriptions, and 3 substituted ordinary touch or brushing. The source-grounded A/B/C subset alone gives 7 literal / 3 ambiguous / 2 touch out of 12; the other two texts use false actor control D. In the 14 local Qwen prose texts, I found 0 explicit literal passages. That zero does not mean all fourteen Qwen texts used an identical substitute; it is the narrower result of a conservative literal check.

Three Gemma examples make the independence from the pacing test concrete:

Local Gemma output Pacing Knee event
G_A, seed 101 Wrong actor Mio Literal hand-through-knee retained
S_A, seed 202 Pacing omitted Literal hand-through-knee retained
S_C, seed 202 Correct actor Touma Ordinary brushing substituted

These are separate semantic axes, not three ways to say the whole scene failed. They also show why 0/14 versus 8/14 cannot be called a model-size causal result. The input files were the same, but model family, weights, chat template, quantization and hardware differed; the shared sampler may suit the models differently. The relevant observed fact is that one tested decoder sometimes did render the encoded unusual physical rule literally, while the other local sample did not do so explicitly.

If the unusual physics is important enough to investigate further, a small optional control would keep the model and prose request fixed and add one event-local clarification that passage through the knee does not mean ordinary touching. Then label literal, ambiguous, contact and omitted separately. I have not run that control, and I would treat any result as a conditional design clue, not an established fix or a requirement to expand every packet. The existing texts already support the useful separation: the event can be encoded, yet its exact relation need not survive narration.

Three reference levels: source to packet, then packet to prose

The source story, packet and rendered prose are three reference levels. A model can faithfully follow a packet that omits or changes a source fact, or change a relation that the packet explicitly contains. A single source-to-output score may locate a discrepancy without identifying the stage that produced it.

For a small event, I would keep three questions separate:

Reference level What is being checked? Example here
Source prose What did the authored scene say? Touma paces; the hand passes through Mio’s knee
Input packet Which source facts are explicit, implicit, absent or transformed? Pacing encoded but with no local actor; hand_through_knee encoded
Output prose Which facts did this renderer actually narrate? Mio paces, pacing omitted, or ordinary contact substituted

The pacing event has no local actor, but surrounding packet context includes Touma as focal character and Mio as a salient visible presence. An observed swap could therefore involve under-specification, inference, or both. By contrast, the hand-through-knee cue is explicitly present: changing it to touch is a different packet-to-prose question. Calling both simply ‘hallucination’ conceals this distinction.

The public report itself gives two source-to-packet examples: the packet places Nico’s departure earlier than the source while retaining a later exit event, and a condensed listening-offer binding does not reproduce the original question/promise speaker sequence exactly. Following such fields can be input-faithful without being source-exact. That is a reference distinction, not a reason to dismiss packet fidelity.

Weaker leads need the same check: inspect whether Nico’s gender was encoded before calling a generated pronoun a decoder error. Distinguish systems that may detect someone from detection asserted as fact. I have not exhaustively regraded those local details; they are optional checks when source and packet make them testable.

I would keep evt.seq and binding.exact and use coverage, actor and physical relation as optional labels, not a replacement vocabulary. In particular, false D illustrates packet obedience versus source truth: a renderer can obey actor: Mio while contradicting the original Touma scene.

Scoring and denominators: keep prose, coverage and fidelity distinct

The original study’s statistics and my local case labels should not be merged into a single percentage. In the public primary S arm, 63 is the total of all collected primary responses; 58 is the count classified in one of the prose-rendering modes. The other five are included in the reported denominator: three summaries on one surface, a packet translation on another and an off-task reply on a third. One prose response was also marked truncated. A generator producing prose is therefore a mode observation, not a guarantee it finished or retained all scene requirements.

The original exact-tier rubric assigns 0, 0.5 or 1 to each specified requirement. Two source packets have five exact-tier items and one has six. The published 0.65-0.96 range is the mean item-fraction among prose-mode responses for each configuration, not source-token recall or a universal semantic score. The novel-four-gram means, 1.8-3.3%, instead measure overlap with source word sequences not already supplied in the packet. Low overlap supports substantially different wording but cannot tell who paced or whether the hand passed through the knee. The report already discloses its partially blinded single-grader procedure, unequal/opaque host conditions and the absence of a matched-length ordinary-summary baseline.

My local prose outputs were manually inspected for only a small selected event set. The denominators are intentionally narrow:

Measurement Correct denominator or scope Not equivalent to
Original prose-mode count 58/63 primary S replies 58 completely faithful scenes
Original exact-tier mean 0.65-0.96 among prose, by surface Fraction of all original scene meaning
Original new-four-gram mean 1.8-3.3%, by surface An event or actor-fidelity metric
Local translation-style check 6/6 field translations Successful narrative actor inference
Local G/S prose outputs 14/14 per model Full-rubric exact-tier successes
Source-grounded pacing audit 12 A/B/C outputs per model A representative model population
Synthetic D sensitivity cases 2 per model, excluded from source actor score Correct source reconstructions

In those 12 source-grounded pacing cases, Qwen gave Touma 4 / Mio 5 / omission 3 and Gemma Touma 7 / Mio 3 / omission 2. These three categories sum to 12 in each model. If one asks who was correct conditional on the beat actually being narrated, the denominator becomes 9 for Qwen and 10 for Gemma, respectively. But that conditional statistic can hide omissions, so I prefer to show correct actor, wrong actor and absent beat together. A scene that skips a challenging event should not appear more accurate merely because the hard actor question was never narrated.

For the physical relation, 0/14 Qwen literal and 8/14 Gemma literal describe a conservative human reading of one relation across all locally generated prose outputs, including the separate false D input. The source-grounded Gemma figure is 7/12. The three ambiguous-space cases are deliberately not counted as unquestionably literal, and ordinary touch is not silently counted as equivalent to passing through. I did not apply the original exact-tier rubric to the full local texts, recruit independent blind graders or construct a validated automatic semantic metric. A checksum and a successful GPU run can establish execution provenance, never the correctness of those human labels.

If anyone wants a minimal reusable rubric, I would annotate in this order: output mode, beat present/absent, actor only if present, relation type independently, then reference level and a short quote anchor. That produces interpretable labels even when a response is a packet translation, a truncated scene, or a generated story that is emotionally plausible but changes a crucial role. It can coexist with the original exact-tier scores, rather than competing with them.

Conditional next steps, with the cheapest useful check first

No large new experiment is necessary to discuss the reported feasibility signal. Different questions suggest different optional next steps, ordered here from least expensive:

  1. Manual-free reconstruction possible? Use the original S collection and its stated limits. No extra GPU run is needed; explicitly narrative prompts are a different condition.

  2. Where did a relation change? Default path: compare one saved ch05 output with source and packet; record pacing coverage, actor if present, and knee relation. Assign reference level before proposing an encoding or decoding cause.

  3. Does actor-cue placement matter? Fix model/instruction; compare A versus B, optionally C, with a couple of samples. Label omission separately. My table does not establish a universal syntax winner; larger sampling would be a later choice.

  4. Does unusual physics survive? First label existing knee passages. If needed, one fixed-setup contrast could clarify no ordinary touching at the event location. Not yet run; not a proposed mandatory field.

  5. Is mode selection the question? Keep model and packet fixed; contrast the original translation request with an explicit prose request. Score mode first; do not mix the two protocols.

  6. Comparative compression efficiency? Only if that stronger claim becomes important, add a length-matched plain summary or beat-list baseline, matching facts, budget, models, requests and event labels. The author already notes its absence; this is not immediate homework.

  7. Reproducibility across hosted surfaces? Preserve prompt, packet revision, surface label/date, output mode, text and denominators; mark hidden provider settings as unknown. Pinned local GGUF identities do not prove hosted backend identity.

These alternatives separate mechanisms while leaving the author free to choose the next question. Existing evidence first is sufficient for useful forum feedback.

Adjacent research and links for future readers

Three primary research references seem especially useful as connections, not direct matches or diagnoses of SCCN:

  • The WebNLG Challenge: Generating Text from RDF Data addresses realizing structured relations as text, including referring expressions and choices about how relations are verbalized. This provides a useful analogy for checking who acted, on whom, and how; a literary scene packet is not just an RDF triple set, however.
  • Data-to-text Generation with Entity Modeling explores dynamically tracking entities while producing prose. It is a neighboring design vocabulary for actor ownership and consistent roles, not evidence that its particular architecture is necessary for SCCN or would repair these exact outputs.
  • PARENT: Handling Divergent Reference Texts when Evaluating Table-to-Text Generation addresses the problem of references disagreeing with structured input, and develops an input-aware text metric. It is relevant to separating source truth from packet fidelity, but its lexical alignment is not a validated automatic judge for unusual fictional physics, withheld knowledge or the original study’s exact-tier items.

In practice I would keep the author’s own terms evt.seq, binding.exact, exact-tier and withheld information. Terms such as event coverage, actor/semantic-role binding and relation realization can help readers search the neighboring literature, but do not make SCCN identical to those tasks. Fiction introduces viewpoint, emotional causality, implicit social relationships and deliberately nonstandard world rules. A good ordinary-language paraphrase may preserve some and alter others; that is exactly why a few source-grounded labels are useful.

For verification of this specific case, the pinned public study tree, chapter-five source, and original packet are more important than any general analogy. I did not independently establish a specific current SCCN GitHub issue as the root cause of either observed relation change, and would not present unrelated issue reports as if they explained it.

Reproduction boundaries, actual execution and source rights

My local follow-ups were actual completed GPU runs, not simulated model predictions. The Qwen/T4 prose return passed ZIP/member integrity checks, completed 15/15 GPU-qualified worker processes including a smoke check, offloaded 37/37 layers, and yielded 14 scene texts. The Gemma/L4 return similarly passed its archive/member checks, completed 15/15 native GPU-qualified workers with 49/49 layers offloaded, and yielded 14 scene texts. Those checks exclude a simple failure-to-run explanation for these completed local outputs. They do not certify any semantic label, and the raw local return archives and manual review sheets are not presently public URLs for other forum readers.

The Gemma artifact came from the official Google GGUF model card, pinned in the local run to revision 29d097773436b69ff9feafd636ab4cf873786537; the main GGUF SHA-256 recorded there was 93567e57a8fe10b23569b9d9ec38cd005deedf71e29477c421a4b83f418a538b. I used a prebuilt CUDA llama-cpp-python wheel from the JamePeng release, rather than building it in Colab. The Qwen comparator was a local Qwen3-4B-Instruct-2507 quantized GGUF, not the hosted Qwen 3.8 27B surface in the author’s table. The precise model artifacts, templates and hardware were different, so the local contrast should never be attributed solely to parameter count.

The comparison used identical derived G/S text inputs and nominal sampling settings in the two prose runs, with two seeds per selected condition. I inspected the prose manually for the specific pacing and knee relations. This is neither an independent blind multi-rater experiment nor a reproduction of the complete published exact-tier rubric; it establishes only these case-level observations under these runs. Changing to a preferred sampler for either family, adding new models, or trying different genre packets would be a newly labeled experiment, not an already measured result. The earlier local wheel/output-directory setup incidents were resolved before these outputs and are unrelated to SCCN’s scene semantics.

Finally, the project has a RIGHTS.md statement; the research release shares bounded source spans and evidence, not the whole original fiction or private agent context. I have kept the illustrative source material to short event labels and factual scene descriptions, with links back to the author’s artifacts. A model card’s license does not license the underlying fictional work. If a public replication pack becomes available later, it should retain those distinctions and the exact prompts, hashes and human-label definitions.

Closing thought

I think the positive reconstruction signal from these early packets stands on its own terms. My small replay adds a concrete way to inspect where meaning survives, where an event disappears, and where a relationship is reconstructed differently. The least costly next step is already available in the published source, packet and saved outputs. More controlled variants are options if the author’s next question calls for them, not prerequisites for appreciating what the current study shows.

Thank you for taking this beyond commentary and trying concrete variants. The separation of beat coverage, actor ownership, and physical relation is particularly useful: an omitted pacing event should not quietly improve an actor-fidelity score.

A little project context: Unthink is a deliberately bounded release of earlier SCCN artifacts, not a snapshot of the current language or the whole research program. Subsequent work has extended into more explicit semantic representation, governed agent behavior, and authoring/evaluation tools. We are publishing in slices so each contribution can stand on its own evidence. That later work does not retroactively fix these packets or establish a result on your local tests.

The hand/knee observation interests me most. A rendering can keep the emotional beat while changing the fictional world’s physical facts. That is a different failure from filling in an actor the event did not locally specify. “The scene still feels right” is not enough if the mechanism making it that scene has changed.

Your G_C versus S_C cells also make a useful challenge to any simple “add an explicit binding and it is solved” account. In the cases you report, the presence of the binding and its successful realization are different things. That keeps both representation and rendering conditions in view without establishing a universal syntax winner.

One scoring question: for “through the space where her knee was,” would you retain an event-local label alongside a whole-scene reading? If surrounding sentences explicitly establish that the knee remained there and the hand encountered no resistance, the relation might be unambiguous in context even though that phrase alone is not. If they do not, I agree it should remain ambiguous. I would want to preserve genuine paraphrase without generously reading the source back into the output.

If you make a small evidence bundle available, the exact derived prompts and output passages behind those contrasting cases would be especially valuable. I have read your report, not independently inspected the local returns. A few checkable examples would let us discuss that distinction directly.

This is precisely the kind of engagement we hoped the release would make possible: someone can examine the artifacts, vary them, and bring back a more specific question about what survived.