The circle
returns open.

Over-Color / Carrier V3 brought colour names, typed phrases and exact objects into one reversible register. This is an earlier format, distinct from the later USL1 message layer.

Recorded finding. The original page explicitly described its longer phrases, checksum limits and the separation between encoding an object and establishing a claim about it.

Boundary. A spoken name and a checksum do not supply truth, authentication or token compression.

A lexical substrate, permanent colour allocation and larger exact-image messages made the layers usable together.

Recorded finding. English Base 1 retained 354,983 source forms. The colour-allocation migration reassigned three duplicate Carrier colours while retaining historical context. A supplied full Bible TXT recovered byte for byte through the saved CVP1 image.

Boundary. Lexical forms are not word senses; exact complete-text transport is distinct from a later selective conceptual atlas.

USL-E01 combined existing Carrier concepts with an explicit experimental grammar, then transported the expression through a CVP1 image. Its contribution was a new composition with roles, conditions, permission, prohibition and aims.

Recorded finding. Two reported receiving systems recovered the same 100,551-byte USL1 packet and the same eleven concept bindings. The pinned context contained 106 checked records. Successful recovered structure coexisted with differences in prose and gaps in one returned validator.

Boundary. One intact artifact; code-assisted exact pixels; English definitions and grammar; differing prior-context conditions. Recovery does not establish semantic compression, unaided visual reading, robustness to transformed images, or superiority to equivalent prose/JSON. Public reports distinguish recovery from validator conformance.

CVP2 RGB24 Exact increased gross payload capacity per data cell from four to 24 bits while preserving USL1 and the frozen E01 composition. Separate exact-text examples carried a retained Bible and the project corpus.

Recorded finding. At matched 4×4-pixel cells, E01 changed from a 2048×2048 CVP1 PNG of 172,392 bytes to an 896×896 CVP2 PNG of 125,718 bytes: 27.1% smaller, with both receivers returning the same 100,551-byte packet. The separate Bible example recovered all 4,455,950 TXT bytes from a 1,750,944-byte PNG, 60.7% below uncompressed TXT.

Boundary. Sixfold gross cell capacity is not sixfold file-size reduction. Book examples use a shared lexical database excluded from PNG size and conventional compression; no advantage over directly gzipped TXT was established. Original digital files, not camera/screenshot recovery; fresh-context conditions were organizer-reported.

Five source atlases moved from recoverable references to authored readings with evidence, qualifications and competing priorities: Bible, Qur’an, Tanakh, Memorandum 11 and Memorandum 7. Later kits supplied full sources so receivers could challenge selections beyond the carried quotations.

Recorded finding. The release 106 retrospective counts five atlases within seven distinct sender compositions. The checked corpus grew from 106 to 1,150 through 765 scriptural-batch and 279 memorandum-batch additions. The atlases selected 182, 247, 220, 207 and 163 distinct checked concepts respectively. Public returns document exact recovery and source-based challenges: competing readings, inherited religious standing, divine conduct, political freedom, and later legislative reversal.

Boundary. These were small contextual, tool-assisted receiving trials and qualitative assessments, not controlled evidence of better reasoning. Checked definitions, new applications and whole interpretations have different status. Navigational coverage is not exhaustive interpretation. In Memorandum 11, the paragraph supporting the political-freedom challenge was already transmitted in the separate source-questions group; its new role was argumentative, not newly discovered uncarried evidence.

A Peace That Leaves the Future Open completed a cycle of expression, recovery, interpretation, disagreement and encoded reply. The response proposed replacing one clause and explicitly preserved the others.

Recorded finding. The letter used 32 existing concepts in 54 concept-token occurrences and 25 clauses. All selected IDs were 36–104, predating the large atlas enrichments. The reply used only two concepts, targeted p05, left 24 clauses unchanged, and occupied a 56×56 CVP2 image of 2,888 bytes carrying a 2,487-byte USL1 packet. No new dictionary entry was needed.

Boundary. The compact reply reused the frozen context and lexical dependencies; it did not carry its own full dictionary. The original expanded object was 1,825,183 bytes, of which 1,802,512 were canonical carried context. This was one contextual supplied-tool-assisted trial. Encoding a counterproposal neither accepts it nor changes the source or authorizes action.

Expression Library broadened source coverage beyond scripture and policy. Source ledgers, overlap review, reuse applications and more explicit reading coverage supported enrichment; Reading Room 2.0 made a growing dictionary browsable, shareable and portable.

Recorded finding. Ten books added 327 definitions, reaching 1,477; fifty more added 693, reaching 2,170; rereading ten represented books added 31, reaching 2,201; ten new sources added 241, reaching 2,442. The last collection recorded 10,071 indexed text blocks inspected in 236 packets and 222 proposed source applications of 171 earlier definitions. The Reading Room added per-concept addresses, local reading marks, graph navigation and a timestamped AI handoff.

Boundary. Earlier batches used selected close reading; later complete indexed-text coverage does not prove exhaustive understanding, and visual formula/diagram coverage remained partial. All used English-readable, nonrandom sources; authoring and peer review shared an AI workflow before human Keeper promotion. No new external receiving experiment or validated yield forecast resulted. Increased entry count and graph degree do not measure understanding.

Organic revision inspected existing definitions and qualified their relationships. Corpus enrichment and three Wikipedia waves then expanded reusable distinctions with source pins, scope notes, contextual links, human review and authored use examples.

Recorded finding. Organic revision reviewed all 2,443 baseline definitions, approved 389 typed relationships and 12 additions, reaching 2,455. A fixed 12,000-request navigation simulation made 114 requests newly reachable and 1,452 shorter. The project corpus added 113 definitions to 2,568; Wikipedia waves added 1,521 to 4,089; 325 to 4,414, then 1,843 to 6,257. Wave III's 32 authored before/after pairs yielded 64 locally recovered messages with 228 verified concept-token occurrences.

Boundary. Shorter graph routes are a structural navigation result, not improved understanding or retrieval quality. Wikipedia sources were selected passages, not exhaustive full-article or primary-literature review. Some example messages add detail and are not equivalent paraphrases. Local codec success supplies no independent reader-benefit result; checked status is editorial admission, not expert consensus or automatic approval of future applications.

A frozen 6,257-record corpus and controlled lexical/graph comparisons tested whether contextual links improve recovery of intended concepts.

Recorded finding. On 240 held-out AI-authored questions, BM25 MRR@10 was 0.914778 versus 0.915184 for the selected graph blend: +0.000405, 95% family-bootstrap interval [0, 0.001157], sign-flip p=0.49435. Both had 214 rank-one and 232 top-ten intended-target hits. Only two of 80 families improved; 78 tied. An exploratory dense model scored 0.814563; the 16-question reader pilot reached the same ceiling under both contexts.

Boundary. No reliable graph retrieval benefit was established. Authors saw target names/definitions; the benchmark measured intended-item retrieval with other concepts unjudged. Split blinding was imperfect and corrected in the record. One dense model and a same-family reader pilot do not represent all embeddings or independent human understanding.

R02 began with information needs and fixed answer obligations, then used pooled relevance judgments and a separate answer endpoint. A qualified typed overlay preserved predicate direction and scope.

Recorded finding. 48 tasks comprised 24 adaptations of public human questions and 24 AI scenarios; 12 were development and 36 final. Two AI assessors graded 666 candidate-task pairs. Dense final nDCG@5 was 0.674080 versus 0.670678 typed: −0.003402, 95% interval [−0.010205, 0], raw p=1.0000. Answer utility was 0.908565 versus 0.922454: +0.013889, 95% interval [−0.003472, 0.031279], Holm-adjusted p=0.2822. Only two final top-five lists changed; 34 were identical.

Boundary. Neither co-primary endpoint established an advantage. A true or well-supported relationship can still promote a general neighbor and displace needed evidence. The 146-relation overlay touched 3.77% of the corpus; pools were incomplete; judges/readers shared a model family and no human validation established ground truth. Final fusion outperformed the selected dense baseline descriptively, but cannot replace it after seeing final results.

R03 tested a fixed requirement-aware selection method on public R02 tasks, a separate bounded-memory handoff, and deterministic reference/cost/scale checks.

Recorded finding. Across 48 reused tasks, coverage rose 73.96%→76.56% while nDCG@5 fell 0.692666→0.604827. On the original final 36 subset, coverage fell 73.61%→72.92% and nDCG fell 0.674080→0.593081. In the separate handoff, 128 actual agent outputs across 64 trajectories matched all 768 scored fields under all four formats. For 32 records across 100 handoffs, inline evidence used 1,088,900 application bytes; warm-cache transfer starting empty used 383,500, but full registry provisioning raised the charged total to 3,338,602.

Boundary. Retrieval reused public tasks and incomplete AI-judged pools; no fresh holdout. Perfect handoff scores were a ceiling on six fields across eight simple synthetic cases, with supplementary definitions and no human receivers/graders. Generic references received matched mechanics, so no Carrier-specific advantage followed. Payload bytes are not model tokens. One-million-key lookup reused 6,257 records with 6,256 distinct definition texts; more keys did not create more knowledge.

The local prototype gives Carrier a concrete role as a shared glossary beside inspectable project memory: statement, evidence, scope, conditions, unknowns, authority and proposed next steps. JSON and AI-brief exports support a reviewable reply.

Recorded finding. The tool uses a frozen 6,257-concept catalog and an R03 publication example. It can show exact field changes, retain checkpoints, compare returned JSON without accepting it, and export selected definitions with source text. Release 1.1 carries the snapshot caveat into text exports and preserves earlier starter records.

Boundary. A working prototype, not a new benchmark or USL1/CVP2 profile. Structural checks do not establish truth, entailment, completeness, authorization or authenticity; a matching source excerpt need not support a claim. No automatic saving or live-registry freshness. Its public account explicitly reports that the retrieval studies found no established advantage.

A concept-first 7×9 composition was frozen before its author's English reading. A fresh pretrained agent received the composite image, layout, frozen dictionary and supplied decoder, recovered the score, froze a first interpretation, then authored a reply. The author's reading was withheld. No hidden English original was being reconstructed.

Recorded finding. One original composition and one reply were recovered with 29 supplied-codec checks. The reply reused 33 selected concepts and introduced 17 outside the original selection.

Boundary. One artistic exchange; no correct-reading key, medium comparison or AI-performance result. The 29 checks use supplied codecs and original digital files with supplied layout; no independent codec, optical/print/scan recovery or verification of every decorative pixel.

Nine fixed fresh-agent reader attempts, three each for Carrier composite PNG, ordinary structured JSON and controlled English prose, arranged in three launch blocks. All share a 6,257-entry dictionary, grammar and optional common inspection adapter. Interpretation and reply freeze precedes sixteen assisted factual probes. Two fresh agent appraisers assess seven qualitative dimensions without a composite winner.

Recorded finding. All nine readers scored 16/16 on fixed probes. Common normalization and one composition prevent a notation-superiority conclusion.

Boundary. Every reader normalized through the common adapter; later performance cannot isolate the original notation's effect. No Carrier superiority or general equivalence follows. One composition, nine attempts, shared dictionary/model conditions; 144 probe items are not 144 independent readers.

A separate exploratory round added after the notation result. Two themes each had one creator assigned to Carrier, free JSON or expressive prose: six fixed creator–viewer pairs. Free JSON/prose were read in native form without normalization. Viewers froze interpretations, chose their reply medium before writing, and could transform or decline. Creator intention was withheld and was not the correctness key. Two judges assessed six artistic dimensions with anchored 0–4/NA judgments.

Recorded finding. Four of six planned creator attempts completed; two stopped at capacity. All four completed viewers chose prose for their reply.

Boundary. Four completed pairs from six fixed attempts; missing work remains NA and was not replaced. Creators, works, viewers, media, resources and affordance limits vary together. Two planned cases per medium cannot establish a causal or population ranking.

Three fresh pretrained creators first composed visual preludes using a 256-colour opaque subset of the 6,257-entry frozen palette and an English procedural grammar, before dictionary meanings were supplied. Each then encountered decoded meanings and created two semantic developments. A fixed creator ring supplied a fourth answering state. Eight deterministic spatial/topology-based steps per prelude formed a separate rule branch. Two fresh interpreters per colony froze readings, alternatives and field replies.

Recorded finding. Three authored trajectories produced 12 authored states; 24 deterministic rule states formed separate branches. Six readers supplied interpretations and field replies.

Boundary. Common pretrained model, grammar and dictionary conditions; no language-free internal thought, consciousness, biological life or Carrier advantage is established. The opaque initial palette delays supplied definitions but cannot prove absence of prior colour/concept knowledge. English instructions and grammar remain scaffolding.

One randomly initialized 625,745-parameter model (32-dimensional embedding, one 64-state GRU, 6,257-way head) learned next-colour identities from numeric walks on the inherited graph. Unordered endpoint pairs, including reverse edges, were split before fitting. One 1,200-update run selected checkpoint 200 by development loss; all 2,038 held-out edges were then scored against the full 6,257 candidates. Three fixed prefixes generated nine learned fields plus three untrained first-field controls. Six separate pretrained readers interpreted the final learned fields and replied; the frozen model continued from non-null reply colours only.

Recorded finding. A 625,745-parameter GRU trained from random initialization achieved modest held-out loss improvement. Simple graph-neighbour and frequency baselines ranked better; the model did not receive the whole spatial field.

Boundary. Selected-model NLL improves modestly, but training-frequency and two-hop graph baselines rank targets better. One graph, one initialization and dependent queries do not support population inference or semantic-understanding claims. The one-hop graph baseline is structurally disadvantaged because direct test edges are held out. The two-hop comparison is more informative.

Can a small numerical model respond through the complete Carrier field rather than only a sequence of colour labels?

Recorded finding. Full typed-link F1 was 0.547 on 600 held-out scenes, versus copying at 0.906. Some channels improved while changed colour identities and graph preservation remained weak.

Boundary. Receiving all channels did not ensure preserving them. Deterministic validity repairs and a copy prior were part of the system.

Can a source-preserving specialist repair the graph channel while the earlier field backbone remains frozen?

Recorded finding. On its new 1,200-scene test, typed-link F1 reached 0.991; copying scored 0.902 and the prior model 0.553 on those same inputs. Newborn-link F1 remained about 0.526.

Boundary. A relation specialist improved the specified task. New connections and near-threshold cases still required work.

Can the instrument improve parent selection, growth feasibility and threshold decisions without hiding failures in an aggregate link score?

Recorded finding. Newborn-link F1 reached about 0.942 on its ordinary test, with nine extra and sixteen missed links. Four-turn colour agreement was 61.7% on its 60 trajectory starts.

Boundary. A strong relationship score could coexist with an evolving field whose colour identities diverged from the reference.

Can the whole numerical field, including retained and legitimately changed colour identities, follow a known reference without accumulated drift?

Recorded finding. A teacher-informed fitted program matched all 2,000 ordinary and 480 focused single-step fields, and every step in 120 eight-turn trajectories.

Boundary. It emulates familiar authored rules with substantial engineered structure. Exactness on these instances is neither a universal proof nor creator approval.

Which replies are permitted or valued when more than one continuation is possible, and can declared conditions survive repeated updates?

Recorded finding. Across 30 first candidates, 26 met explicit charters; seven compliant replies were nevertheless declined by creators. Each source admitted multiple accepted alternatives.

Boundary. Six purposely composed works, unequal inputs between numerical and AI responders, related model sessions and no human reception study.

An actual alternating conversation becomes an inspectable animated artwork, with each proposal and counterpart decision retained.

Recorded finding. 24 proposals; 22 adopted, including one accepted unchanged pause; two declined. Six fixed protections survived. The final field retains an unresolved interpretive disagreement.

Boundary. One AI artistic trajectory with companion prose; no medium comparison, human reception measure or inner-experience claim.