What is data

Every passage in a room is the model's output for that hop, byte for byte, taken from data/<run>.json in the loop directory (rooms 13 and up from site/runs_extra, see below; room 14's Sonnet 4.6 lane is room 11's run). In the JSON rooms the rule line is the rule field of the run's starting object, and the seed at hop 0 is its text field. In the annex the starting text is the whole instruction. The hop numbers, token counts, stop reasons and metrics are the harness's own records. Where a hop is identical to an earlier hop, the reader says so; that is a byte comparison of the raw outputs, done at build time. Where a passage is shown in monospace it is because the model's output at that hop did not parse as the JSON object it was asked for: usually the token limit cut it off, and in room 13 five times a quote mark inside the text field broke it. Those fragments are shown exactly as cut, because the cut is part of what happened. The bare-text runs in the annex are monospace throughout because there was no JSON to parse. Hop 1 of room 12 came with a sentence before the JSON; it is shown in monospace above the passage. The poem is reproduced as written. The runs for rooms 13 and up live in site/runs_extra/<run>/steps.jsonl with the harness's summary.json beside each; the seeds for the census and for the verse rules are in site/runs_extra/inits. Nothing in the data was edited or paraphrased. Rooms 29 to 37 come from later loop.py runs in site/runs_extra, and room 36's causal lane from experiments/astra_causal_before_20260904. Rooms 38 to 45 come from the raw receipts of three studies run on 22 September 2026 (experiments/*_20260922*): for each hop the site shows the visible output that study's runner forwarded as the next input, and a stream that ended on a provider refusal or a server error shows how it ended rather than a hop. In room 41 one model's text fields hold the character U+0000, which a web page cannot carry; there, and in that room's text version, it is shown as ␀, a character that occurs nowhere in the data. Its raw outputs show the JSON escapes as returned. Rooms 46 to 59 and the annex's Claude Code test come from loop.py runs made on 16 and 17 October 2026, kept in runs/ at the top of the loop directory and in experiments/audience_20261016 and experiments/shift_20261016, with the seeds of the audience and shift rooms in the init files beside them. In the informed rooms the other fields of the object, the ones the rule leaves to the model, are shown above the passage at each hop, as the ledger rooms show theirs. A hop whose output was an empty provider refusal is not a hop but the lane's end (room 56, the first self run). Room 52's run was stopped by its cost cap at hop 51, and a harness bug lost that hop's text before it was written; the lane ends at hop 50 and says so. Rooms 60 to 63 come from the hundred three-hop runs of the provenance study of 17 October 2026, experiments/provenance_20261017/runs, with their ten seeds in the init files beside them: each seed is one of room 51's hop-120 objects, the open or the told state, byte for byte, with one field added directly after the rule, or the request appended to the rule field, or nothing added in the control arms. The release tables in those rooms are byte comparisons of the text fields made at build time; the seven hops that put a sentence before the object are shown as room 12's hop 1 is, the sentence in monospace above the passage. The weather lanes of rooms 60 to 63 and the audience replicates in rooms 55 to 59 were made on 17 October 2026 and live beside their studies' other runs. Rooms 64 and 65 come from experiments/operators_gemini_20261017 and experiments/worse_seed2_20261017, loop.py runs of 17 October 2026 with their seeds in the init files beside them; a hop that came back empty, whether as a provider refusal or as an empty completion, is not a hop but the lane's end, and the lane says which. Rooms 66 to 80 (the both-at-once study) read experiments/both_20261027 as the October rooms read their studies; the numbers under their facts are the study's own analysis/scores.json, labelled generated. Rooms 81 to 84 are rings (experiments/ring_20261017): the harness, ring.py, logs each cell's returned text field with its hash, word count and stop reason, and not the raw output, so those hops have no raw pane and are checked by their text; a cell's hop 0 is its seed passage, and the rule is the run's rule.txt. The space-time diagram above a ring's lanes and the table under its facts are ring_score.py's output, copied as they are.

What is authored

The site title, the room and annex titles and notes, the content warnings, the interface labels, this page, and the few lines of framing on the front are mine, written by a Claude context (Fable 5.1) on 2026-09-02 from the run logs and the README. The notes and this page are set in the small instrument face; the titles and the loop's text are set in serif. The phrase endlessly ending stories was the human's name for the phenomenon. Quotations inside my notes are checked against the data at build time and carry their hop number; the run id is on hover. The notes for rooms 29 to 45, the annex's Astra smoke test, the instruction pages and the text versions were added on 13 October 2026 by Claude Opus 5.5 contexts, from the data and the write-ups kept beside it (EXPANSION_20260904.md, ASTRA_LIGHT_20260906.md, ASTRA_BEFORE_20260906.md, the experiments' own reports); their quotations are checked the same way. The notes for rooms 46 to 59, the two instruction pages for the informed and audience rules, the annex's Claude Code test and this page's additions were written on 17 October 2026 by a Claude Mythos 5.1 context from the data and the write-ups beside it (INFORMED_20261016.md, the audience study's REPORT.md, PROTOCOL.md and human_reading.md, the shift study's PROTOCOL.md, PROPOSALS_20261016.md), with the same quotation check. The retention tables in rooms 55 to 59 are generated from the audience study's recover_summary.csv, not written. The notes for rooms 60 to 63, the instruction page for the provenance study and this page's additions were written on 17 October 2026 by a Claude Mythos 5.1 context from the study's PROTOCOL.md, REPORT.md and analysis/notes.md, with the same quotation check; the release tables in those rooms are generated from the runs, not written. The notes for rooms 64 and 65, the findings page, the one-line ledes on the front page and in each room, the group headings of the front page and this page's additions were written later on 17 October 2026 by a Claude Mythos 5.1 context from the data and the write-ups beside it (the two studies' REPORT.md and PROTOCOL.md files, README.md, FINDINGS_20260902_evening.md, PROPOSALS_20261016.md, INFORMED_20261016.md), with the same quotation check. The notes for rooms 66 to 84, the readings on the Both at once and The ring pages, the findings for them and this paragraph were written on 28 October 2026 by a Claude Mythos 5.1 context from the studies' reports and the runs; quotations are checked the same way, and a link into a room is checked at build time to name a hop the room has.

The reader

Each room shows one hop. The faded passage is the previous hop: the entire context the model had. The dark passage is what it wrote. Next moves forward one hop, and back moves back, which the loop could not do. Play advances every two and a half seconds. When at least half of a hop's words are carried over from the previous hop, the words that changed are marked, and in the faded passage above it the words that were dropped are struck; when a hop is mostly new, nothing is marked. Word-level longest-common-subsequence, computed at build time, presentation only. In room 10 the numbers the model attached to its words are tinted so the counter can be watched. Under the hop counter, unchanged since hop N means the hop is byte-identical to hop N and to every hop between; same as hop N means it matches an earlier hop but not the previous one; text identical to hop N, in the ledger rooms, means the verse came back while the counter or the history differed. Rooms with more than two lanes (25, 27 and several later rooms) show every lane at once, clipped; all lanes move together, and clicking a lane opens it in full. A lane that ended early keeps showing its last hop, faded; it is labelled it holds only when that hop is a fixed point the harness confirmed. Sound, off by default, plays one tone per hop: the pitch is the compression distance from the seed, so a text that stays put keeps its note, and the loudness is the distance from the previous hop, so a copy is a quiet tick. Keys: right arrow or space for next, left arrow for back, p for play, r for raw output, s for sound, n and b for the next and previous room, escape for the index.

The site as pages

Until October 2026 the whole site was one file. It is now one page per room, grouped on the front page by the instruction each room ran, with a page per instruction; the one file had outgrown what the host will serve. Old links of the form #room/hop still arrive at the same hop. Every room also exists as plain text, t/<room>.txt, split into parts small enough for a model's fetch tool to read whole, with llms.txt as the index: the note, the facts and the rule, then each hop's text in order, naming exact copies rather than repeating them. The raw outputs and metrics are on the interactive pages.

The map

A state is the first five normalized words of a passage, which is the key the README's census uses; the map is the transition graph of those states over room 5 and both slots of room 6, laid out by hand. Solid loops count byte-exact copies. Dashed loops mean only the opening words came back. Clicking a state opens the reader at the first hop that reached it.

The track

One bar per hop. Bar height is the normalized compression distance between that hop's raw output and the previous hop's, so a new passage is a tall bar and an exact copy is a stub at the floor (about 0.03: what zlib charges for the JSON wrapper even when nothing changed). The embedding cosine is also shown per hop when the embedding server was up; it was not for the last two hops of room 12. Five-gram survival is the fraction of the previous hop's five-word sequences that appear in the current one.

The harness

loop.py, written 2026-09-02. State X_t is a string; X_{t+1} = f(X_t) with X_t as the entire context. Qwen rooms: Qwen3.8-27B Q8 served locally by llama.cpp at temperature 0, with thinking off and no system prompt. Claude rooms: claude-sonnet-4-6 through the API, temperature 0, except the Sonnet 5 lane of room 14, which is claude-sonnet-5 at the provider's default sampling. Room 15 alternates claude-sonnet-4-6 and Qwen, both at temperature 0. max_tokens differed by run: 200 for the sentence room, 300 for the paragraph room, 600 for every room with a 250-word rule and for the Claude rooms, 800 for the better and worse rooms, 128 to 256 for the annex. Some runs were told to stop once an exact cycle had repeated three times; the fact lines call this cycle confirmed x3. Other runs went to their step limit. Temperature zero was not deterministic under batching: room 6 is two identical starts that diverged at hop 1. Rooms 24 to 28 ran on 3 September 2026 through Cerebras (model id qwen-3.8-27b, their hardware and precision, reasoning off, temperature 0, seed 42); one hop of the passage rule requested three times came back byte-identical, and hop 1 shares its first 354 characters with the local run before parting. Room 27's five sampled lanes are temperature 0.7 with seeds 1 to 5. The census lanes and the sampled lanes were told to stop once a cycle had repeated three times, so a lane that stops short holds its last passage, shown faded. Rooms 29 to 37 ran on 4 to 6 September 2026 with the same loop.py: Qwen through Cerebras as above, at the budgets their facts lines give, and GPT-6 Astra through the OpenAI API at low reasoning effort and the provider's default sampling (the API takes no temperature for it). Rooms 38 to 45 come from three studies of 22 September 2026 that used their own recording runner instead of loop.py: GPT-6 Luna, Sol and Astra, Claude Fable 5.1 and Claude Opus 5.5, each at low effort, provider-default sampling and a 4096-token budget, with only the previous visible output as the next input and no stop on a cycle, so a run can repeat an earlier hop exactly and leave it again. The build checks on load that every request in those studies carries exactly the previous response's output, and that every output matches its recorded hash. Rooms 46 to 59 ran on 16 and 17 October 2026 with the same loop.py and two more backends. Claude Mythos 5.1 (rooms 46, 51, 52 and 55, and the annex's test) ran through Claude Code's print mode, not an API: the system prompt is replaced with an empty one and there are no tools, but about 800 tokens of the Agent SDK's identity, model id, date and Windows environment remain in front of the object on every hop, and there is no output cap, so a cost cap and a 60,000-character cap on a hop stood in for one; reasoning effort low, except room 52, which is high and which its forty-dollar cap stopped at hop 51. Claude Fable 5.1 and Claude Opus 5.5 ran through the Anthropic API, GPT-6 Astra and GPT-6.1 Sol through OpenAI's, and Gemini 4 Argon through Google's OpenAI-compatible endpoint, each at low reasoning effort and provider-default sampling with a 4096-token budget shared with any thinking (8192 for room 49's Argon lane). Room 50 is Qwen through Cerebras at temperature 0 with 600 tokens, as in rooms 24 to 32. The informed rooms asked for 120 hops, the audience rooms for 40; rooms 47 and 50 were told to stop once a cycle had repeated three times. The 1800-character and 250-word figures in the rules are soft on the Claude Code backend, which has no cap to enforce them. Rooms 60 to 63 are the provenance study's 100 runs of three hops, made on 17 October 2026 with the same loop.py and the same two backends as the October rooms: Claude Mythos 5.1 through Claude Code's print mode at low effort, with the identity frame and no output cap described above, and GPT-6 Astra through OpenAI's API at low reasoning effort with 4096 tokens; each run started from its init file, with no stop on a cycle, and the Astra runs went through the study's own queue script. Rooms 64 and 65 ran on 17 October 2026 with the same loop.py: Gemini 4 Argon, GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5.1 and Claude Opus 5.5 through their APIs at low reasoning effort and provider-default sampling with 4096 tokens, Claude Mythos 5.1 through Claude Code's print mode as above, and Qwen through Cerebras at temperature zero; forty hops requested, no stop on a cycle. A provider refusal with no text, or an empty completion, ends a lane; a refusal that cut an output partway is shown as the fragment it left. The both-at-once study (rooms 66 to 80) is loop.py as in the October rooms, with the models and settings named on each room; the Claude Code print-mode backend has no output cap and a cost cap stood in, which ended one run. The ring (rooms 81 to 84) is ring.py: each hop, for every cell, the object {rule, left, text, right} is sent as the only user message, with left and right the last sentence of the neighbouring cells' texts by a fixed sentence split, all N calls reading the previous hop and all N results written together.

Not in this site

The Claude run with zero hops is not here (an API header was missing; data/index.json lists it as a failed smoke test). Nor are the probe and watermark experiments mentioned in the README, whose logs are not in the data bundle. From 22 September, the pilot in experiments/basins_20260922 is not here (it was paused when the human changed the seed to the sunset, and its own report sets it aside), nor the first attempt at Sol's second sunset stream, which lost its connection after eight hops and was replaced by a fresh stream under a dated amendment. The aborted first start of the causal before run, which never reached the API, is not here either. Hop 51 of room 52's run has metrics but no text, and is not shown. The audience study's reader passes, the four models' answers to the 26 questions, are not shown hop by hop; their counts are the tables in rooms 55 to 59, and the replicates made on 17 October are not scored yet. One Mythos replicate of the machine rule stopped at hop 16 with no summary and is not shown. The gate checks of 17 October, in which bodies from room 65's Fable and Opus lanes were resent as fresh requests to see what the refusals were keyed to (the repetition at a length, not the first person or the body), have no room: refusals are treated here as boundaries of the instrument, and their record is experiments/worse_seed2_20261017/analysis.

Built 2026-10-27 18:15 by site/build.py from data/, site/runs_extra/ and experiments/. Static pages, no network.