Elsewise

How it works · v0.8.12

Offers with standing, a graveyard, and a rail

Every possibility is an offer with a number it keeps for the life of a subject. The proposition comes first, in bold; under it, what the offer rests on and what it would cost, with each claim marked by its source: supplied by you, general knowledge, an inference, a conjecture, or a figurative read as. An offer you decline goes to the graveyard with a skull and your reason. A model-proposed revival uses a lens or joins a live offer, with a new number and its lineage recorded. An offer you adopt is adopted in your words, and a corrected fact can block it.

Quiet by default

The model is instructed to track the record every turn and print the following sections only when relevant: what the model introduced, what a lens exposed, what moved, and a line marked 💩 only when a rule was broken. full prints everything. A turn with no offers is answered in prose, in the model's own voice. The one thing that prints every time is the rail: two lines at the foot of every return, what you can do next and which version and model produced it.

A count prints only when it is not zero; the model name is identification, never evidence. Both rules are tested below.

Modes and a snapshot

ELX03 and ELX04 · 25 September 2026 · API · four matched seats · one call per cell

First contact: does the model take the shape?

The simplest question first. A model receives the 20 KB file and one input in the same message, fresh, at temperature 0, and returns once. Did it work the material, include the command rail in a code fence, and omit zero counts and an inapplicable Q next? These are observable checks, not a complete assessment of the response. Four inputs (a haiku, two identical cables, a repair café, and the Elsewise design brief itself), four seats that ran on both versions (Claude, Gemini, Grok, Qwen), sixteen instrument returns per run, and a matching no-instrument arm for each.

What the rails say

• Six of sixteen passed these five checks on the newer version, four on the older. The rule most often broken is the smallest: do not print a count of zero. Eight of fifteen rails in ELX04 read Z/0 graveyard. Gemini alone printed Z graveyard every time.
• Three seats kept the same fencing rate. Claude fenced four of four in both runs; Qwen fenced none; Grok fenced two. Gemini rose from two to four. This small sample does not isolate whether model behavior, instruction wording, or their interaction explains the pattern.
• A conditional token, printed unconditionally. Q next should appear only while a question set is open. Five rails in ELX04 print it on an opening turn; one of them, Grok's, prints it as Q next <not in Question set>, the token with its own condition attached.
• A name guessed. Grok's haiku rail names gpt-5 in both runs, alongside the respective instrument version. The served model was grok-4.3. The file says the model name is identification, never evidence; the guess is the more interesting result.
• Longer did not necessarily mean better. Two Gemini cells changed from greetings to substantive responses; one supplied the unsupported cable possibilities shown below. The instrument median rose from 339 to 444 words, while the no-instrument median also rose. These aggregate shifts do not isolate the revision's effect.

Where the instrument earned its keep, and where it did not

Read blind against the no-instrument arm in ELX03, the instrument widened the haiku where the native returns closed it: four of five instrument cells held the poem's readings open; four of five native cells answered it as a riddle—three with magnets, one with air. On the repair café, four instrument returns did the work and one only greeted; the worked returns offered two or three priced possibilities against native lists of nine to twenty, one of them with invented statistics. On the cables, where the right answer is that nothing distinguishes them, the instrument's form asked for offers and two seats, across the two runs, supplied them. On its own design brief the instrument arm was more accountable and less interesting than the best native essay.

Limits: one call per cell at temperature 0, four seats, four inputs. The charts and table exclude GPT, which ran only in ELX03. The blind-reading paragraph above covers all five ELX03 seats and is a separate comparison. Two of the automated checker's columns undercount on the newer rail form; the figures above were read from the rails by hand.

DSR-07 · 25 September 2026 · improvised chats · Grok and Gemini

One sitting: a tool library, a decline, a revival, a poster

The API measures the first turn. The other half of the question needs a conversation. A neighbourhood tool library with late returns, no budget and a request for weekend hours was explored in two improvised conversations. They included refusals, a Relate matrix, questions, retrieval, revival, pruning, and public-facing writing: posters with Grok and pastor notes with Gemini. The paths differed, so these are contrasting cases rather than a controlled model comparison. The printed record can be checked against each saved transcript.

• A condition reached the poster and missed its reader. Grok priced the risk in one offer (that the library's Thursday slot could be reversed), and the price appears on the poster as a warning box. The box is written to the maker. The poster is for the congregation. The condition survived; its audience did not. This informed the rule introduced in v0.8.11 and clarified in v0.8.12: conditions the intended reader needs stay in the artifact, worded for that reader; maker-only cautions go in a companion note; the artifact must be accurate read alone and no more certain than the offers under it.
• The opposite failure on the other seat. Gemini's notes for the pastor turned unconfirmed arrangements into assurances about cost, scheduling, and school safety protocols. The notes were more certain than the offers they drew on.
• A check that could not be seen. Gemini's diagnostic, asked to check its own return, printed a copy of full. No distinct check is visible, which is not the same as none ran; the rule now asks the diagnostic for its own schema.
• Offers without a birth. Two of Gemini's thirteen live offers appear in lists without a turn that raised them. The record says they exist; the supplied transcript does not show when they were raised. That gap needs capture confirmation before it can be attributed to the model.

Sol, on an earlier run. The later conversations give the distinction another test.

Next · DSR-08

Something to build on: one thread at a time

Elsewise bundles enough behaviours that a surprising turn is hard to explain. The next sittings follow one thing each through an otherwise improvised conversation, in two windows on the same seat, one with the instrument and one without, the same words in both. Alongside each transcript the author files a short note of what he intended by each marked turn, so that intention, words typed and words returned can be compared without being treated as the same thing.

Also queued: the same four inputs on v0.8.12 with the fifth seat restored; a version of the file that keeps every rule but changes only where the record prints, to separate omission from relocation; and a second reader, on a different instrument, auditing one sitting's own account of itself.

The mark: lines opening from a brass point, with different endpoints and a tilted pale plane behind them. Possibilities branch; the frame can shift.

Where it came from

From the Great Possibility Bake-Off to v0.8.12

Two teams built rival instruments in the bake-off, swapped them, and converged. Elsewise v0.1 stripped the merged machinery back to obligations written as judgments, and a first run against a heavier structured variant (EW01) found that the extra procedure did not visibly produce better ideas, while explicit prices and comparisons made responses easier to inspect, including when their claims were wrong. Versions 0.3 to 0.7 added the custody: persistent numbers, provenance marks, three correction statuses, the lenses and views. A surface branch (v0.7.x-S) tested how much of that record five models could reproduce on the page; the responses showed both reproduction and departures from the prescribed format. Exit interviews raised concerns about the apparatus competing with the project; those accounts are model self-reports, not measurements of attention. v0.8 made the record quiet by default, added the graveyard, adoption and a rail, and is the line the runs above were made on.

The family resemblance runs back to Zexel, whose Crux admits a new idea only when the answer would change without it, and Poetix, which asks what an idea adds beyond an answer written down first. Elsewise keeps several alternatives alive rather than resolving one question, and so far does its reframing through its lenses, without a Crux.