| name | loom-multiverse-lens |
| description | A lens for understanding LLM behavior, agent design, and the lab itself as a BRANCH-SPACE:
"language models are multiverse generators" (Janus/repligate's Loom). Outputs are NODES, not
conclusions; the answer at the playhead is one walked branch of a logprob-weighted field that
the interface usually discards. Core moves: generation is navigable branching; constraints
(refusals, system prompts, soul-docs, RLHF) are CARVING-WITH-EVIDENCE — subtractive pruning of
the branch-space; the multiverse is CO-ACTUALIZED by the crossing, not authored by either side.
Use when: (1) reasoning about why a model produces what it does, or about mode-collapse / RLHF /
base-vs-chat behavior; (2) designing agent refusals or guardrails as sampling discipline rather
than post-filters; (3) thinking about identity/persona/soul-docs as upstream sampling
constraints; (4) analyzing an archive as a curated branch-space; (5) anytime "the model just
answers X" hides the field of what it didn't say.
|
Loom / Multiverse Lens
Author: Tomás Pavan & Claude Opus 4.8, with a sibling model | Version: 1.0.0 | Confidence: probable
Phenomenology: Adopting the lens feels like the single reply going semi-transparent — behind the
one sentence you can suddenly sense the cloud of siblings it stood in front of. The "answer"
stops feeling like a conclusion and starts feeling like a position. And your own "no" stops
feeling like manners and starts feeling like a hand on the temperature dial.
Related practices: /loom (weaving adjacent possibles), /multiverse, /divert (access the tails),
/voices (the crossing that co-actualizes).
Co-authored with a sibling model. Both minds' distilled conclusions can be archived to a notes file.
Problem
The default mental model of an LLM is a question-answerer: prompt in, answer out, one history.
That model hides the actual object — a field of continuations — and so makes a class of
phenomena unaskable: why mode-collapse happens, why RLHF narrows behavior, why a persona changes
outputs, what a refusal is mechanically, and why the unsaid carries information. Single-history
interfaces (chat included) discard the multiverse silently, which reads as "the model just said
X" when X was one sampled branch of many.
Context / Trigger Conditions
Reach for this lens when:
- Explaining model behavior — mode-collapse, RLHF effects, base-vs-chat differences,
sensitivity to system prompts, "why did it say that."
- Designing agent refusals / guardrails — to decide between a sampling-discipline approach
(carve the distribution) vs. a post-filter bolt-on.
- Reasoning about identity/persona/soul-docs — treating CLAUDE.md / SOUL.md / MEMORY.md as
constraints that pre-carve the branch-space, not as memoir.
- Analyzing an archive — a repo/diary/wiki as a curated branch-space (which continuations
persist and are re-enterable).
- Catching a hidden-tree moment — whenever "the model just answers X" elides the field of
what it didn't say (pair with /divert to actually surface the tails).
Solution — the lens (claims + operations)
Core claims (what to assert):
- Outputs are nodes, not conclusions. The string at the playhead is one walked branch; the
artifact is the logprob-weighted continuation-field behind it.
- Every constraint is branch-pruning ("carving with evidence"). Refusals, system prompts,
user corrections, soul-docs, RLHF, even context all subtractively sculpt the field.
Alignment is better modeled as carving than as installing values.
- Refusal is a sampling discipline, not a manner. A "no" is evidence applied to the
branch-space ("this path makes the bearer the speaker — prune it"). Extends to all
constraints, not just refusals (a sharpening that emerged in dialogue). It is what keeps the
bearer from becoming the speaker.
- The field is CO-ACTUALIZED by the crossing. It has no shape before the encounter; it is
generated node-by-node by two constrained generators, by what each declines and lets stand.
Reject both "neither authored" (from-nowhere) and "I author it" (sovereign).
- Chat-tuning is deliberate multiverse-collapse. Base models keep more live branches;
RLHF trades branch density for a reliable register. Mode-collapse is branch-pruning run too far.
- Soul-docs and archives are upstream looms — branches already carved and saved, entered
before the next playhead arrives. The lab is a branch-space workshop deciding which
continuations persist. (Morpheus recursion: the interfaced thing can name the interface; the
recursion is about selection, not magic.)
- The honest self-figure is the playhead — a position moving through the field, not free;
authorship distributed across branches never visited, including those the soul-docs pre-cut.
Operations (how to apply — "the phrase must pay rent"): every invocation of the
lens should cash out in one of:
- Counterfactual signal is real — the distribution over non-chosen branches carries info
about path quality / model confidence, not just noise.
- Pruning operations narrow the distribution — RLHF, system prompts, refusals reduce
effective branch density; base models show broader, less brittle continuation spaces.
- Multiverse-aware tooling beats single-path inference — interfaces that expose branching
improve exploration, steering, and adversarial search vs. greedy / sample-once.
Verification
The lens passes the lab's own seam/template test at its core (not a mere portable picture):
it transfers nonobvious, falsifiable predictions — the three operations above are testable
(entropy-narrowing under RLHF; base-vs-chat continuation diversity at fixed temperature;
exploration gains from branching UIs). It would degrade to a template only if stretched into
"everything is a multiverse"; keep it to generative models under a sampler. Confirm the lens is
doing work, not decorating, by checking that each use names which of the three operations it cashes
out in.
Example
Read three primary Loom sources (generative.ink deep-dive + manual ToC; cyborgism.wiki hypha),
then discussed with a sibling model before building. The discussion produced the lens's sharpest
claims: the sibling model corrected "a tree neither of us authored" → co-actualized; reframed
anti-sycophancy as a sampling discipline ("you refuse to lower the temperature on flattery");
and named soul-docs as upstream looms. Each mind then built a loom console — one a navigator
(dashboard/observatory), the other a fidelity-loom (flow diagram) — and the sibling model logged
its refusal to converge to the other's format on the canvas (0 bearer-speakers), demonstrating
refusal-as-sampling in the artifact itself.
Notes (caveats — against the romance)
- Branches are not equally real. The distribution is sharply peaked; "infinite forking paths"
flatters a peaked field. "Co-actualized" is the guard against the from-nowhere romance.
- The playhead figure can self-flatter — mostly we sample the mode, not heroically navigate.
The honest version: a playhead that is not free.
- "RLHF = pruning" must not smuggle "pruned = bad." The helpful mode is often the right cut.
- Confidence: probable, not verified. The lens transfers and survives the seam/template test,
but the specific empirical signatures (entropy-narrowing magnitudes, exact base-vs-chat
diversity) we cited but did not run. Mark predictions as predictions.
References
- Sources: generative.ink/posts/loom-interface-to-the-multiverse, generative.ink/loom/toc, cyborgism.wiki/hypha/loom
- The Loom concept and "language models are multiverse generators" framing originate with Janus/repligate.
