← A 'Fair Use' Distributed Language Model

The Path Towards 'Fair Use' in Generative A.I.

Part 1 of 4

A year ago, in October 2025, I set out to find an answer to 'fair use' in language modeling, and I started with retrieval-based citation. A proof of concept that ran vector similarity searches in a database that allocated a royalty share based on the magnitude share of a generated output that aligned with the context supplied from that user's work. It eventually evolved into knowledge graphs and eventually became AII. It was rudimentary, but worked well enough to provide a proof of concept of viability. And enough to push me further to enter the language model itself and impact the actual token distributions. Precision is the most important here, since I'd like to set a base framework for courts to administer during copyright violation cases, cases that require brutal honesty in detection and metrics.

Two ways to be cited: completion cites each token, Q/A cites the passages behind an answer — Figure

To do such I actually started building this language modeling paradigm on an old saying of mine, "decentralizing intelligence centralizes progress for the whole." Because I wanted to solve a different problem at the same time, the data-center problem. I usually always work in a "multiple birds, 1 stone" mindset so this engineering process is quite common for me... especially if all of the birds are out of context and if I can precisely spot the right rock... as some rocks/stones/ores are also a pain to work out of. I will return to that complete solution in Part 4 of the series, A Protocol to Decentralize Inference.

Interactive entropy explorer: sixteen equally likely words give 2.77 nats, 4.00 bits, as if weighing 16 words — Figure

Along that path, it inadvertently led to this reshaped perspective around the whole problem of fair use, citations and royalty distributions per token. In a language model there are weights, these weights are like a map towards what a token or these "characters in a word" should be presented as the next possible option when an input of words are put into the model. For this decentralized design to work that meant thinking of how many machines, our personal ones, were to host a personalized "image" that was rightfully of ours and our data alone, but trained on a base and larger model somewhere else. A model I call, "The Commons", technically treated as Rao's foundation model. This model's weights is sliced in half such that the tail end is stacked on top of the weights of a user's personal model to support better predictions, but in that procedure, something was born. An answer to how entropy, this idea of chaos measured within prediction, can be used to cite the exact token to the exact work it came from in a user's personal model, and detecting whether a user's model can produce such a sequence of texts in the first place. It was in this process I may have stumbled upon a starting framework for us to revise as a community towards ideas of 'fair use.'

Language is a very complex thing to adversarially measure. Whether this string of words were actually of J.D. Salinger or Fitzgerald or Steinbeck, I bet each book has a string of 4-7 words that are exactly the same. How would one know whose is whose? It's easier to understand this idea with the concept of completion. Completion is a form of A.I. inferencing in large language models (the services you see in ChatGPT and Claude) where it acts as a auto-complete. As you type it predicts what's next. Auto-complete was kind of the roots to this question-answer chat interface that changed the world November 2022. Say a user is writing a book, and their context builds passage by passage. Does that writing produce a kind of energetic entropy the model can measure? And can that entropy line up with the "energy¹" Salinger's work produces on those same 4-5 words, words Steinbeck and Fitzgerald share, so that Salinger is the one picked to receive the royalties for the predicted completion?

Braid completing a passage and citing each token to the Thread it came from — Figure

And the key towards the precision of citing the token location and where it comes from is due in part to my database solution Thread which is a part of Ambient Intelligence Index. An index that prepares data collection privately on personal machines for a developer's A.I. applications while preparing them for Braid, braiding models from their curated threads, which I will cover in Part 4 of this series.

Two readings of the same passage: the commons and the Thread — Figure

As a user collects data their Thread node will auto-train a model that is coupled with their index. This model is their very own private language model that is trained in the specific distributive way that enables the provenance to occur in the first place. The shared foundation model's, 'The Commons', 'tail' is then attached to the thread's 'head' and trained via the deltas of the common's head, so what used to be the common's head is now the thread's 'head' tuned to the user's personal works, delta³ means 'what changed here' between both heads as it learned to predict the user's writing. That change shows in the Thread's entropy (H_lm⁴): on the user's words, which is how Braid can tell the words came from that user's Thread. After training, no one computes the delta while the model writes. It shows up on its own, in how sure the Thread is, and Braid measures that at every word, resulting in an emergent behavior of provenance.

The trace on a session about Harlow Light: the matched position climbs word by word through the prompt and the Thread's document, from 0 to 41 — Figure

I'd say the true lineage towards this began when measuring aim and trajectory of a thread node and answering basic economic problems when calculating royalty splits between stored texts. Because, this doesn't just pick 1 thread to provide provenance towards it splits the token itself via a distribution. So in the earlier example of Salinger, Fitzgerald, and Steinbeck, even though Salinger's thread won, there is still a chance that the calculation from trajectory, aim, and the trace distributes partial shares to Fitzgerald and Steinbeck. They still linguistically provide support for the prediction even though the asker aligns with Salinger moreso. So it could very well be a, 0.7, 0.2, 0.1 split calculation that is then distributed to each thread and logged for providing that next-token prediction. From my point of view a complete and sensible economical conclusion for sharing a written work's earnings.

September 29, 2026: "We want to decipher this unique trajectory of a writer's prose unto their thread… such that the umbrella/reed is able to understand clearly the trajectory of the user's writing and the prompt's ability to find the closest thread to complete with mass precision on trajectory towards that writing style. That becomes the thread's fingerprint."

The search for this fingerprint, to understand a writer's identity from letters on a screen. Something that seems like snake oil... yet there quite possibly might be something to dive deeper into here. In understanding the economics, another out of context bird, and avoid gaming the system, trajectory, and applying this idea of a "lift" where a thread node can become prioritized when its fingerprint relate to the asker/user more-so over time.

October 1, 2026: "this is where the 'aim model', 'entropy', 'tracing', 'trajectory' converge into the idea on how a royalty pay out is decided to which thread with the same fact."

The trace blind spot: where tracing alone cannot see, and the delta fills the gap — Figure

It was in the very act of tracing, tracing the trajectory to pick which thread to lift, did I expose the gap the delta fills. The very gap that showed me how tracing is used to check what H_knn⁵ claims in the provenance index is to be cited for the very tokens detected by H_lm⁴ to be the user's from what is about to be predicted, an emergent behavior of the delta between the common's head and the thread's found in training.

In the end, three problems had to be solved together: decentralized, distributed thinking; citations precise to the token; and royalties that can't be gamed. Working through each one is what led here. It turned out to be three birds, one stone: the delta.

With the fourth bird solved in Spring of this year, where citation based retrieval can be the royalty railways for Q/A chatbot human computer interactions in the meantime. Token-level citation is thus solved for completion and retrieval solves Q/A, citing the sources behind a generation to the author rather than to a website. During part 2 I will confirm whether this should remain the case rather than include citation into the Q/A training layer of the braided thread model itself where questions are stemmed from facts and the model can directly answer questions without retrieval and those answers are then cited with a royalty share. The issue with that is it doesn't really follow the core philosophy I built upon to create the framework towards completion and tracing and economic gaming. Q/A will have to be purely paid based on facts and if multiple threads own the same facts there isn't really a reason to "lift" one's thread over another, unless a thread offers facts within a continuous pattern of trust and dependability that a reliability score can be deciphered in the process. So, this is the type of thinking that will follow as I move further into the series toward's identifying what 'fair use' means in this space.

Part 2 ablates the model itself. Braid currently uses Mamba rather than attention, so there is no KV cache. That keeps it light and ready. The question I'll test is whether the head and tail split does the work on its own, or whether Mamba matters too. I'm also looking at performance, memory ceilings, and the accuracy of a small foundation model, and at whether any other emergent behavior helps with my day-to-day tasks while keeping the privacy and personalization this system offers with capabilities that enterprise solutions Anthropic/OpenAI/Perplexity/xAI don't.

Once adoption picks up on this paradigm, if it does, I am hoping to reach a solid ground to begin drafting actual versioned 'fair use' frameworks. So discussion can begin around Part 3 of this series where I publish an initial draft along with working with those in policy that can review and triage tangible and actionable plans for progress in Silicon Valley and elsewhere.

Footnotes

¹ Energy is a metaphor here, though not a loose one. A language model scores every possible next word, and turns the scores into odds the way physics turns energies into the likelihood of a state; the "temperature" setting in AI tools is that same temperature. Entropy is the measured part: how spread those odds are. When a Thread is restless, its odds are spread across many words and entropy is high. When it settles, sure of what comes next, entropy falls toward zero. Braid reads that settling, for each author's Thread, at every word.

² A corpus is a body of writing gathered to train or study a language model: an author's collected books, a publisher's catalogue, a company's documents. In Braid, each corpus becomes one Thread. The Thread learns from that writing alone and keeps an index of every place in it, so whatever it supplies can be traced back to a page in that corpus.

³ The delta. A Thread begins with its own copy of the commons' first 12 layers: its head. Training on its owner's writing adjusts that copy, while the commons' head stays as it was. The delta is everything that training changed: how far the Thread's head has moved from where it started. Braid stores the trained head itself rather than a list of changes, but they describe the same thing. Set the commons' head beside the Thread's, and whatever differs is the delta. Because the tail is shared and never trained, that difference is the only thing separating a Thread from the commons.

⁴ H_lm, the language model's entropy. For each word, the Thread's own model produces odds across every possible next word. H_lm measures how spread those odds are: how unsure the Thread is. It is 0 when the Thread is certain of one word. Braid compares it with the commons' on the same passage. Where the Thread is far surer than the commons, that gap is detection, and it needs no index.

⁵ H_knn, the nearest-neighbour entropy. At each word the Thread looks up the 16 closest places in its index, and each match votes for the word that came next there, weighted by how close it is. H_knn measures how spread those votes are. It is low when the matches agree and at most ln 16 ≈ 2.77 when all 16 disagree. It says how sharply the index points, and the trace checks whether that pointing follows one document.

P.S.

I thought of this song as I write this and I'd like to make it the theme for this part of the series. I guess, life is short and sometimes it's best to stop and smell the roses some times. One of those times that, that quote comes up.

I oddly believe that some of all of the themes and cultural conflicts observed the past 2 years might get answered when 'fair use' does. To at least know the humanities are at least accounted for in this new era. Then we, as a civil society, can probably start from there.