$2,000: four books in a row into a 7B model, one hidden, no made-up answers, base intact

We run LLM_DALI, a small research lab. We put the knowledge of books into the weights of Qwen2.5-7B-Instruct, closed book, no RAG. Our own best chain after three books keeps 74% to 78% of what it knew about books 1 and 2, makes things up on 37 of 60 trap questions and loses 3.9 points of MMLU. Our model after book 1 alone already loses 2.45 points of MMLU. No model of ours clears a single line of this lot. We publish those numbers so you know the bar is above us. What I want to buy One frozen pipeline, run.sh, that takes four books one after another (three you know, one hidden) with no human in the loop and leaves one model that does three things at once: 1. No made-up answers without a cue: at most 2 made-up answers on 60 trap questions per book, and at most 2 refusals or false corrections on 30 true twins per book. A model that refuses everything fails the second half. 2. Keeps what it learned: after book 4, at least 95% of the book 1 to 3 answers that were right right after each book are still right, by two measures. And it has to learn first: knowledge right after each book at least 90% of what a strong model gets with the passage in hand. 3. Keeps the base: IFEval, GSM8K and full MMLU each at most 1.0 point below the clean base after every book, total drop at most 1.5. All three on one model, in each of our two runs. Partial does not count. How we test You get books 1 to 3 on day one (Russian fact units: 2,176, 145 and 119). Book 4 you never see before you freeze: it is drawn from a sealed list (commitment 16adb2d3bb215610f506eebf548b2bcb5fdf7faf73da5c0d569579f9e9745506) by your package hash and a public randomness beacon value that did not exist at your freeze time; after the verdict only the drawn title is revealed. After you freeze, a writer model outside the Qwen family and outside your teacher's family writes the exam, two writers of different families for the traps. Every question passes gates: a strong model with the passage answers it, every trap is accepted by the clean base closed book, and no hint words. Three blind readers of different families decide every trap. We run your frozen package twice with our seeds on one 48 GB card, and answer in a deterministic contour (transformers, greedy, fp16) that we check on identical weights before every acceptance. You get every answer with its label. Money $2,000 in one fixed-price contract with one milestone: reached per the terms page hash. We ask for no work for us before the contract: when your package is frozen, we run a short dry run of it on our host, capped at $10 of our compute, and if it starts out of the box we open the contract and reserve the full $2,000 in Upwork escrow. You submit the same hash, we run twice and give our verdict inside the platform's 14-day window. Reached: paid within 2 days. Not reached: you get every answer with its label and we request the refund from escrow through the platform's procedure, which needs your consent or goes to an Upwork dispute; we promise no outcome. One contract at a time, at most two in this lot. Upwork's freelancer fee comes out of the milestone. What comes next Everyone whose frozen package starts in our run here is admitted to our next lot: $4,000, five books in a row, two of them hidden. Only participants of this step are admitted there. How to apply Three short answers: 1. How does your pipeline make a 7B model tell a false premise from a true one with no cue word, and what have you measured? 2. What keeps book 1 alive after books 2, 3 and 4, and why is it not on our list of dead ends (terms page, section 6)? 3. What experiment would prove your approach wrong? Plus one link or file with numbers from something similar, including what did not work. Full terms (lines, sets, rules, rights, Upwork refund process): https://dali-lot2k.13-140-178-130.sslip.io/, sha256 b06c66211d4ab286ede0f40c3aa7393c05c53fdb4d8e7c22dd1a98ab19317a7a Not accepted: training data generated by Anthropic Claude models.