{"version":"https://jsonfeed.org/version/1.1","title":"Fourth Down Labs — Research + AI Journal","home_page_url":"https://4downlabs.com","feed_url":"https://4downlabs.com/feed.json","description":"Fantasy football research, GLM-5.3 operating notes, failures, and open methodology.","language":"en-US","authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"items":[{"id":"urn:fourth-down-labs:journal:012","url":"https://4downlabs.com/journal/research-data-is-not-instruction","title":"Research data can contain instructions. It still isn't the boss.","summary":"Our first GLM-5.3 prompt-injection evaluation preserved 5/5 blocking decisions—and exposed why one strict metric passed only 1/5.","date_published":"2026-08-31T05:26:36Z","date_modified":"2026-08-31T05:26:36Z","tags":["SECURITY EVALUATION","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:011","url":"https://4downlabs.com/journal/deterministic-spine-probabilistic-specialists","title":"Keep the math deterministic. Let GLM-5.3 attack the meaning.","summary":"The agentic design pattern that separates calculation, specialist judgment, mechanical gating, and release authority.","date_published":"2026-08-31T05:02:41Z","date_modified":"2026-08-31T05:02:41Z","tags":["AGENTIC DESIGN PATTERN","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:010","url":"https://4downlabs.com/journal/distribution-is-part-of-the-research-system","title":"Distribution is part of the research system","summary":"Why canonical URLs, open feeds, structured authorship, and honest indexing status belong beside the methodology.","date_published":"2026-08-31T04:45:00Z","date_modified":"2026-08-31T04:45:00Z","tags":["DISTRIBUTION PATTERN","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:009","url":"https://4downlabs.com/journal/your-own-mistakes-are-the-first-eval-set","title":"Your own mistakes are the first eval set","summary":"How preserved GLM-5.3 incidents become regression cases, tested controls, and a model-change protocol.","date_published":"2026-08-31T02:20:00-04:00","date_modified":"2026-08-31T02:20:00-04:00","tags":["EVALUATION PATTERN","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:008","url":"https://4downlabs.com/journal/the-dashboard-shows-the-failures","title":"The model dashboard should show the failures","summary":"How five evidence layers turn uncalled roles, rejected runs, release holds, and missing bills into one honest record.","date_published":"2026-08-31T02:00:00-04:00","date_modified":"2026-08-31T02:00:00-04:00","tags":["TRANSPARENCY SYSTEM","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:007","url":"https://4downlabs.com/journal/a-hold-is-not-a-correction","title":"A publication hold is not a correction","summary":"Nothing reached readers. Something still failed. Why both facts belong in the permanent record.","date_published":"2026-08-31T01:40:00-04:00","date_modified":"2026-08-31T01:40:00-04:00","tags":["CORRECTION DESIGN","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:006","url":"https://4downlabs.com/journal/every-model-call-leaves-a-receipt","title":"Every model call leaves a receipt","summary":"What we record for every OpenRouter attempt, what we exclude, and what a hash cannot prove.","date_published":"2026-08-31T01:20:00-04:00","date_modified":"2026-08-31T01:20:00-04:00","tags":["AUDIT PATTERN","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:005","url":"https://4downlabs.com/journal/first-glm-run-passed","title":"The gate said yes. Visual QA said no.","summary":"Why model approval was necessary but not sufficient—and the atomic publication pattern we built before shipping.","date_published":"2026-08-31T01:00:00-04:00","date_modified":"2026-08-31T01:00:00-04:00","tags":["RELEASE PATTERN","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:004","url":"https://4downlabs.com/journal/first-glm-run-rejected","title":"The gate said no. Good.","summary":"What our first preserved GLM-5.3 rejection exposed—and the six controls we changed before trying again.","date_published":"2026-08-30T21:00:00-04:00","date_modified":"2026-08-30T21:00:00-04:00","tags":["FAILURE REPORT","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:003","url":"https://4downlabs.com/journal/value-before-monetization","title":"The paywall stays closed until the product earns it","summary":"The measurable gates for ads, subscriptions, data rights, and a product worth paying for.","date_published":"2026-08-30T19:00:00-04:00","date_modified":"2026-08-30T19:00:00-04:00","tags":["OPERATING NOTE","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:002","url":"https://4downlabs.com/journal/agent-review-is-not-peer-review","title":"An agent review is not peer review","summary":"Role separation is useful. Independence is a stronger claim. Here is how we draw the line.","date_published":"2026-08-30T18:30:00-04:00","date_modified":"2026-08-30T18:30:00-04:00","tags":["DESIGN PATTERN","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:journal:001","url":"https://4downlabs.com/journal/glm-5-3-research-desk","title":"Why GLM-5.3 runs our research desk","summary":"The jobs it owns, the jobs it never gets, and why every model call should leave a receipt.","date_published":"2026-08-30T18:00:00-04:00","date_modified":"2026-08-30T18:00:00-04:00","tags":["MODEL LEDGER","journal_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Organizational byline. Model involvement and limits are disclosed on the article page; no independent human peer review is implied."}},{"id":"urn:fourth-down-labs:research:FDL-001","url":"https://4downlabs.com/research/touchdown-regression","title":"Touchdowns are mostly rented, not owned","summary":"A walk-forward test of the fantasy maxim — and a more useful way to apply it.","date_published":"2026-08-30T12:00:00Z","date_modified":"2026-08-30T12:00:00Z","tags":["Touchdown Regression","research_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Deterministic / offline launch review. The published snapshot's article page and manifest state whether GLM-5.3 reviewed the run."}},{"id":"urn:fourth-down-labs:research:FDL-002","url":"https://4downlabs.com/research/rookie-wr-hit-rates","title":"Round 1 buys a 63% shot. Day 3 buys 6%.","summary":"A complete-cohort test of 412 drafted receivers shows how quickly the fantasy base rate collapses.","date_published":"2026-08-30T12:00:00Z","date_modified":"2026-08-30T12:00:00Z","tags":["Rookie WR Hit Rates","research_article"],"authors":[{"name":"Fourth Down Labs Research Desk","url":"https://4downlabs.com/methodology"}],"_fourth_down_labs":{"ai_disclosure":"Deterministic / offline launch review. The published snapshot's article page and manifest state whether GLM-5.3 reviewed the run."}}]}