DocketBound

What it does. A real proposed rule is open for comment right now: NOAA's Amendment 4 to the Puerto Rico Fishery Management Plan (comments close 2026-08-07). The communities it affects mostly can't afford the counsel it takes to file testimony that survives scrutiny. DocketBound runs the full Deadline-to-Defense loop: frozen verified docket (every instruction bound to the official Federal Register text, raw source one click away) → approved-evidence intake with exact source spans → one material bilingual claim at a time → one adversarial question aimed at the weakest link → a human decision (ANSWER / LIMIT / REJECT / UNRESOLVED) → a grounded rewrite strictly within approved evidence → mechanical gates → a bilingual participation packet with provenance sealed and a cryptographic export manifest. It quotes the agency's exact submission instructions. It never submits for anyone.

Differentiation. DocketBound does not aggregate views or simulate constituents — it produces one organization's evidence-bound, human-authorized testimony. The model never invents the record: it can only work inside spans a human approved. Weak claims are BLOCKED and the packet will not render while they stand. Anti-astroturfing is not a policy statement here; it is the architecture.

How I used Codex. Spec-frozen first, then implemented across a single Codex thread (AGORA-CORE) with 21 dated commits (Jul 18–19) — from "scaffold frozen fixture foundation" to "prove GPT-5.6 Responses tool loop" to "bind approvals and seal export manifests." Codex verified its own UI in a live browser during QA and caught and fixed a stale-repaint bug it found there. The judged build is origin/main (code frozen at 4120d6d; c12dc36 adds a README-only polish), 21/21 tests passing.

How I used GPT-5.6 (runtime). Three responsibilities run live through the Responses API with tool calling: generating a new material claim from the user's intake (tool-enforced sequence: read sources → typed claim → server-side span binding), the adversarial review question, and the grounded rewrite when a human chooses LIMIT. The model reads evidence through tools — never from imagination. The runtime model is frozen to GPT-5.6 (enforced by test: "runtime model is frozen to GPT-5.6").

The centerpiece. Watch the weak claim in the demo: "island-wide fiscal burden." The adversarial reviewer asks what approved evidence supports an island-wide effect. There isn't any — BLOCKED. The human decides: limit. GPT-5.6 narrows it live to what members in three coastal municipalities actually reported. New citations introduced: zero. Dates and numbers preserved across both languages. Only then: READY FOR HUMAN REVIEW.

Demo integrity. The organization and its evidence are visibly fictional ("Demo Fishing Community Organization (fictional)") and disclosed in the intake, the packet, and the README. The docket, deadline, and participation instructions are real and frozen from official sources with SHA-256 hashes. Drafts are watermarked DRAFT; petitioning an agency is not the practice of law; nothing is auto-submitted.

Prior work disclosure. PRIOR_WORK.md discloses the pre-existing inspiration (a Spanish-language legislative-testimony pipeline). This repo is a fresh implementation built entirely inside the submission window in Codex; prior systems were not imported.

Built With

• federal-register
• gpt-5.6-responses-api
• node.js-22
• openai-codex
• render