2026-07-27 — GRILLED itd-101/102/103/104 to settlement (six decisions, each now in its intent's Grill Settlements + acceptance criteria): itd-103 guard fails open loudly, blocker/warn tiers with committed-only overrides, shell-token-aware command-position matching with an itd-81-style TNR-floored corpus; itd-101 refresh is manual-with-nagging (scheduled CI later, own sign-off), staleness warns at 180d and blocks releases at 365d, human verifications age on the same clock; itd-102 identity lives in a parseable markdown block (config points at it), check is warn-tier and never rewrites; itd-104 ideate is an optional verb, never a pre-capture gate, with a fresh-context off-policy adversary. All four drafts are grill-settled and await intent plan.

Frontmatter

.abcd/work/decisions/dcn-2607270000000005-grilled-itd-101-102-103-104-to.md
open on GitHub ↗ · commit history ↗

mentions5

• itd-81No Judge Ships Unmeasured, And A Clean Diff Is A Result
• itd-101**Every citation abcd publishes is provably alive and honestly labelled — and the gate that enforces it never flakes.** A cited reference page is only as trustworthy as its links, and links rot silently: pages retitle, URLs redirect, whole platforms announce their own shutdown. abcd splits citation validation across a deterministic gate and an explicit refresh. An offline lint family checks structure and source policy on every commit — footnote markers and definitions in bijection, URLs and DOIs well-formed, aggregator domains refused — with zero network in the gate. `abcd docs cite refresh` does the live fetching on demand and writes a committed baseline: per URL, the final resolved address, when it was last checked, and whether verification was automatic or manual — never how. Sources that block automated fetchers join a manual queue the maintainer clears link by link: printed as a checklist first, later as a generated, disposable checklist page that hands back a receipt file the verb ingests. The lint then enforces the baseline offline — no broken entries, no stale entries, no cited URL without a receipt, no citation whose recorded final address has drifted. The engine is native and dependency-free first; a specialist link checker can slot in later behind the same seam. "My release gate stays deterministic and my citations stay alive," said Kira, an open-source maintainer. "When a source needs human eyes, abcd hands me the exact list to click through and records that I did — not how I did it."
• itd-102**Your repo says the same thing about itself everywhere — because abcd asked you once and holds every surface to it.** A project's positioning fragments silently: the README strapline, the package or plugin manifest description, and the agent conventions file are each edited at different moments, and soon three surfaces tell three stories about what the project is. When abcd prepares a repo it interviews the user for the project's identity — title, tagline, and a short elevator pitch — and records the answers in one canonical home in the development record. Every surface rendering derives from that home, and a deterministic positioning check compares the surfaces against it on every audit: a mismatch is highlighted with the exact drifted line, and the fix is either re-render from the record or a deliberate, recorded identity change — never a silent extra variant, and never a silent rewrite by abcd. "I changed my tagline once, in one place, and abcd chased it everywhere else," said Alice, a solo founder. "When a stray edit crept into the README, the next audit pointed at the exact line instead of letting the drift settle in."
• itd-103**abcd teaches repo agents the shell commands they must never run — and blocks them when they try.** An agent that runs `cd scratch && rm -rf *` is one failed `cd` away from deleting the working tree, and the facilitator watching the session usually has no way to know that. abcd ships a hazard registry: each entry a dangerous command pattern, its severity, the safe successor, and a plain-language why. One registry, two planes. The rules loader injects the matched safety rules before shell-heavy work, so the agent is taught the safe form up front; a deterministic guard — a core verb any harness hook can call — refuses a matching command at execution time and replies with the safe successor, so the block itself is the lesson. Hosts without hook support still get the teaching plane; hosts with hooks get both, wired in at install. The registry grows from reality: a facilitator who sees something scary needs to know exactly one move — capture it — and recurring captures promote patterns into the bundled defaults, each entry shipping with a fixture that proves its guard fires. "I couldn't have told you why that command was dangerous," said Nia, a facilitator. "I didn't have to. The guard refused it, told the agent what to run instead, and told me why in words I understood. The one time something new scared me, I captured it and moved on."
• itd-104**abcd gates a new idea before it becomes a record entry: research it against primary sources, grill it against the existing record, and let an independent adversary try to kill it.** A product thinker's exciting idea is cheapest to kill in its first hour — and most expensive to kill after it has minted intents, specs, and branches. `/abcd:ideate` runs the admission interview: SOTA research in which every load-bearing claim is checked against its primary source; a record grill — does the brief, an intent, an ADR, or a principle already cover, contradict, or supersede this?; and an independent adversarial review, evaluator outside the loop. The output is a verdict plus a recorded decision with rejected alternatives; only a surviving idea graduates to a draft intent. The protocol is already proven by hand: three ideas entered it on 2026-07-14 — two were killed (one by a measured null result its premise could not survive, one by three independently fatal methodological defects) and one was reframed by adversarial review before a line was built. The manual run also named the two load-bearing steps automation must keep: check the record first (one idea was already written and superseded), and open the primary (three secondary-source claims were falsified, one of them fabricated by tool extraction). "The ideas I'm proudest of are the ones abcd killed in an afternoon," said Iris, a product thinker. "Each one left a recorded reason behind, so I never talk myself back into them six months later."