**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also named the two load-bearing steps automation must keep: check the record first (one idea was already written and superseded), and open the primary (three secondary-source claims were falsified, one of them fabricated by tool extraction).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."

Press Release

abcd gates a new idea before it can mint a single record — researching it against primary sources, grilling it against the existing record, and letting an independent adversary try to kill it. An idea is cheapest to kill in its first hour and most expensive to kill once it has spawned intents, specs, and branches. /abcd:ideate runs the admission interview: state-of-the-art research in which every load-bearing claim is checked against its primary source; a record grill that asks whether the brief, an intent, an ADR, or a principle already covers, contradicts, or supersedes the idea; and an independent adversarial review by an evaluator outside the loop. The output is a verdict and a recorded decision that keeps the rejected alternatives, and only a surviving idea graduates to a draft intent — so the ideas abcd kills leave a reason behind that no one has to re-litigate six months later.

Why This Matters

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."

Acceptance Criteria

• Given a one-line idea, when the user runs the intent or capture verb, then no ideate step is required or suggested as blocking — ideate is an optional verb, never a pre-capture gate; the routing help names it for big, unproven ideas.
• Given /abcd:ideate runs, then it executes three legs in order: primary-source research (every load-bearing claim checked against its primary), a record grill (does the brief, an intent, an ADR, or a principle already cover, contradict, or supersede this?), and an adversarial review that is fresh-context and off-policy — conducted by a session that did not do the research, receiving the idea as an artefact of unknown authorship.
• Given a verdict, then the decision is recorded with rejected alternatives whether the idea survives or dies; only a surviving idea graduates to a draft intent.
• Given the record grill leg runs, then a hit on an existing record entry (covered, contradicted, or superseded) is cited by id in the verdict — the check-the-record-first step that killed one of the three ideas in the proven manual run.

Open Questions

• The verdict's schema and where it is stored (ledger entry vs a dedicated ideate record family) — a spec-time decision.
• Whether ideate can consume a lifeboat or external document as the idea source.

Grill Settlements (2026-07-27)

• Optional verb, never a gate: capture friction stays at one line, and the routing help carries the nudge.
• The adversarial leg codifies the fresh-context, off-policy, unknown-authorship pattern — the only measured debiasing effect, per the salvaged 2026-07-14 research record.

Audit Notes

Fidelity review — receipt rcp-37a162874f75 (verifier abcd:intent-fidelity-reviewer claude-fable-5).

Provenance: abcd:intent-fidelity-reviewer@claude-fable-5 · rubric_hash sha256:1aa9117683cdc84ea4bcbae37c95edab738938331fdb6d39d5d00ea051db4141 · prompt_hash sha256:95792472ae74ca0469f69a51c618946e0d33cb1380032460099ed4b469d67e86 Input attestations: diff:PR 157 main..e05a58ef44b970f7d8db65d7502675bf9a46c3d9: internal/core/ideate/ (ideate.go, record.go, render.go), internal/core/recordid/resolve.go, internal/surface/cli/ideate.go + cli.go ideate wiring, commands/abcd/ideate.md, routing help in commands/abcd/{intent,capture}.md@sha256:e05a58ef44b970f7d8db65d7502675bf9a46c3d9; rubric:.abcd/.work.local/reviews/rcp-37a162874f75.request.md@sha256:1aa9117683cdc84ea4bcbae37c95edab738938331fdb6d39d5d00ea051db4141;

Acceptance rollup: MET 3 · MET_WITH_CONCERNS 1 · NOT_MET 0 · INCONCLUSIVE 0

Per-criterion verdicts:

• ac-1 — MET: The bare intent and capture status renders both print ideateRoutingRule (an optional pointer for a big, unproven idea) with no gate wording, and nothing in the ideate package is reachable from capture or intent-create; TestIdeateRoutingHintIsAPointerNotAGate enforces both facts and passes. evidence: internal/surface/cli/cli.go:1255 — "const ideateRoutingRule = " a big, unproven idea? abcd ideate runs the optional admission gauntlet and records the verdict either way\n"" evidence: internal/surface/cli/cli.go:1119 — "fmt.Fprint(w, ideateRoutingRule) // bare abcd intent render; mirrored at cli.go:1746 for abcd capture" evidence: internal/surface/cli/ideate_surface_test.go:177 — "for _, forbidden := range []string{"must", "required", "before you", "warning"} { ... presents ideate as a gate"
• ac-2 — MET_WITH_CONCERNS: The three legs and their required order are binary-enforced — legOrder is [research, record-grill, adversarial-review] and validate() refuses any leg out of position or carrying another leg's evidence — but the adversarial leg's defining fresh-context/off-policy/unknown-authorship property is delegated to the ideate.md prompt and is by design unverifiable at ingest (validateAdversarial only checks non-empty kill attempts and outcomes), so that debiasing property rests on prompt compliance, not on the recorded artefact. evidence: internal/core/ideate/ideate.go:145 — "var legOrder = []LegKind{LegResearch, LegRecordGrill, LegAdversarialReview}" evidence: internal/core/ideate/record.go:274 — "if leg.Kind != legOrder[i] { ... the legs run in order (%s), and the order is what each leg is looking at" evidence: internal/core/ideate/record.go:367 — "The leg's fresh-context, off-policy, unknown-authorship conduct is the orchestrating prompt's obligation — the binary cannot observe how an agent was run" evidence: commands/abcd/ideate.md:71 — "Strip every authorship signal before handing the artefact over ... This is the only measured debiasing effect in the whole protocol"
• ac-3 — MET: Record() writes the durable verdict for every registered outcome (survives/killed/reframed) and refuses a payload whose rejected-alternatives arrive empty by omission (the explicit no_rejected_alternatives marker is required), while graduation is gated on survival — Result.Graduates is verdict==survives and renderOutcome offers the draft-intent path only for survives, ideate minting no intent itself. evidence: internal/core/ideate/record.go:405 — "if !p.NoRejectedAlternatives { ... the verdict records no rejected alternatives and does not say so explicitly — nothing was written" evidence: internal/core/ideate/record.go:474 — "Graduates: v.verdict == VerdictSurvives," evidence: internal/core/ideate/render.go:119 — "case VerdictSurvives: ... may graduate to a draft intent through\nthe ordinary quoted-text create (abcd intent \"<text>\"). Ideate mints no\nintent itself"
• ac-4 — MET: resolveCitations builds one recordid.Resolver over the repository and refuses the whole verdict (CitationError, nothing written) if any grill hit cites an id that does not resolve; each hit is shape-checked against CitedIDRe first, and the resolved set is carried into the record as CitedRecords — the check-the-record-first gate, exercised green by the surface test that feeds an unresolvable itd-9999. evidence: internal/core/ideate/record.go:454 — "if _, ok := r.Lookup(h.Record); !ok { unresolved = append(unresolved, h.Record); continue }" evidence: internal/core/ideate/record.go:348 — "if !recordid.CitedIDRe.MatchString(h.Record) { ... which is not a record id (want adr-N, itd-N, iss-N, or spc-N)" evidence: internal/core/recordid/resolve.go:76 — "func NewResolver(repoRoot string) (*Resolver, error)" evidence: internal/surface/cli/ideate_surface_test.go:117 — ""ideate", "record", "the-ideate-gate", "--verdict-json", writeVerdict(t, ideateVerdictJSON("itd-9999"))"

Gap audit:

• honoured:
  • Ideate is an optional verb named in the routing help, never a pre-capture gate evidence: internal/surface/cli/cli.go:1255 — "a big, unproven idea? abcd ideate runs the optional admission gauntlet" evidence: internal/core/ideate/ideate.go:19 — "NEVER A GATE. Nothing here is reachable from capture or intent-create."
  • The record is written whether the idea survives or dies, and rejected alternatives cannot arrive silently empty evidence: internal/core/ideate/render.go:123 — "case VerdictKilled: ... The idea is closed. Before proposing it again, read this record" evidence: internal/core/ideate/record.go:405 — "the verdict records no rejected alternatives and does not say so explicitly — nothing was written"
  • Every record-grill hit is cited by id and validated against the live repository, refusing the whole verdict on an unresolvable id evidence: internal/core/ideate/record.go:460 — "if len(unresolved) > 0 { ... return nil, &CitationError{Unresolved: unresolved}"
  • Only a survivor graduates; ideate mints no intent itself evidence: internal/core/ideate/record.go:474 — "Graduates: v.verdict == VerdictSurvives,"
  • Wired on both planes: the abcd ideate verb (CLI) and the /abcd:ideate command surface evidence: internal/surface/cli/cli.go:150 — "root.AddCommand(newIdeateCommand(&asJSON))" evidence: commands/abcd/ideate.md:7 — "# /abcd:ideate — the idea-admission gauntlet"
• diverged:
  • The adversarial leg's fresh-context/off-policy/unknown-authorship conduct is codified in the orchestrating prompt but not enforceable by the binary at ingest — the recorded artefact cannot prove the evaluator was a fresh, off-policy session evidence: internal/core/ideate/record.go:367 — "the binary cannot observe how an agent was run, and pretending to check it would be theatre" evidence: commands/abcd/ideate.md:66 — "The evaluator must not be the session that ran legs 1 and 2. Dispatch it as a separate agent with its own context."
• missing: (none)