--- name: dpd-verify-edge description: Externally verify a DPD edge that claims a logically-necessary implication (layer='necessary'), using a context-stripped prompt so an independent verifier judges the implication on its own merits rather than rubber-stamping the agent's framing. Records the verdict via record_edge_verification. Invoke when the user runs /dpd-verify-edge, or proactively suggest at a natural pause when list_unverified_edges is non-empty. Requires the dpd-mcp-server MCP server. --- # /dpd-verify-edge — External verification of necessary edges (#42, v0.6) Announce: "Using dpd-verify-edge skill." This skill implements the verification half of **proof-tree discipline** (see the main `/dpd` skill, "Proof-tree discipline" section). It checks edges classified `layer='necessary'` — the claim "these premises *necessarily* imply this conclusion" — by handing a **context-stripped** prompt to an independent verifier. It is the opposite-incentive companion to `/fcot`: `/fcot` *constructs* an argument using full context; this skill *checks* one necessary step in isolation, skeptically. **Compose them, don't merge** — run `/fcot` while building, `/dpd-verify-edge` after, on the necessary edges that resulted. ## When to invoke - User invokes `/dpd-verify-edge [edge_id?]`. - Claude MAY suggest it at a natural pause when `list_unverified_edges` returns a non-empty set — custodially ("necessary edge が N 本未検証です。verify する?"), never auto-firing the external call. ## Step 1 — Select the target edge(s) - If `edge_id` is given, use it (confirm it is `layer='necessary'`; if not, ask whether to proceed anyway). - Otherwise call `list_unverified_edges(session_id)` (necessary edges with no verification record, ordered critical→standard→low→unset). Present the queue and confirm which edge to verify. Default to the highest-priority one. ## Step 2 — Build the context-stripped prompt (the core asset) Fetch the edge and its endpoints (`list_edges` / `get_node`). Build a prompt for an independent verifier. **One prompt_builder** — this exact construction is reused whether delivered by paste or auto-invoke. The stripping line: **include what is needed to UNDERSTAND the proposition; strip what ARGUES for it.** - **INCLUDE**: the from-node text, the to-node text, the claimed relationship (`from` necessarily implies / supports / requires `to`), the edge `type`, local definitions needed to interpret the terms, and any explicit domain assumptions already encoded as nodes in the graph. If a node is shorthand for a formal premise, include that premise. - **STRIP**: the agent's chain-of-thought or proof attempt, the surrounding sibling/parent derivation, *why the agent believes* the edge holds, author confidence / `verification_priority` / prior verdicts, and downstream consequences. Prompt skeleton: ```text You are verifying a single logical-implication claim, in isolation. Do not assume the claim is true; judge it on its own merits. CLAIM: ⟹ (necessarily implies) [edge type: ] DEFINITIONS / EXPLICIT ASSUMPTIONS (only those needed to read the claim): - - ... QUESTION: Does the premise NECESSARILY imply the conclusion as stated? Answer with a first line exactly of the form: VERDICT: holds | holds-with-caveat | refuted If holds-with-caveat, add a second line: CAVEAT: Then a short justification. ``` ## Step 3 — Deliver (paste baseline, optional auto-invoke) Two transports, **same prompt, same parser** — they differ only at the edge: - **Paste (always available)**: emit the rendered prompt in a code block and ask the user to run it in whatever model they trust, then paste back the verifier's reply. - **Auto-invoke (optional)**: if an external verifier CLI is configured/available (e.g. `codex exec`), you MAY run the prompt directly and capture the reply. If it is missing or errors, **fall back to paste** — do not hard-fail. Not vendor-hardwired. Record which transport was used: `method = 'external:'` (auto-invoke) or `method = 'paste'`. `verified_by` is the verifier identity (the model name, or a human name in paste mode) — distinct from `method`. ## Step 4 — Parse the verdict Parse the reply's first line: `VERDICT:` then one of `holds` / `holds-with-caveat` / `refuted` (case-insensitive). If a `CAVEAT:` line is present, fold it into the notes. Everything after is the justification → notes. `holds-with-caveat` is **distinct** from `holds` — never collapse it to "holds + a note". A caveated necessary edge is operationally different (it follows only under an explicit assumption) and warrants a structural fix in Step 5. ## Step 5 — Record + act on the verdict Always: `record_edge_verification(session_id, edge_id, verdict, verified_by?, method?, notes?, prompt_hash?)` — `verdict` is required and comes right after `edge_id`; the rest are optional. Append-only — re-verification adds a row, preserving history. Compute `prompt_hash` from the rendered Step-2 prompt (drift audit). Then, by verdict — **propose, never auto-mutate the graph's epistemic structure**: - **holds** → done. The necessary claim stands. - **holds-with-caveat** → propose one of: (a) rewrite the from-node to fold in the assumption, (b) add an intermediate `assumption` node + edge making the dependency explicit, or (c) weaken the edge to `layer='selective'` via `set_edge_layer`. Let the user choose. - **refuted** → the necessary claim failed. **Do NOT auto-downgrade** (a verifier can misread terms, miss an implicit assumption, or be weaker than the original model). Instead: the record is written (marking the edge as having a refutation = "needs adjudication"), and you *propose* downgrading via `set_edge_layer(layer='selective')` or `set_edge_layer(layer='invalid')` (kept for audit, not deleted). Only auto-apply a downgrade if the user has set an explicit policy (e.g. two independent refutations). ## Composition & boundaries - Separate from `/fcot` (construct vs. check have opposite incentives). A higher-level flow may chain construct→verify, but the primitives stay separate. - This skill never edits End nodes (§5.0 End modification gate still applies to any proposed structural fix). - The verification obligation is **edge-local**: this skill only ever concerns the specific edge(s) selected, never an inferred surrounding region. --- ## Feedback footer After completing a meaningful response (not for trivial status output), print exactly one line at the very end: > 💬 Hit a bug or have feedback on DPD? Run `/dpd-feedback ""` or open an issue at https://github.com/o3co/agent-dpd/issues/new Keep it to one line. Once per skill invocation is enough.