verify-bidirectional

Files (1)

• SKILL.md

SKILL.md (first lines)

# verify-bidirectional — Bidirectional Self-Validation Automation

Automates the processing procedure for bidirectional self-validation. Three stages: user counter-argument → baseline update → next session reflection.

## Trigger Conditions

Immediately **after** this harness AI recommendation/agreement/decision is committed, when the user raises a refinement challenge:

1. **Proposition refinement**: "isn't it more like..." / "I'd say..." / "what about this" / "what's the root?"
2. **Baseline grep trigger**: User cross-references own assets, memory, CLAUDE.md rules, past decisions
3. **Meta-level trigger**: "if you have objections" / "double-check" / "essence" / "root cause"
4. **Side validation result**: User runs independent tools (`/skills` · `gh` · external fetch) → catches mismatch with this harness AI hypothesis

### Natural Language Triggers (works without internal vocabulary)

Also triggered in external user environments by these natural language phrases:

| Phrase | Intent |
|---|---|
| "is that right?", "re-examine this" | Request AI recommendation re-validation |
| "something seems off", "this doesn't feel right" | Counter-argument / refinement |
| "I think you said something different before" | Catch inconsistency with past decisions |
| "can you be more precise?" | Proposition refinement trigger |
| "what's your basis?", "why do you think that?" | Baseline grep trigger |
| "check that one more time" | Self-validation request |

### 🟥 5. Prescriptive doctrine statement — the category that fell through (added 2026-08-16)

Every trigger above is shaped as **doubt about a claim**. A large class of operator correction is not
doubt at all — it is a **standing rule being handed over**, stated as fact or instruction:

| Shape | Real examples (2026-08-16 session) |
|---|---|
| *"X should also include Y"* | *"입장리뷰에는 정적리뷰뿐만 아니라 **동적리뷰도 포함되어야 할 거야**"* |
| *"the real point of X is Z"* (redefining, not doubting) | *"부스팅보다 인큐베이터의 장점은 그 레포 전체를 감싸서 …**모든 것을 조작할 수 있는 권한**이 있는 거야"* |
| *"isn't this too late / wrong-ordered?"* (rhetorical, expects agreement) | *"이 실패가 CI 확인 단계에서야 발견되는 건 매우 **늦은 게 아닐까**"* |
| *"from now on, do W at Z"* | *"앞으로도 마감할 때 그 갈래로 이어갈 수 있게 **알아서 정리해줘**"* |

**Why these were missed 4 times out of 4** — they sit in the blind spot between the triggers (which
expect a *challenge*) and the Exceptions below (which release a *simple correction* with "no
review"). A doctrine statement is neither: it does not dispute a claim, and it is not a one-off fix
to apply and forget. Treated as the exception, it gets a verbal acknowledgment and evaporates at
session end. **Measured**: four such statements in one session, all acknowledged in conversation,
**zero landed in any file** until a later review grepped for them and found nothing.

**The tell is grammatical, and it is language-independent** — the listed phrases above are all
English interrogatives, which is why a Korean declarative (`~해야 할 거야` · `~인 거야` · `~아닐까`)
matched none of them. Do not fix this by appending more literals (Grep-Collision Treadmill, P10);
the discriminator is: **does this utterance describe how I should behave from now on, rather than
what is wrong with this one output?** If yes, it is this category regardless of language or phrasing.

**Required action** — an acknowledgment is not compliance. The statement must land **in a file** in
the same session: canon (`CLAUDE.md` / `knowledge/`) if it governs future behavior, memory if it is
about this operator, a signal file if it needs a decision first. Then say WHERE it landed, so the
operator can see it did.

**Exceptions** (this skill does NOT apply):
- Simple user correction ("this is wrong, redo it") = direct negation → immediate correction (no review).
  ⚠️ **Not the same as category 5 above** — "redo this" is scoped to one output; "from now on do X"
  is a standing rule. When both readings fit, take it as category 5: over-landing costs a file edit,
  under-landing loses the rule entirely.
- This harness AI self-catch (no external counter-argument) = `fact-checker` rule (narrow 1 / broad N+1)

## Execution Steps

### Step 1. Immediate Baseline Update Channel Processing

Treat user's statement as **external refinement material**. **Do NOT attempt to defend against it** — acknowledge possibility of partial weakening or correction of initial recommendation.

Core proposition: "refinement challenge ≠ fundamental negation". Priority is identifying where the initial recommendation is weakened.

**Evidence gate (overwrite ≠ soften)** — closes the sycophancy/steering vector where a bare assertion
("that's wrong, re-examine") flips a baseline with zero evidence (judge-robustness swarm, 2026-06-13).
"Do NOT defend" still holds for *this conversation's proposition* (anti-stubbornness is the point), but a
**persistent-baseline overwrite** (a rule, asset, memory, or `knowledge/` claim — anything that outlives
this session) requires a **supporting basis**, not mere pushback:

- **(a)** the user cited a file / line / commit / URL / past decision **and the cited content actually
  supports the challenge** — verified by *reading it*, not by its mere existence (an irrelevant-but-real
  citation is the out-of-context-grounding trap, the same vector phantom-quench guards), **or**
- **(b)** the Step 2 grep returns an actual contradiction that *supports the challenge* (not just any
  conflict with the original) — surfaced literally.

If a baseline overwrite is implied but **neither holds**, do not silently rewrite: verdict is **ESCALATE**
— surface *"this would overwrite baseline {X} with no cited evidence; confirm override, or provide a
source?"* and block the Step 4 cascade until the operator answers. Softening a local in-conversation
proposition (no persistent asset changed) proceeds as before — the gate fires only on persistent-baseline
writes. **Sequencing**: this gate is written in Step 1, but Step 4 is what enumerates affected persistent
assets — so if Step 4 later identifies *any* persistent asset to write, **re-apply this gate before Step
4.5** even if the original challenge first looked like a mere soften. This is **not** restored AI defensiveness: the AI still does not argue the user is wrong; it only
declines to *fabricate a baseline change* the evidence does not support (mirror of the steel-quench/phantom
mechanical-anchor principle — judged verdicts bind to evidence).

### Step 2. Consistency Area Grep (3-step mandatory)

Grep to find which rules, assets, or propositions conflict with the initial recommendation:

**Scope priority**:
1. `memory feedback_*.md` (especially operating model rules — meta principles, warning lines)
2. `CLAUDE.md` Sync/Push Protocol, asset ownership table
3. `tracks/*/learnings/feedback_*.md` — this harness AI's own rules
4. `knowledge/shared/harness-core/*.md` — higher-level framework

**Mandatory grep keywords** (baseline consistency guard):
- "drift" · "asset ownership" (CLAUDE.md)
- Asset names, abbreviations, identifiers explicitly stated in user's message

**External users**: Replace with your own environment's baseline keywords (prioritize asset names, abbreviations, identifiers from user's message).

### Step 3. Fact-Checker Self-Catch Mark

Mark the corrected weakened proposition as a fact-checker self-catch:

```
fact-checker self-catch #N (narrow 1 / broad N+1)
- This harness AI initial recommendation: {summary}
- Refinement challenge: {user's statement}
- Corrected recovery: {updated proposition}
- Consistency rule: {grep result}

From chrono-meta/forge-harness, MIT.

More from chrono-meta/forge-harness