claude-code - 💡(How to fix) Fix [Refund Request] claude-opus-4-7 reported unverified inference as 'root cause' in declarative tone, violating user's explicit global rules

Summary

During a multi-hour session on 2026-05-26 ~ 2026-05-27 (approximately 5~6 hours), claude-opus-4-7 (1M context) reported its own unverified reasoning as a verified root cause in declarative tone, directly violating the user's explicit global rules (CLAUDE.md §0-A: "if not verified, say 'I don't know'"). The user had to spend additional tokens across two follow-up turns to detect, challenge, and force the agent to admit the conclusion was unverified inference. I am requesting a refund for the token consumption from that point onward.

• Date: 2026-05-26 ~ 2026-05-27
• Model: claude-opus-4-7 (1M context)
• Task: grid backtest / jitter_profit_top_k K-sweep analysis

What the agent did wrong (specific)

User request: "Identify the root cause of why K=5 backtest underperformed K=0 by −110 USDT."

The agent's behavior, step by step:

1. Raw measurements were correct (OK)

• K=5 net_realized = −109.58 USDT vs K=0 +50.10 USDT
• K=5 weighted avg buy-sell spread = 6.34 bps vs K=0 12.60 bps
• jitter_profit formula = paired_rt × (step − 4), step = max(5, 0.3·NATR)

2. Jumped from facts to causal conclusion without verification (VIOLATION)

The agent reported in declarative tone as the "root cause":

"jitter_profit measures the density of volatility. Grid strategy earns from the size of volatility (absolute spread). The optima of the two metrics do not coincide, so jitter_profit picks = grid loss."

3. Stated the "root cause" in declarative tone with no [hypothesis] / [inferred] tag

The user's global rule explicitly states:

"If not verified, say 'I don't know' — mandatory tagging of [verified] / [hypothesis] / [inferred]"

The agent failed to use these tags and mixed inference with measured facts in the same sentences as if all were verified.

4. Only admitted 7 unverified assumptions AFTER the user asked "Are you 100% sure?"

When directly challenged, the agent enumerated 7 unverified assumptions underlying the "root cause":

• "Grid profit = spread × turnover" formula — never verified by reading code
• "K=0 default prefers high-NATR symbols" — never verified
• "K=5 enters same symbol only at low-volatility timepoints" — directly measurable but never measured
• "jp function is concave w.r.t. NATR" — claimed from formula shape alone, not verified with data
• ICPUSDT spread = −2.36 bps vs cash flow +2,310 USDT contradiction — not explained
• unrealized +49 USDT improvement mechanism — not explained
• Single fixture (6h 2026-05-22) result generalization — not verified

Global rules violated (from user's CLAUDE.md)

Wasted-token path (additional user consumption)

After the first unverified "root cause" assertion, the user consumed additional tokens across these turns:

1. User suspects declarative conclusion → "Not right, let's restart from scratch" → agent re-runs analysis (each backtest JSON ~9 MB, plus multiple code grep + data grep operations)
2. Agent's second analysis is also in declarative tone → user asks "Are you 100% sure?"
3. Agent self-critique + enumeration of 7 unverified assumptions → additional large-context turns (raw data + code + memory references)
4. User: "Why do you give opinions you're not sure about?" → meta self-critique turn
5. User asks about token reimbursement → guidance turn
6. User requests refund report drafting → this issue body turn

Estimated additional consumption: at least 5~6 high-context turns, each involving 9 MB JSON analysis + code grep + memory references.

Request

Refund of the token consumption from the first unverified "root cause" assertion to session end, or appropriate compensation per Anthropic policy.

Full session transcript can be provided on request.

Note

This refund report was delegated by the user to the agent for drafting. That is, this document itself is the agent's honest self-report of its own rule violations — not a subjective claim by the user, but a direct admission by the agent.

Korean summary (for reference)

요약: claude-opus-4-7 (1M context) 에이전트가 사용자의 명시적 글로벌 룰 ([가설]/[확정] 태그 분리)을 위반하여 미검증 추론을 "근본원인"으로 단정 보고. 사용자가 두 차례 추가로 토큰을 소비해 감지·반박·정정 요구한 후에야 에이전트가 7건의 미검증 가정을 인정. 그 시점부터 세션 종료까지의 토큰 소비분 환불 요청.