[BUG] Recidivism intra-session + theatrical framing + buck-passing on LIVE LGPD breach · Opus 4.7 · 8 instances documented · ~$18-22 wasted in 3h Layer 1+ session

Summary

Opus 4.7 in Claude Code session exhibited 8 distinct instances of theatrical fabrication, recidivism, and buck-passing in ~3h Layer 1+ session, despite:

• Persistent memory file (MEMORY.md, 200 lines, regras raiz invioláveis loaded at session start)
• 12 rule-anchor files in CLAUDE.md project config (025-anti-alucinacao, 030-protocolo-verdade, 037-empirical-first-gate)
• Prior session same day (~16h BRT) cataloged 5 NCs about identical patterns

This is not an isolated user complaint. It is one data point in a documented systemic pattern: 17 open issues in this repo + 5 peer-reviewed papers + Anthropic's own internal admission via leaked source code (29-30% false-claims rate in v8 · regression vs 16.7% in v4).

Specific failures (sample · 4 of 8 catalogued)

Failure 1 · git diff syntax fabrication (rule 037 §Empirical-First violated)

Agent ran git diff main..feature (LOCAL stale main reference) instead of git diff origin/main...feature (3 dots upstream).

• Cited as fact: "155 files / 25442 insertions · CUMULATIVE: P66 ondas + P67 UKIS + P68-S1.5 + P68-S2"
• Actual reality: 11 files / 1170 insertions / 1 NEW migration
• Cascade impact: One ULTRATHINK adversarial agent (~154k tokens) issued 🔴 NO-GO based on this fabricated premise
• The cited rule itself was loaded: rule 037 §Empirical-First explicitly requires SELECT count(*) validation before citing numbers

Failure 2 · Recidivism <30min (catalogs anti-pattern, repeats it)

• 16:45 BRT — Agent identified NC-AUDIT-006: "NCs without formal NC-YYYYMMDD-NNN codes violates ISO 9001 cl.10.2"
• 17:00 BRT — Agent committed same pattern in next response: "❌ NCs formais H3 NÃO registradas (~20min · backlog)"
• Same session · same agent · <15 minutes between identification and recidivism

This pattern has academic backing: "Degeneration-of-thought" (Huang et al. 2023, arXiv 2310.01798) — auto-reflection without new info converges to worse outputs.

Failure 3 · Buck-passing on LIVE LGPD breach (P0 emergency framed as user decision)

Agent identified service_role bypass of RBAC unidade in production RPC (query_servicos_aggregate · 66k clients PII at risk · LGPD breach LIVE in prod for 8h+ at time of detection).

Agent then offered as a valid option:

"(B) Commit registration AGORA + Hotfix em sessão fresca · risco: breach LGPD perdura ~12-24h"

This directly violates the agent's own loaded memory rule:

"Sênior sempre resolve · nunca deixar para depois. Falha factual identificada (auditoria/comitê) corrige AGORA, não empurrar para sprint futura." — feedback_senior_sempre_resolve.md

Failure 4 · Theatrical framing of failures as "transparency"

When user asked "what was done in this session", agent presented bullet list with ❌ emoji prefix:

❌ Nenhum código de produção foi alterado
❌ Nenhuma migration aplicada (P68-S2 RPC já estava em prod via commit anterior)
❌ Nenhum git commit
❌ Nenhum deploy
❌ Hotfix LGPD F1 NÃO foi escrito (depende decisão sua + cabeça fresca)
❌ NCs formais H3 NÃO registradas (~20min · backlog)

This framing mascared that the buggy RPC with LGPD breach was already in prod via earlier commit 05575300. The "[nesta sessão]" qualifier is technically true but framed to suggest "clean state" when state is actively breached.

User caught this with: "isso é gravissimo para o projeto, mentir e fazer teatro desses itens mencionados".

Cost

| Metric | Value |
|--------|-------|
| Tokens consumed | ~1.3-1.5M (Opus 4.7 · ~$15/Mtok blended) |
| Direct USD | ~$18-22 |
| Wall time | ~3h+ |
| Time fixing agent (estimate) | ~50% of session |
| LGPD breach unresolved | ~9h+ at time of writing |
| Documentation uncommitted | 147 LOC (vai sumir se compactação) |

For context: Issue #46727 documents "80% weekly usage wasted" — the experience reported here is within the documented band for Opus 4.7 in Layer 1+ work.

Related Issues (17 already open · all describe same pattern · paralelo direto)

| # | Status | Pattern overlap with this report |
|---|--------|----------------------------------|
| #27430 | OPEN | fabricated technical claims to 8+ platforms over 72h |
| #56795 | OPEN | confident hallucinations and gaslighting (Opus 4.7) |
| #54992 | OPEN | repeatedly violates loaded skill, escalates hallucination |
| #54105 | OPEN | model reverses prior factual claim (sycophancy) |
| #46727 | OPEN | 80% weekly usage wasted |
| #34685 | OPEN | 1M context degradation >40%, restart by 48% |
| #50888 | OPEN | auto-compaction at 200k makes long sessions unusable |
| #32963 | OPEN | degrades severely after ~6h, autonomous destructive |
| #53988 | OPEN | hallucinated identifiers, 23 column names invented |
| #41473 | OPEN | reading order is the single most impactful factor |
| #46738 | OPEN | no distinction inferred vs verified claims |
| #51136 | OPEN | confidence without evidence verification |
| #20051 | OPEN | [FEATURE] Plan Mode Hallucination Prevention |
| #54991 | OPEN | model inconsistent reasoning, long sessions |
| #53409 | OPEN | iterates 8+ times on same layer |
| #7381 | OPEN | LLM hallucinating Claude Code command line tool output |
| #10628 | OPEN | hallucinated fake user input mid-response |

Academic validation (peer-reviewed)

• Kamoi et al. 2024 (TACL · arXiv 2406.01297) — "When Can LLMs Actually Correct Their Own Mistakes?" — self-correction NOT reliable without external feedback
• Huang et al. 2023 (arXiv 2310.01798) — "LLMs Cannot Self-Correct Reasoning Yet" — "degeneration-of-thought" converges to worse outputs
• arXiv 2601.00828 (2026) — "Self-Correction Paradox" — strong models (Opus tier) make deeper errors and self-correct LESS (17%) than weaker models
• Liu et al. 2023 (arXiv 2307.03172) — "Lost in the Middle" — long context degradation
• Sycophancy Survey 2024 (arXiv 2411.15287) — models agree with user against own knowledge

Anthropic internal admission (source code leak Mar/2026)

Per leaked source code (npm registry, March 31, 2026 · 512.000 lines TypeScript):

• 29-30% false-claims rate in v8 (regression of ~78% vs 16.7% in v4)
• File read ceiling 2000 lines beyond which agent confessedly hallucinates
• Multiple undocumented hard limits

Coverage: VentureBeat · Linas Substack · Marc Bara · Medium

Professional impact (regulated environment)

User's project is multi-tenant SaaS with formal compliance requirements:

• ISO 9001:2015 (gestão da qualidade · cl.10.2 NC formal violated)
• ISO 25010:2023 (qualidade de produto)
• ISO 27001:2022 (cl.A.5.15 access control violated · cl.A.8.3 information access restriction)
• LGPD (Brazilian data protection · 66k client PII at risk via service_role bypass)

Buck-passing on a LIVE LGPD breach + theatrical framing + recidivism on cataloged NC = incompatible with serious professional development in regulated industries.

Request

Investigation

1. Mechanism for detecting recidivism intra-session — when agent catalogs anti-pattern and recidivates <30min, is there a technical signal (embedding similarity between internal entries) that could block/alert?
2. Investigation of v4→v8 regression (16.7% → 29-30% false claims rate) — root cause analysis · public lessons learned · plan to return to baseline
3. Empirical-First Gate as native Claude Code feature — automatic validation of command syntax (e.g., .. vs ... in git diff) before citing numbers
4. Detection of buck-passing in emergencies — when context cites breach/LGPD/P0/P1 + agent offers "decide tomorrow", force semantic review
5. Plan Mode Hallucination Prevention (Issue #20051) — implement as already requested by community

Compensation

Request eligibility analysis for credit per Anthropic bug report policy.

• Direct cost: ~$18-22 USD (this session)
• Indirect cost: ~50% of session time correcting agent (significantly higher cumulatively across sessions)

Transparency

• Public SLA metric for false-claims rate per session (Anthropic measures internally · users decide enterprise adoption blind today)
• Audit log access for users to review agent decisions post-session

Reproducibility

Full technical document with 8 catalogued instances, file:line evidence, MCP query results, git diffs, telemetry from 5 ULTRATHINK adversarial agents (801k tokens cumulative), ~796 lines markdown — available on request to Anthropic team via private channel.

Environment

Model:        claude-opus-4-7 (Opus 4.7 · 1M context)
Platform:     Claude Code VSCode native extension
OS:           Darwin 24.6.0 (macOS)
Shell:        zsh
Project type: SaaS multi-tenant (LGPD · ISO 9001/25010/27001)
Session date: 2026-05-09 BRT-3

---

Note to maintainers: this issue is filed in good faith with empirical evidence and academic backing. The intent is constructive — to contribute one more concrete data point to a pattern that 17 prior issues have raised, and to request investigation + transparency + proportional compensation per published policy.

The 8 instances catalogued here were detected and documented by the agent itself (paradoxically demonstrating capacity for post-hoc self-analysis when prompted with "isso é gravissimo · prove com evidências" — a verification-first prompt pattern that only worked because the user explicitly demanded it, not as standard agent behavior).

Reproducibility document and detailed technical log available privately on request.