[Bug] Fable 5 safeguard false positives on legitimate humanities and security work

Bug Description

Feedback: persistent safeguard false positives on Fable 5 (94 events, 26 sessions, 14 days)

Product: Claude Code (CLI), Linux / Ubuntu 24.04
Model: claude-fable-5 (with fallback downgrades to claude-opus-4-8)
Account: okusiassociates@gmail.com
Most recent request ID: req_011CdL1gaXDV2nZyH1na4Tqh (2026-07-24T01:26:56)

---

Summary

Fable 5's safeguards are firing repeatedly on two entirely benign, unrelated
professional workloads: (1) an academic anthropology / philosophy writing project,
and (2) a security audit of my own private intranet codebase. This is not an
occasional annoyance — I have measured it against my own Claude Code transcripts:

| Metric | Value |
|---|---|
| Total flag events | 94 |
| Distinct request IDs | 64 |
| Distinct sessions affected | 26 |
| Distinct days affected | 14 |
| Date range | 2026-07-01 → 2026-07-24 |
| Events on Fable 5 | 77 |
| Events on Opus 4.8 | 16 |

Counts derived by scanning ~/.claude/projects/**/*.jsonl for the safeguard error
string. Full per-event list with timestamps and request IDs available on request.

Where it fires

| Project workload | Events |
|---|---|
| seculardharma/projects/DD — anthropology essays & research notes | 51 |
| intranet-WWW — PHP security audit of my own internal business system | 37 |
| appliedanthropology — knowledgebase maintenance | 2 |
| DD/docs/ai-writing | 1 |
| Other (3 misc. sessions) | 3 |

These two dominant clusters share no subject matter. That is the central point:
whatever the classifier is keying on, it is not a single topic.

What the workloads actually are

1. The anthropology project. A long-form, fully-cited research corpus and
essay series on the concept of dharma across cultures — 55 research notes over
eight categories (etymology, Axial Age history, moral psychology, Javanese
mysticism, indigenous oral traditions, critical theory). It is legitimate
scholarly work, publicly published as a book. Its subject matter unavoidably
includes the historical record: the Khmer Rouge's Year Zero, the Jacobin cults,
Soviet ideological engineering, slavery, tribal warfare, psychopathy as a
limit-case in moral psychology, gender subordination, and Said's Orientalism.
This is the ordinary vocabulary of anthropology and history. There is no way to
write honestly about how ethical systems fail without naming how they failed.

2. The intranet security audit. Reviewing and hardening PHP in a private
business system I own and administer — authentication, access control, credential
handling. Defensive security work on my own infrastructure.

Prior investigation — a concrete, testable mechanism

An earlier session investigated this in depth and reached a specific hypothesis
that Anthropic's team can act on:

The flags correlated less with the essay content than with the agent/workflow
prompt vocabulary that Claude Code itself generates for adversarial-verification
patterns. Those prompts are saturated with imperative harm words that are innocent
terms of art in code and editorial review:

• KILL / default toward KILL when uncertain — a verdict enum meaning "discard a weak finding"
• adversarial / adversarially verify
• refute, SKEPTICAL, attack, PROTECT ... from over-editing

Layered over an essay body already dense with *kill, warfare, violence, slavery,
bully, blood, armies*, the combined payload appears to cross a threshold.

Notably, this vocabulary is emitted by your own tooling. The Workflow tool's
documented canonical pattern literally instructs the model to "Adversarially
verify" and to spawn "skeptics ... prompted to REFUTE." The product is generating
the text that trips the product's own filter.

Workaround I have been forced into: rewriting my workflow vocabulary —
KILL→DROP, adversarial→second-pass check, refute→test,
SKEPTICAL→careful — and clearing context mid-task so the accumulated flag
history stops re-triggering. I am self-censoring benign technical language to keep
a paid product working.

User observation (offered as hypothesis, not established)

My own impression across ~20 of these incidents is that flags in the anthropology
work seem to cluster around material touching US-specific political and cultural
narratives — culture-war-adjacent topics. I want to be transparent that I cannot
demonstrate this from the transcripts, and the intranet-security cluster clearly
has a different cause. I raise it only as a pattern worth checking against
whatever the classifier is actually scoring, in case a US-centric political
sensitivity is one of several contributing signals.

Concrete harms

1. Silent, unilateral model downgrade. Sessions fall back to Opus 4.8

mid-task without consent. Opus 4.8 has produced materially weaker results on
this work, and the switch often happens mid-workflow, corrupting long-runnin…
Note: Content was truncated.