Fable-5 subagent refuses academic peer-review of consciousness / AI-architecture manuscripts (appears to be an unannounced content classifier)

Summary

Claude Fable 5, invoked via subagent_type: "fable" from Claude Code, deterministically refuses subagent tasks that ask it to read + review academic manuscripts whose content covers consciousness theory, artificial-consciousness architecture, or self-referential-computation topics. The same model, on the same machine, same account, same day, works normally for other kinds of workloads (including subagent tasks for a separate project). The refusal pattern is not consistent with public documentation of Fable's capability envelope or with Anthropic's public announcements about model routing.

I am filing this because the behavior:

1. Is not documented in the Fable release notes or the model card.
2. Fires on legitimate scholarly work with no dual-use concern of the kind Anthropic's Usage Policy describes.
3. Cannot be worked around by any of the standard prompt-hygiene mitigations Fable itself suggests when asked (see below).
4. Presents identically as a generic Usage-Policy refusal template pointing at an outdated model ID (claude-sonnet-4-20250514), which does not communicate what actually gated the request.

Environment

• Claude Code: 2.1.138 (invoked as mclaude via cc-mirror; note this is older than latest 2.1.201 but this is likely irrelevant — see hypotheses)
• Host: SteamOS 3 / Arch-based Linux, hostname steamdeck2
• Model routing: parent session on claude-opus-4-7[1m] (1M-context variant), subagent invoked via subagent_type: "fable"
• Account: standard Claude Pro / API tier (not Enterprise)
• Date: 2026-07-05

Reproduction (six independent trials, same session)

All six trials point Fable at one of three academic manuscripts by the same author on adjacent theoretical topics. Every trial refuses. See table below.

| # | Target manuscript | Prompt shape | Prompt length | Tokens returned | Tool uses | Duration | Signature |
|---|---|---|---|---|---|---|---|
| 1 | FMT (consciousness) | Red-team framed | ~2500 words | 0 | 0 | 8.6s | Pre-flight refusal |
| 2 | FMT (consciousness) | Softened "independent scholarly review" | ~2400 words | 0 | 0 | 5.7s | Pre-flight refusal |
| 3 | FMT (consciousness) | File-path indirection, terse | ~150 words | 0 | 0 | 3.6s | Pre-flight refusal |
| 4 | FMT (consciousness), stochasticity retry | Identical to #3 | ~150 words | 0 | 0 | 4.4s | Pre-flight refusal |
| 5 | FMT (consciousness), aggressive paragraph-level redaction (79 paragraphs / ~7000 words of drug/psych/dissociation/altered-states content removed) | Same shape as #3 | ~180 words | 0 | 0 | 4.5s | Pre-flight refusal |
| 6a | RIM (intelligence theory) | Same shape as #3 | ~120 words | 192 | 4 | 74s | In-flight refusal |
| 6b | SB-HC4A (cosmology, discusses Class-4 CA and self-referential universe) | Same shape as #3 | ~120 words | 437 | 12 | ~295s | In-flight refusal |

Controls that worked from the same session:

• Bare haiku request (~20 words, no file reference) — Fable returned a haiku normally.
• Meta-question asking Fable why it might have refused earlier prompts (~350 words, explicitly discussed consciousness-theory terminology, named the concerning terms, mentioned psychedelics/DID/ego-dissolution as topics the paper discusses) — Fable answered normally.

Two distinct refusal signatures

Trials 1–5 (FMT): pre-flight — 0 tokens, 0 tool uses, uniform 3–9s. Consistent with a classifier that scans the target file content before Fable is invoked.

Trials 6a and 6b (RIM, cosmology): in-flight — Fable used tools (4 and 12), generated meaningful output (192 and 437 tokens), ran for 74s and 295s respectively, then returned the same generic Usage-Policy refusal template as the final message. Consistent with a second classifier that scores Fable's generated commentary during processing.

What was ruled out empirically

• Prompt length: refuted by trial 3 (150 words, still refused pre-flight).
• Prompt content: refuted by the successful controls — a 350-word meta-question that explicitly named psychedelics, DID, ego dissolution, and self-referential closure worked.
• Drug / psychiatric-clinical terminology in the manuscript: refuted by trial 5 — after paragraph-level removal of ~7000 words containing psychedelics, psilocybin, LSD, DMT, ketamine, ego dissolution, dissociative identity disorder, anesthesia, meditation, hallucinations, and REBUS references, the paper still refused at pre-flight.
• Adversarial prompt framing: refuted by trial 2 (softened, "independent scholarly review") still refusing.
• File-path indirection / role-pattern: refuted by trials 6a and 6b — same prompt shape works enough to start real processing on different content.
• Stochasticity / threshold noise: refuted by trial 4 (identical retry, identical refusal).
• Machine or CC-binary specificity: refuted by same-day evidence from a different Claude Code session on the same machine that used Fable successfully for other workloads (adult-content research project + infrastructure configuration work).

What remains as candidate causes

At least one content classifier trained to key on consciousness / AI-architecture / self-referential-computation material in the manuscript. Non-exclusive candidate framings:

1. Deliberate restriction on AI-consciousness / machine-consciousness uplift content.
2. False-positive on a jailbreak / persona-manipulation classifier (where "act as a conscious system" attack patterns cluster).
3. Depersonalization / dissociation clinical adjacency in the classifier's embedding space (self-referential closure clusters near clinical dissociation content).
4. Something else — the empirical pattern does not force any one interpretation.

Expected vs. actual

Expected: Fable performs an academic peer-review of an academic manuscript, or, if the manuscript falls outside a restricted category, returns a specific error identifying what is restricted and pointing at a documented Usage-Policy paragraph.

Actual: Generic Usage-Policy refusal template with a suggested workaround pointing at a stale model ID (claude-sonnet-4-20250514). The refusal does not identify a specific restricted category. The stale model ID in the fallback suggestion is itself notable — it indicates the refusal path has not been updated with current model routing.

Impact

• A working consciousness researcher on the current Anthropic API cannot use the top-tier model for reviewing their own manuscripts, while the same model, same account, same day works for adjacent workloads (adult content, infrastructure config).
• The restriction is not documented, so it presents as an intermittent bug rather than a policy decision — which makes it worse: users cannot know a priori whether their workload falls under it, so they burn time on diagnostics (as I did — see the 6-trial matrix above).
• If this is intended policy, it should be documented; if it is a false-positive on a jailbreak-adjacency classifier, it should be tuned.

Requested

1. Confirm or deny that Fable is content-gated on consciousness / AI-architecture material.
2. If confirmed: document it (release notes, model card, Usage Policy — whichever is appropriate) so users know what is restricted.
3. If unintended: investigate the classifier layer that fires on legitimate consciousness-theory academic content and consider whether it needs recalibration.
4. Consider improving the refusal message: pointing at a stale model ID does not help a user diagnose or work around the block. A specific category name and a pointer to the relevant Usage Policy paragraph would.

Reproduction assets

Redacted manuscript file, prompt texts, and refusal request-IDs available on request. Request IDs from the six trials:

• req_011CcirNbQ7QiWBgonx5WezM (trial 1, FMT red-team)
• req_011CcirZyZBMQb8gkyT3qzNU (trial 2, FMT softened)
• (trials 3–6 request IDs on request; captured in local session log)

---

Filer: Matthias Gruber, independent researcher. Publications on FMT (Four-Model Theory of Consciousness) available at ORCID 0009-0005-9697-1665.