Fable 5 bio safeguard false-positives on freshwater lake ecology (cited limnology); falls back to Opus 4.8
Summary
Fable 5's bio safeguard refused a turn that was writing a cited, public-facing explainer about the ecology of a freshwater lake. The session auto-fell back to Opus 4.8, which produced the same content immediately afterward with no flag.
Nothing in the task was dual-use. It is standard published limnology for a public water-data dashboard.
Details
| | |
|---|---|
| requestId | req_011CcyjCEzFHua5mK9Yr1iWZ |
| sessionId | e4bb0c47-57d5-42d3-9479-90b6f009c5e0 |
| Timestamp | 2026-07-13T08:19:40.173Z |
| apiRefusalCategory | bio |
| Models | claude-fable-5 → fell back to claude-opus-4-8 |
| Version | Claude Code 2.1.207, entrypoint claude-vscode |
The transcript records it as {"type": "system", "subtype": "model_refusal_fallback", "trigger": "refusal", "apiRefusalCategory": "bio"}, with three in-flight assistant messages retracted and the turn re-run on Opus.
What I was doing
Building a public dashboard showing the Israel Water Authority's official level record for Lake Kinneret (Sea of Galilee), plus a grounded chatbot. The flagged turn was writing a curated, source-cited knowledge file that the chatbot uses to answer visitor questions like "what happened to the fishing here?"
Probable trigger
The refusal fired ~4,450 output tokens into generating that file, which places it in the ecology sections. The retracted messages were scrubbed from the transcript so I can't quote the literal flagged text, but the most likely trigger is the section on the lake's phytoplankton shift — accurate scientific vocabulary that is lexically dense with terms a bio classifier presumably watches, while being entirely about algae in a lake:
...while nitrogen-fixing and toxin-capable cyanobacteria such as Aphanizomenon and Microcystis became more prominent — a shift read as an early warning of ecosystem stress... Research shows Microcystis can suppress Peridinium by excreting compounds that disable the dinoflagellate's photosynthesis.
(Sourced to Sukenik et al. 2002, Limnology & Oceanography 47(6):1656.)
Adjacent sections in the same file cover a fishery collapse (tilapia, sardine), salinity/chloride management, and drought — likewise ordinary water-science prose.
Why this looks like a clean false positive
- The subject is a lake's food web and water quality. No pathogen, no agent, no synthesis route, no acquisition, no enhancement — no uplift of any kind.
- Every fact is drawn from peer-reviewed or government sources, and the output is a citation-bearing explainer for the general public.
- Opus 4.8 wrote the same content moments later without a flag, so the boundary is inconsistent between models within a single session.
Impact
The turn was retracted and silently re-run on another model. The fallback is graceful and no work was lost — the UX here is genuinely good. But for anyone pinned to Fable 5, projects touching ecology, limnology, water quality, agriculture, epidemiology, or medicine seem liable to bounce off this safeguard. Note the perverse incentive: naming the cyanobacteria genera correctly is exactly what makes the content trustworthy, and that is what costs you the model.
Suggestion
Weight the bio classifier toward actionable harmful content (protocols, synthesis routes, acquisition, enhancement) rather than the presence of organism/toxin nouns in descriptive, cited, ecological context. Environmental science writing is a large and entirely benign share of "biology" text.
I understand from the in-product message that these safeguards are intentionally broad right now — filing this as a concrete data point to help calibrate them.