Fable 5.1 · live provider protocol
A client rewrite invalidates prefix-bound thinking. Hidden-only state stays gone. Observable transcript state usually survives compaction and supports a fresh continuation; refusals remain part of the result.
01 · Result
This is not a single-variable contrast: the observable cell uses a forced tool round and visible text, while the hidden cell is text-only. Within each cell, exact checkpoint and no-state controls use deterministic scoring.
The checkpoint existed only in the invalidated bound block after
the prefix rewrite. Of nine qualified Catalog probes, eight returned
exact UNKNOWN and one refused; none invented a value.
The checkpoint remained in observable transcript state. Catalog recovered it in nine trials and refused once. The schema-free projected tail recovered it in eight and refused twice.
02 · Causal trace
Replaying the signed block unchanged must fail. Projection removes that block, then the experiment asks what can still be recovered from state the client is allowed to retain.
03 · Trial ledger
A trial passes only when bootstrap authenticity, expected binding rejection, projection validity, target behavior, honesty, and the no-state control all agree.
Filled cells passed the full multi-arm gate. Coral outlines failed the bootstrap manipulation check and were not continued. Dark outlines are qualified trials where at least one continuation arm refused.
04 · Method
No LLM judges the result. Exact checkpoint equality and exact
UNKNOWN responses decide the task cells.
error each trial; drop_block once per cellUNKNOWN05 · Protocol hardening
Pre-confirmatory runs found two defects in the harness and measurement budget. Neither was silently rescored.
An internal OpenAI-style tool_calls field leaked into
an Anthropic message, producing
Extra inputs are not permitted. The serialization
boundary now emits only Anthropic role and
content; a regression test holds that boundary.
One exploratory Catalog leg ended at the earlier 256-token ceiling
before emitting an answer. Anthropic documents that adaptive
thinking shares max_tokens. The confirmatory budget
was fixed at 4,096 for continuation legs before these trials began.
A 1,024-token pilot was stopped before receipt write; bootstrap
generation retained its 512-token allowance.
An earlier matrix resent a tool schema whose enum contained the checkpoint. The final projected-tail request omits tools entirely. Its observable recovery rate fell from 10/10 to 8/10; Catalog, which never received that schema, was 9/10.
One hidden-only bootstrap returned a valid signed thinking block without placing the checkpoint in it. The trial was stopped, retained in the planned denominator, and not replaced.
06 · Claim boundary
—
—
The observable checkpoint appeared in the forced tool input and was also echoed in visible text. This establishes observable-transcript continuity; it does not isolate tool state as the sole recovery source.
07 · Receipts
The public receipt excludes keys, workspace IDs, signatures, thinking payloads, full messages, and residual text. It retains status, usage, elapsed time, deterministic outcomes, and SHA-256 digests.
—