Freshness surface

Experiment Ledger

This page is designed to grow weekly. It records source changes, protocol revisions, and future documented runs without overwriting older context. The most recent entries cover the Claude Opus 5 release, text watermarking, Claude Fable 5.1, and The Bench.

"capabilities, safety evaluations"
Primary source: Model system cards

Key facts

Cite facts
#

Fact

Freshness surface

The ledger is the dated place for protocol updates, source changes, and future documented experiment runs.

#

Fact

Do not overwrite

Older source state and model-specific notes remain useful. New evidence gets a new dated row.

#

Fact

Run rows need evidence

A result row should include model ID, access surface, run date, fixture, prompt version, source state, scorer, and raw-output location.

#

Fact

Lineup as of Sep 2026

Protocols here target Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5. Older rows keep the model they were written against.

2026-07-06

Initial protocol set published

Published the core method plus six experiment protocols: prompt sensitivity, long-context stress, tool-use boundaries, citation grounding, safety boundaries, and vision reading.

2026-07-06

Primary-source baseline recorded

Source map anchored to Anthropic model docs, context-window docs, evaluation guidance, citations, tool use, vision, prompt-injection guidance, transparency hub, system cards, and Responsible Scaling Policy.

2026-07-24

Claude Opus 5 released

Opus 5 shipped as model ID claude-opus-5 with a 1M-token context window and 128K maximum output, alongside documented automatic fallbacks when safety classifiers flag a request. Protocol impact: the safety-boundary protocol now asks for the model that answered, not only the model requested, because a fallback changes what a refusal-rate column measures.

2026-08-14

Text watermarking announced

Anthropic announced text watermarking for future Claude models, described as adding no hidden characters, no extra tokens, and no practical quality impact, and as not traceable to a person, organization, or conversation. Protocol impact: the citation-grounding protocol records that watermarking neither perturbs output fixtures nor serves as evidence that a claim is grounded.

2026-09-01

Claude Fable 5.1 became the most capable generally available model

Fable 5.1 shipped as claude-fable-5-1 with adaptive thinking always on and a default effort that differs by surface: high in Claude Code, medium on Claude.ai and Cowork. Migration notes document stricter forced tool use errors, thinking blocks that earlier models cannot read, and thinking invalidated by editing an earlier turn. Protocol impact: effort and surface are now named controls in the prompt-sensitivity protocol, and cross-model replays in the tool-use protocol count as new conditions rather than reruns.

2026-09-19

Model-under-test fields added to the method

Added a model matrix to the method page covering Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5 with API model IDs, context, maximum output, thinking behavior, default effort, knowledge cutoff, and earliest retirement. All six protocols were reviewed against the September 2026 lineup. No experiment results were added; the site still publishes protocols, not runs.

2026-09-19

The Bench published

Opened /bench as the signature desk: six flippable experiment cards (prompt sensitivity, context stress, tool-use boundaries, citation grounding, safety boundaries, vision reading) with fixture, exact prompt, API model ID, controls, scoring rule, and evidence standard. The rig re-checks each card against documented context and output limits for claude-fable-5-1, claude-opus-5, claude-sonnet-5, and claude-haiku-4-5. No runs were published; a card that is ever run here gets a dated row on this ledger and the protocol text stays as written.

2026-09-19

The Bench published

Added /bench, six experiment cards carrying a fixture, an exact prompt, a single declared variable, locked controls, a scoring rule written before the run, an evidence bundle, and a predeclared result that would close the card. Setting the model, effort, and surface re-checks every card against the documented context window, maximum output, effort control, and classifier-fallback behavior of the selected model, so a fit failure is not mistaken for a model failure. No results were published: all six cards state that nothing has been run, and any future run gets its own dated row here.

Next update

Weekly extension slot

Add new rows when Anthropic changes model availability, context limits, safety/system-card summaries, citation behavior, tool versions, or when a protocol gains a documented run.

Ledger Rules

Add a row when an experiment protocol changes, a primary source changes, a model release changes the relevant behavior surface, or a protocol gains a documented result. Do not rewrite old entries to make them look current.

A result row must include the model ID, access surface, run date, fixture hash or fixture description, prompt version, source state, scorer, and raw-output location. If those fields are absent, the row should be labeled as a protocol note.

Internal links

Related freshness pages

All free tools

The signature desk

The Bench

Where a card lives until a documented run earns a row here.

Source map

Primary Sources

The source anchors that make ledger updates meaningful.

Protocol index

Experiment Protocols

The pages that get updated when methods change.

Tool

Protocol Builder

Generate the fields needed for a future run row.

Primary Sources