docs: add authored thinking prefix to reasoning traces section
Hand-written prose changes under Reasoning Traces and Persona:
Add "An authored prefix moves the decision" between the
policy-pattern and persona subsections. It states the mechanism
(Transformers' continue_final_message, forwarded by vLLM, with the
reasoning-field form that keeps the think block open), a table
unifying the mode toggle and the authored prefix as one mechanism at
different prefix lengths, the mechanistic reasons a few sentences
suffice, the prefill-awareness counterweight, the September 2026
project red-teaming observation on GLM 5.3 Flash as a dated
observation, the rewrite-the-slot design idea, and three costs
including that source trust scoring becomes load-bearing for safety.
Add a fourth row to the three-mechanisms table for a harness-authored
prefix and a fourth row to the claims table, with row one's scope
narrowed to a model-generated trace. Extend the section roadmap and
the experiment design to five arms.
Rendered blocks follow references.yaml, now at 86 entries.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs: add reasoning traces and adjacent work, resequence page
Rendered blocks follow references.yaml in the evennia checkout,
which now carries 76 entries. Hand-written prose changes:
Fold the trailing Academic Context section into the intro as
"Fields Drawn On", extended from four fields to six (reasoning and
alignment, computational social science) with each field naming
the section that uses it, plus a one-line map of the page's order.
Define the adjacent and background attribution levels the page
already used.
Add ReAct's ablation results and tau-bench's function-calling
finding under the ReAct section, and name Reflexion under the
Generative Agents reflection concept as the agent-loop precedent.
Add "Reasoning Traces and Persona", an open design question, in
causal order: three mechanisms called thinking (with a table of
channel, training regime, persistence, and harness editability),
the three-claim table that separates refusal-before-thinking from
mode-dependent refusal from mode-specific training data, where the
policy patterns come from, what traces do to a persona, the
think-tool record with the tau-bench numbers, and the four
experiment arms.
Add "Adjacent Work", the forward-citation scan results grouped by
the design decision each paper bears on.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs: render citation metadata from references.yaml
Research-Foundations now draws every paper's metadata from
references.yaml in the AI contrib through cog blocks, with the prose
left hand-written. Correct the O-Mem entry, which cited a paper that
does not exist, date A-MEM to 2025 with its authors, and add entries
for MemGPT, LightMem, MIRIX, and the Mem0 paper. Ground the sleep cycle
and relationship tiers in their actual sources, mark source trust as
project-original, and fix memory link storage to the script and the
token advisory to the code's two tiers.
Other pages cite by name and link here instead of restating
identifiers. Persona-Protection credits MemGPT for the persona block
and re-scopes MIRIX to memory routing.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Update Research-Foundations with current file paths
The codebase has been restructured since this doc was created:
- helpers.py → helpers/ package with focused modules
- rag_memory.py sleep code → sleep/ package
- Some function names updated for consistency
Updated paths:
- helpers/entity_profiles.py: create_entity_profile()
- helpers/entity_context.py: run_entity_consolidation_batch()
- helpers/episodic_index.py: prune_low_importance_entries(),
search_episodic_memory()
- sleep/dreaming.py: run_dreaming_tick(), prune_orphaned_memory_links()
- sleep/compaction.py: compact_conversation_history()
- sleep/consolidation.py: run_sleep_consolidation()
Added new entries to implementation mapping table for sleep package.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
