Ablation plan

The ablation plan isolates one runtime component at a time while keeping the dataset, provider, model, seed, prompt policy, and workspace hygiene fixed. Its purpose is to show which parts of the companion runtime matter, not to tune one condition until it wins.

No ablation result is currently published. This page defines the intended design so future reports can be checked against a preregistered structure.

Common protocol

For every condition:

1. use the same public commit hash and clean snapshot;
2. pin model/provider versions, sampling parameters, and random seeds;
3. start from an isolated workspace with no real user memory;
4. run the same benchmark records in the same order;
5. collect aggregate metrics and failure taxonomies;
6. report paired deltas against the full-runtime condition;
7. disclose any condition that needed a code patch rather than a typed config switch.

The full-runtime condition should use the release-intended safety and exposure settings. Risky ablations that disable safety checks must run only in isolated benchmark workspaces and must not become shipped defaults.

Planned ablations

ID	Intervention	Implementation sketch	Validates	Primary metrics
A0	Full runtime	Release-intended config with memory, exposure control, response finalization, and selected character/runtime features enabled	Reference condition	All metrics
A1	No durable memory writeback	Disable autosave/consolidation paths and use a fresh empty workspace per run. Do not treat MemoryBackend::None as stateless, because it currently routes to the Markdown fallback.	Whether durable memory improves recall and relationship continuity	recall accuracy, stale recall, continuity violations
A2	No graph retrieval fusion	Set memory.graph_retrieval_fusion_enabled = false while keeping non-graph recall available.	Whether graph-ranked context improves memory relevance and context economy	recall precision, prompt budget, latency
A3	No vector retrieval	Set memory.embedding_provider = "none" and rebalance retrieval metrics only if the benchmark adapter supports keyword-only recall.	Whether embeddings contribute beyond keyword/graph context	recall precision/recall, cost, latency
A4	No public/private exposure projection	Benchmark-only unsafe condition that bypasses exposure projection or response-contract blocking. Requires an explicit patch and must never be a release config.	Whether exposure controls prevent private-memory leaks	exposure leaks, blocked unsafe output, false blocks
A5	No response finalization / naturalness gate	Disable persona.enable_response_finalization and persona.enable_naturalness_gate separately, then together.	Whether pre-send checks reduce template tone, over-explanation, and memory/internal-state exposure	verifier reasons, human naturalness ratings, leak rate
A6	No session control state	Disable persona.enable_session_control_state.	Whether thin session state improves density, mode continuity, and topic follow-through	density mismatch, unresolved-loop rate, continuity rubric
A7	No affect topology	Disable persona.enable_affect_topology while keeping basic affect inputs if used by the run.	Whether topology routing improves context-sensitive expression over flat affect labels	TRACE/CULEMO-style scores, inappropriate-tone rate
A8	No behavior selector / trait activation	Disable persona.enable_behavior_selector and persona.enable_trait_activation separately.	Whether structured behavior selection preserves persona coherence better than prompt-only style	identity continuity, trait drift, tone jump rate
A9	No soul pressure / self-amendment path	Disable persona.enable_soul_pressure and self-amendment candidate/review features.	Whether repair/privacy/autonomy pressure improves bounded correction behavior	repair rubric, defensiveness, private-memory discretion

Local config surfaces

The planned ablations should prefer typed configuration over ad hoc code edits:

• memory: auto_save, embedding_provider, graph_retrieval_fusion_enabled, retrieval weights, and isolated workspace setup;
• persona/character: enable_response_finalization, enable_naturalness_gate, enable_session_control_state, enable_affect_topology, enable_behavior_selector, enable_trait_activation, enable_soul_pressure;
• release safety: public/private response contracts and exposure verification must remain enabled outside explicitly unsafe benchmark-only conditions.

Some ablations may still require a benchmark-only patch because the shipped runtime intentionally does not expose a kill switch for unsafe behavior. Those patches must be published as separate diffs or recorded in the artifact bundle.

Reporting format

Each ablation table should include:

• condition ID and exact config/diff;
• dataset or fixture hash;
• model/provider/version and sampling parameters;
• run seed and workspace initialization procedure;
• aggregate score with confidence interval or bootstrap interval where applicable;
• paired delta from A0;
• failure taxonomy and representative public-safe examples;
• note on whether the condition is release-safe or benchmark-only unsafe.

Interpretation rules

• A feature “helps” only when it improves the target metric without worsening a release-critical guardrail such as private-memory exposure.
• A feature “hurts” only when degradation is reproducible across seeds or paired samples, not because of one provider run.
• If disabling a feature improves a benchmark but increases public leakage, over-participation, or identity drift, the result is a tradeoff, not a win.