Skip to content

Constitutional evolution — research compilation

Status: deferred (see ADR 0003) Sources: four exploratory notes from 2026-05-22, archived in _archived/ Purpose: distill what was explored, what was found, and why the mechanisms didn't ship

The problem

Closed-form derivations D4 and D5 (see closed_form_analysis.md) established two structural problems:

  • D4: in ATTENTION frame, constitutional bedrock with thematic overlap ≥ 0.75 cannot be beaten by any new block under current weights — mathematically inevitable
  • D5: PERMANENT tier half-life is 47.5 years; constitutional cannot evolve via decay

These compound into a long-term failure mode: as the agent's identity drifts over years, constitutional blocks ossify and dominate retrieval inappropriately.

Four mechanism families explored

Architecture M — structural separation

Idea: constitutional content is a different kind of content from task knowledge. Don't force them into one ranking pool. Filter constitutional out of ATTENTION candidates; inject as a preamble at frame render time.

Implementation cost: one-line filter, plus design for preamble injection (template, token budget, query-relevance filtering).

Simulation result (4 strategies × 5 scenarios × 365 days):

scenario          baseline   M
─────────────────────────────────
stable             85.3%    78.4%   ← −7pp tax
slow_drift         62.5%    72.4%   ← +10pp
regime_change      44.5%    77.6%   ← +33pp
quiet_burst        76.6%    77.0%   ← 0pp
adversarial        65.7%    77.3%   ← +12pp

Verdict: big help under drift, real cost under stability. Without telemetry on real-user drift distribution, shipping M is a bet that helps some users while harming others.

Model C — Darwinian identity (ego_strength mechanism)

Idea: constitutional should earn persistence through positive outcomes, not be immortal by decree. Add a per-block ego_strength that modulates effective decay rate. SELF-frame queries provide a dedicated reinforcement channel.

Implementation cost: 4 magic numbers (EGO_POS_RATE=1.0, EGO_NEG_RATE=0.3, EGO_TIME_DECAY=0.05, EGO_LAMBDA_ALPHA=0.05), 1 new schema table (block_ego_state), new vocabulary.

Simulation result:

strategy            qratio  ego_moat  hoarding (es_concentration)
─────────────────────────────────────────────────────────────────
baseline             45.0%   98.4%     n/a
Architecture M       78.2%    0.0%     n/a
Model C              75.3%    0.0%    HOARDING (9.5×)

Verdict: marginal over M alone (75.3% vs 78.2%). Adds 4 magic numbers and exhibits a clear failure mode: ONE constitutional block dominates ego_strength accumulation (concentration=9.5× across every scenario). The Critic agent's charge against FSRS-5 — "fashion, not calibration" — applies cleanly to ego_strength.

Model D — distributed feedback

Idea: Model C's hoarding is winner-take-all. Distribute SELF-query feedback across top-3 constitutional with softmax weights.

Implementation cost: one more parameter (distribute_n), modest logic.

Simulation result (vs Model C):

  • Hoarding fixed: concentration 9.5× → 3.3×
  • Quality cost: 3-7pp drop in 3 of 5 scenarios

Verdict: fixes one failure mode by introducing a quality cost. Net unclear.

Self-architecting agent — meta-layer

Idea: the four named architectures (baseline, M, Model C, Model D) are not discrete; they're settings in a 4D parameter space (attention_const_weight, ego_alpha, distribute_n, self_check_freq). Hill-climb to learn the right configuration per workload.

Implementation cost: new table (architecture_history), 5+ API methods, hill-climbing subsystem with shadow eval, regime detection, collaborative milestones — significant.

Simulation result (hill-climber starts M-like, adapts every 28 days):

scenario          baseline   M       Model C   Model D   self_architect
────────────────────────────────────────────────────────────────────────
stable             81.1%    77.3%    76.5%     76.1%     76.5%
drift              77.8%    71.9%    74.9%     75.9%     73.0%
regime_change      74.0%    76.9%    79.1%     70.0%     72.3%

The agent correctly identified the right direction for each scenario (stable → raise atw toward 1.0; drift → keep atw=0, raise distribute_n; regime change → raise ego_alpha). But it underperforms every fixed strategy in quality.

Verdict: directionally correct, quantitatively underperforming. By P2 of elfmind: a miscalibrated self-model is worse than no self-model. The hill-climber is the miscalibrated self-model.

Why all four are deferred

Common pattern: each mechanism is intellectually interesting but doesn't produce a decisively-better-than-baseline result across the simulated user mix. Shipping any of them is a bet on a workload distribution we don't yet have data for.

Specific failures of project discipline: - Model C introduces 4 magic numbers + new vocabulary — the exact shape the plan_memory_scoring.md 3-agent review rejected for FSRS-5 - Model D adds yet more parameters - Self-architecting adds an entire meta-layer that the simulation refutes - Architecture M is conceptually clean but trades one failure mode (drift) for another (stability tax)

See ADR 0003 for the formal decision.

What we keep from this research

  • The structural insight that constitutional and task content are different kinds of content — this remains true even if Architecture M doesn't ship
  • The simulation harness in scripts/longitudinal_sim/ is a permanent fixture
  • The empirical refutation of these mechanisms saves future contributors from rediscovering them

Trigger to revisit

Any of: 1. Production reports from multiple users that constitutional dominance is causing retrieval failures (currently: only Dmitry's report, addressed by v0.15.3) 2. Telemetry on real-user drift rate distribution 3. Benchmark evidence (MemoryAgentBench / LoCoMo) showing one of these mechanisms wins on agent workloads 4. Higher-confidence simulation (N≥10 seeds across more scenarios) confirming the marginal results

Archived raw notes

For full reasoning history (not required reading; the above is the distilled synthesis):

  • _archived/mc_evolution_findings.md — update-rule tweaks don't work
  • _archived/constitutional_architecture.md — Architecture M discovery
  • _archived/ego_feedback_findings.md — Model C synthesis
  • _archived/full_scenario_findings.md — cross-scenario evaluation