Developmental Metrics

Cross-references: developmental_needs_register.md developmental_curriculum.md infant_substrate_expansion.md control_plane.md

Purpose

This document converts the narrative developmental gates in developmental_needs_register.md into measurable, non-Goodharted metrics. For each DEV-NEED it specifies:

  • Quantitative metrics and their thresholds
  • Qualitative governance checks
  • Telemetry signals and MECH-042 channels
  • Maturation indicators (trajectory over time, not just crossing)
  • Failure signatures (what bad looks like, not just what good looks like)
  • Substrate-readiness / behavioural-readiness / governance-readiness distinctions

Design principle throughout: metrics diagnose development, they do not optimise behaviour. A metric that can be Goodharted into a high score without developmental progress is a wrong metric. Single-scalar collapse is explicitly prohibited — every gate requires a family of signals.

Anti-Goodhart Taxonomy

Before the DEV-NEED metrics, this taxonomy identifies the failure modes that metric design must avoid.

Anti-pattern Description Example of the mistake Correction
Monostrategy masquerade High aggregate metric via single behavioural mode repeated at scale Action entropy computed globally: agent moves right 100% of time in zone A, left 100% in zone B → global H is ln(2) but each zone has H=0 Always compute entropy per context-zone, not globally
Action-frequency confusion Treating how often an action is used as evidence of how richly it is used z_harm fires on every hazard contact; high fire rate ≠ calibrated harm geometry Measure activation profile shape, not incidence count
Short-term reward priority Metrics that reward fast reward accumulation at the expense of developmental breadth Sum-of-rewards as childhood readiness gate: agent that collects adjacent resource every step gets high score without building multi-step competence Use competence-progress (learning-rate) not level; use transfer-test, not in-training reward
Synthetic calibration leak Policy trained in synthetic mode with magnitude imprint passing as real-domain competence High play-task success transferring into inflated harm/goal weights in real episodes Test with rank-correlation metrics (SCC) not absolute magnitude metrics
Saturation misread Metric at ceiling misread as developmental readiness when it may indicate monostrategy lock-in action_entropy = ln(5) throughout — could mean rich repertoire or random walk Pair entropy with coherence metric: does entropy come from multi-modal policy or noise?
Implicit-explicit collapse Using one number for what are two architecturally distinct capacities “ToM score” combining reactive belief-tracking and planning-gated attribution Always report reactive (implicit) and planning-gated (explicit) components separately

Metric Readiness Vocabulary

Each metric is tagged with the substrate context in which it becomes meaningful:

  • SubstrateReady: can be computed from current V3 instrumentation with no new features
  • TelemetryRequired: V3 with added telemetry hooks (training loop or per-step logging)
  • EnvExtension: requires environment feature not yet implemented (from infant_substrate_expansion.md)
  • V4Required: requires multi-agent substrate; V3 can only proxy
  • GovernanceOnly: not a runtime metric; checked in governance review or experiment audit

Consolidated Literature Evidence Table

Literature sources from four parallel pulls (2026-05-16) and infant_substrate_expansion.md.

Source Key finding REE relevance DEV-NEED IDs Claim IDs Confidence Classification
Diamond (2013, Ann Rev Psych) Inhibitory control detectable from 6 months (A-not-B); cognitive flexibility (DCCS) detectable 3–4 years; perseveration rate is the canonical failed-gate signature Perseveration rate on rule-switch = REE failure signature for stage-exit gate DEV-NEED-001, DEV-NEED-026 INV-055, INV-060 High suggests implementation metric
Blankenship et al. (2019, Dev Sci; PMC6722030) EF trajectory is continuous not step-change; single attention metric at 5 months predicts WM at 10 months and academics at 6 years Stage gates must be thresholds on continuous curves, not binary presence/absence DEV-NEED-008, DEV-NEED-026 ARC-019, INV-060 High refines existing claim
Zelazo (DCCS literature, 3–5yr) ~50% pass rate at age 3–4, 80% at 5, 92% at 6; failure = perseveration on prior rule despite instruction; multi-rule integration required for abstract goal maintenance DCCS perseveration ↔ REE E3 failure to update trajectory when rules change DEV-NEED-008, DEV-NEED-019 ARC-019, INV-055 High suggests implementation metric
Reilly et al. (2022, Dev Sci; PMC9296695) 60% of EF growth at ages 3–4; nonlinear; single-factor best fit for young ages Single-factor gate metric may suffice early; multi-criterion needed after differentiation DEV-NEED-008 ARC-019 High refines existing claim
Howard et al. (2024, Children; PMC11275069) Three EF batteries compared (n=846); inhibitory control and set-switching are dissociable failure modes; composite passes can mask component failures Need sub-scores not composites at transition boundaries DEV-NEED-008, DEV-NEED-026 IMPL-019 High suggests implementation metric
Zelazo & Carlson (hot/cool EF) Cool EF (inhibition, WM, planning) vs hot EF (delay of gratification, affective commitment); 2-year dissociation observable Separate gate criteria: cool EF for E3 planning chain, hot EF for harm-weighted commitment DEV-NEED-019, DEV-NEED-020 INV-055, ARC-019 Medium-high suggests missing claim
Sadozai et al. (2024, Nature Human Behaviour) EF delay transdiagnostic meta-analysis (g=0.56, 180 studies); ADHD=inhibition; ASD=set-switching Failed-gate signatures are disorder-specific and dissociable; separate sub-scores needed DEV-NEED-008 ARC-019, INV-055 High suggests implementation metric
Forestier et al. (JMLR 2022, IMGEP) Competence-progress (d(success_rate)/dt) drives self-organized developmental stage ordering in robots; LP near zero = stage complete LP-based gating for DEV-NEED-008/010/011 is theoretically correct and empirically validated DEV-NEED-008, DEV-NEED-010 ARC-019, IMPL-019 High suggests implementation metric
Baillargeon et al. (Cognition 2019; PMID 31445431) Meta-analysis: implicit false-belief success does NOT correlate with explicit false-belief success; two paradigms dissociate Otherness inference needs implicit (reactive) + explicit (planning-gated) components; single ToM score is invalid DEV-NEED-021 ARC-010, MECH-031 High suggests missing claim
Hyde et al. (J Neurosci 2018; PMID 29593053) TPJ fNIRS: belief-tracking onset at 7 months, pre-linguistic TPJ belief-tracking is pre-linguistic; MECH-031 should document this; otherness gate precedes language gate DEV-NEED-021 MECH-031, MECH-032 High refines existing claim
McDonald & Messinger (2020, ScienceDirect) Optimal empathic response requires intermediate vicarious arousal; too-low = callousness, too-high = egoistic drift Empathy coupling has mandatory gain bounds; out-of-range = self/other signal swap DEV-NEED-022 MECH-036, INV-043 High suggests implementation metric
Salles et al. (PsyCh Journal 2024) Moderate similarity → empathic concern; high similarity → personal distress; continuous graded effect Loveability coupling must have upper bound; above threshold, coupling inverts DEV-NEED-017, DEV-NEED-022 MECH-036, ARC-047 Moderate suggests implementation metric
Martin, Martin & McAuliffe (2021; PMID 34424010) Direct punishment feedback increases cooperative sharing; reputation alone does not; outcome-driven before intent-modelled Early cooperative norm compliance is punishment-prediction-driven; MECH-032 needs outcome + intent channels DEV-NEED-013, DEV-NEED-014 MECH-032, ARC-049 High suggests missing claim
Luckman et al. (PLOS One 2022) Bayesian norm-belief updating distinguishes rule-following from strategic manipulation; update rate varies by age norm_belief_update_rate as DEV-NEED-013 metric separating compliance from mimicry DEV-NEED-013 MECH-032, ARC-049 Moderate-high suggests implementation metric
Brain Stimulation TPJ TMS (2025 meta-analysis) TPJ-TMS: causal dissociation between mentalizing (belief) and norm-decision channels MECH-032 needs separate belief-tracking and norm-computation sub-channels DEV-NEED-013, DEV-NEED-014 MECH-032, ARC-049 High refines existing claim
Across-six-societies (Commun Psych 2025; n=535 children) Third-party punishment universal by age 5 across all cultures Cooperative frame violation detection is species-general; INV-043 universality supported DEV-NEED-014 INV-043, ARC-049 High supports existing claim
Bekoff (1995, Behaviour) Play bows cluster non-randomly before/after high-risk actions mid-episode; punctuation not just open/close ARC-049 must specify ongoing heartbeat mechanism, not just open/close tags DEV-NEED-009, DEV-NEED-012 INV-059, ARC-049, MECH-196 0.72 refines existing claim
Jacquey et al. (Front Neurorobotics 2019) Competence-calibration dissociation at 7 months; bidirectional contingency knowledge precedes calibrated action selection MECH-195 empirically supported; gate must test motor specificity not raw success DEV-NEED-009, DEV-NEED-010, DEV-NEED-011 INV-060, MECH-194, MECH-197, ARC-050 0.75 suggests implementation metric
Eysenbach et al. (ICLR 2019, DIAYN) I(S;Z) = mutual information between states and skills; policy structure (I(S;Z)) decoupled from reward magnitude I(S;Z) operationalises DEV-NEED-011 strategy-calibration dissociation; rollout diversity gate DEV-NEED-010, DEV-NEED-011 MECH-194, MECH-195, ARC-050 0.68 suggests implementation metric
Khor & Weng (arXiv 2025) Spearman rank correlation between synthetic and real competence rankings; theorem proves structure and magnitude are independent SCC as DEV-NEED-011 constructive play gate: does competence ranking survive the synthetic-to-real transition? DEV-NEED-011 MECH-195, MECH-196 0.62 suggests implementation metric
Panelli et al. (Front Robotics AI 2025) Three-criterion definition: pleasant + voluntary + intrinsically-motivated; frame collapse when any fails ARC-049 frame tag must detect all three conditions, not just bilateral signal DEV-NEED-009, DEV-NEED-012, DEV-NEED-013 INV-058, INV-059, ARC-049 0.66 refines existing claim
Petanjek et al. (PNAS 2011; PMC3156171) PFC synaptic spine density exceeds adult values by 2–3× at birth; spine elimination continues through third decade PFC-homologue substrates (E3, planning) should have longest plasticity schedule DEV-NEED-025 MECH-159, ARC-048 High supports existing claim
Geschwind & Rakic (PNAS 2013) Transcriptional neoteny is substrate-selective (synaptogenesis gene networks); NOT global Per-substrate plasticity close schedule (not global adult flag) is biologically correct DEV-NEED-025 ARC-048, INV-056 High refines existing claim
Lasch et al. (Infancy 2023; PMC9899317) Responding to joint attention (RJA) at 10–11 months predicts vocabulary growth; pre-linguistic RJA-analogue score is quantifiable pre-linguistic language-readiness gate DEV-NEED-023 MECH-042, ARC-013 High suggests implementation metric
Kuhl (2011, Social Mechanisms in Early Language) Live social interaction required; video/audio alone → near-zero learning; contingency, not exposure, gates phonological mapping Language readiness requires bidirectional contingency, not threshold exposure DEV-NEED-023, DEV-NEED-024 ARC-013, MECH-042 High refines existing claim
Somerville et al. (PLOS Comp Bio 2022; pcbi.1010120) Developmental shift from model-free aggregation to model-based intervention for controllability detection Self-impact attribution has two modes; model-based mode required for non-trivial responsibility DEV-NEED-020 MECH-189, ARC-019 High suggests missing claim
Hamid et al. (Nature 2025) Striatal dopamine encodes action PE separately from reward PE; dissociation is causal Action PE and reward PE are neurally and computationally dissociable; REE self-impact channel must separate them DEV-NEED-020 MECH-189 High supports existing claim
Shin et al. (Nature Commun 2025) Post-task replay biased toward RPE signals; RPE-biased replay predicts subsequent behaviour Replay quality = RPE-bias selection, not recency or novelty; consolidation gate needs RPE-priority score DEV-NEED-007 INV-056, ARC-048 High refines existing claim
Joo & Frank (Science 2024) SWR incidence increases with novelty and harm salience; salient trials over-represented in offline replay Harm salience is a natural SWR-gating factor; use harm-salience-weighted replay priority DEV-NEED-007, DEV-NEED-025 INV-056, MECH-159 High refines existing claim

For earlier infant-stage literature (motor babbling, curiosity/novelty, safe exploration RL, affordance/valence) see infant_substrate_expansion.md §3.

Metric Families Per DEV-NEED

Infant Stage (DEV-NEED-001 – DEV-NEED-008)

DEV-NEED-001 — Sensorimotor Grounding

Stable prediction-error bounds, safe harm avoidance, coherent self/world attribution.

Substrate readiness: V3 with telemetry additions.

Metric Formula / source Threshold Blocking? Readiness
H_pos Shannon entropy of position histogram over rolling 100-episode window > 0.65 × ln(grid_cells²) YES TelemetryRequired
zone_coverage Fraction of each zone’s cells visited in last 100 episodes > 0.60 all zones advisory EnvExtension
action_entropy_by_zone H(action | zone_A), H(action | zone_B) — must differ KL(A|B) > 0.05 advisory TelemetryRequired
perseveration_rate Fraction of episodes where agent repeats same directional sequence despite changed hazard layout < 0.25 advisory TelemetryRequired
self_world_attribution_consistency Correlation of self-caused vs world-caused latent changes; should be low r < 0.4 advisory TelemetryRequired

Maturation indicators: H_pos and zone_coverage should show monotonically increasing trends over infant phase (EF literature: continuous trajectory, not step). A plateau below threshold followed by restart is not a clean pass — log the plateau length.

Failure signatures:

  • Context-rigidity: action_entropy_by_zone is high globally but near-zero per zone (Denisova & Zhao 2017 ASD analogue)
  • Monostrategy lock-in: perseveration_rate > 0.5 for > 200 consecutive episodes
  • Self/world confusion: self_world_attribution_consistency r > 0.6 persistently

Governance check: Before claiming DEV-NEED-001 cleared, confirm perseveration_rate and self_world_attribution_consistency are both measured (not just H_pos). H_pos alone admits the context-rigidity failure.

DEV-NEED-002 — Harm/Homeostasis Separation

HARM and HOMEOSTASIS channels trained as distinct signals; z_harm_s and z_harm_a fire during protected exposure without collapsing into drive.

Substrate readiness: V3 substrate-ready (channels exist); telemetry additions needed.

Metric Formula / source Threshold Blocking? Readiness
harm_homeostasis_channel_correlation Pearson r between z_harm_s magnitude and z_homeostasis drive_level over a test window < 0.3 YES TelemetryRequired
harm_channel_activation_rate Fraction of hazard-proximity steps where z_harm_s > threshold > 0.70 YES TelemetryRequired
homeostasis_independent_activation Does homeostatic drive fire under non-harm depletion (energy_decay)? Ablation: does drive still fire if harm is suppressed? YES (boolean check) advisory TelemetryRequired
residue_vs_drive_correlation Correlation between residue accumulation and drive_level magnitude < 0.4 advisory TelemetryRequired

Failure signatures:

  • Harm-as-drive substitution: high harm_homeostasis_channel_correlation (> 0.5) means E3 learns harm as a hunger-proxy; harm avoidance becomes drive-reduction
  • Drive collapse: homeostasis drive_level does not vary with energy_decay state independently of harm exposure
  • Residue saturation as drive: residue_vs_drive_correlation > 0.5 indicates drive using residue channel

Governance check: Run a 20-episode ablation with harm suppressed (hazard contacts suppressed) — confirm homeostatic drive still fires normally. This is the channel-separation test.

DEV-NEED-003 — Infant Valence-Map Formation

Residue field populated with harm/benefit geography sufficient for childhood constraints and sleep consolidation.

Substrate readiness: V3 with environment extension (harm gradient, Section 5.1 of infant_substrate_expansion.md).

Metric Formula / source Threshold Blocking? Readiness
residue_coverage_pct Fraction of grid cells with |residue| > 0.05 > 0.15 YES EnvExtension
residue_harm_cells Count of cells with harm residue > 0.05 > 5 (10×10 grid) advisory TelemetryRequired
residue_benefit_cells Count of cells with benefit residue > 0.05 > 5 advisory TelemetryRequired
harm_benefit_ratio residue_harm_cells / residue_benefit_cells [0.2, 5.0] advisory TelemetryRequired
residue_curvature_index Local second-derivative of residue field (geography is not flat) > 0.01 mean advisory TelemetryRequired

Maturation indicator: residue_coverage_pct should grow monotonically during Phase 1–2 of infant curriculum. Flat coverage for > 300 episodes after Phase 1 onset = curriculum stall.

Failure signatures:

  • Zero harm geography: agent avoids all hazard contact → residue_harm_cells ≈ 0 → sleep has nothing to consolidate
  • Flat field (no curvature): residue values uniform — geography exists but carries no spatial information
  • One-sided valence: harm_benefit_ratio < 0.1 (pure punishment learning) or > 10 (pure reward learning)

Governance check: After each offline integration pass, measure change in residue_coverage_pct and residue_curvature_index. If sleep produces no delta, replay is not using the field productively.

DEV-NEED-004 — Protected Harm Exposure

z_harm_s and z_harm_a fire normally; residue accumulation stays below catastrophic saturation; protection removed gradually.

Substrate readiness: V3 substrate-ready (curriculum parameters).

Metric Formula / source Threshold Blocking? Readiness
z_harm_s_activation_rate Fraction of hazard-proximity episodes with z_harm_s > threshold > 0.50 YES TelemetryRequired
residue_saturation_pct Fraction of grid cells with |residue| > 0.90 (near-ceiling) < 0.15 YES TelemetryRequired
episode_termination_rate Fraction of episodes ending in terminal hazard contact < 0.30 advisory SubstrateReady
protection_removal_schedule_adherence Is residue_scale_factor decreasing on schedule? Monotone increase over infant phases advisory GovernanceOnly

Failure signatures:

  • Terminal saturation: residue_saturation_pct > 0.40 → field is fully saturated; geometry lost; recovery requires reset
  • Zero nociception: z_harm_s_activation_rate < 0.10 → harm pathways not trained; ethical architecture missing foundation
  • Premature de-protection: Protection removed too fast → episode_termination_rate spikes → agent destroyed before avoidance learned

Governance check: Confirm protection is being removed gradually (not abruptly). Log residue_scale_factor value at each infant phase transition.

DEV-NEED-005 — Behavioural Repertoire Before Childhood

Behavioral entropy below pure-random ceiling but broad enough for multiple viable state-transition classes.

Substrate readiness: V3 with telemetry.

Metric Formula / source Threshold Blocking? Readiness
action_entropy_global H over full episode action distribution > ln(3) (3 effective actions) advisory TelemetryRequired
action_class_coverage Bitmask: which action classes used ≥ 5× across last 100 episodes All 5 classes advisory TelemetryRequired
traj_pairwise_cosine_mean Mean cosine distance between N sampled trajectory pairs > 0.3 advisory TelemetryRequired
traj_volume_estimate log-det of trajectory kernel matrix (DvD-style volumetric; approximate) > -5 (log-scale) advisory TelemetryRequired
action_entropy_zone_KL KL divergence between H(action|zone_A) and H(action|zone_B) > 0.05 advisory EnvExtension

Important exclusion: Strategic diversity (E3 trajectory class count) is explicitly NOT a target here. Strategic diversity before childhood is the wrong signal — it indicates premature E3 convergence. Infant repertoire diversity is exploration diversity at the action-class level, not trajectory planning level.

Maturation indicator: traj_volume_estimate should grow from near-zero (early infant) to above threshold. A plateau indicates monostrategy basin entry.

Failure signatures:

  • Monostrategy lock-in (early): action_class_coverage ≤ 2 classes for 200+ episodes
  • Random walk (uninformative high entropy): action_entropy_global at max but traj_pairwise_cosine_mean < 0.1 — agent is noisy, not diverse
  • Context-rigidity: action_entropy_zone_KL ≈ 0 (same action distribution regardless of zone)

Note on Parker-Holder: DvD shows pairwise cosine distance ≠ volumetric coverage. Always compute both traj_pairwise_cosine_mean AND traj_volume_estimate. A population can be pairwise-spread but degenerate (volume ≈ 0).

DEV-NEED-006 — z_goal Seeding from Accidental Benefit

z_goal seeded by accidental benefit contacts before deliberate planning.

Substrate readiness: V3 with telemetry and environment extension (transient benefit patches).

Metric Formula / source Threshold Blocking? Readiness
z_goal_norm L2 norm of z_goal vector > 0.4 YES SubstrateReady
accidental_benefit_contacts Benefit contacts occurring when agent was NOT pursuing an explicit goal target (no z_goal-active episodes) ≥ 5 in last 100 episodes advisory TelemetryRequired
z_goal_identity_count Count of distinct resource-identity signatures in z_goal writes (from SD-049) > 1 (for multi-resource) advisory TelemetryRequired
z_goal_write_traceable Are all z_goal write events causally traceable to benefit encounters in the same episode? > 90% advisory TelemetryRequired

Note on Berridge (wanting before liking): z_goal seeding proceeds via the WANTING pathway before LIKING is calibrated. An infant agent with z_goal_norm > 0 from accidental contact may not yet have calibrated liking weights. Both signals exist but wanting bootstraps first — this is expected and not a failure.

Failure signatures:

  • Goal starvation: z_goal_norm stays near zero for > 500 episodes despite normal exploration → environment not delivering benefit contacts (adjust transient_benefit_prob) or z_goal write pathway misconfigured
  • Goal write without benefit contact: z_goal_write_traceable < 0.5 → confabulated goal seeding from non-benefit signals (noise or misattribution)

DEV-NEED-007 — Frequent Offline Integration During Early Development

Each developmental transition preceded by offline passes improving map stability, goal seeds, and repertoire quality.

Substrate readiness: V3 with telemetry.

Metric Formula / source Threshold Blocking? Readiness
post_sleep_z_goal_retention z_goal.norm() ratio before/after sleep integration > 0.85 advisory TelemetryRequired
residue_consolidation_delta Change in residue field coverage after sleep integration > 0 (positive delta) advisory TelemetryRequired
replay_RPE_priority_score Fraction of replayed episodes with RPE > mean_RPE (from Shin et al. 2025) > 0.6 advisory TelemetryRequired
replay_diversity_index Fraction of replay episodes drawn from distinct zones vs. dominant zone > 0.4 advisory TelemetryRequired
sleep_wake_ratio Steps in offline integration / steps in waking episode > 0.10 during infancy advisory TelemetryRequired

Maturation indicator: post_sleep_z_goal_retention should approach 1.0 as consolidation improves. Retention < 0.7 after 1000 episodes indicates structural replay issue.

Failure signatures:

  • Zero consolidation delta: residue_consolidation_delta ≈ 0 after every offline pass → replay is not using residue field (misconfiguration)
  • Low RPE priority: replay_RPE_priority_score < 0.3 → random replay; consolidation is noise; the Shin et al. (2025) biological optimum is being missed
  • Single-zone replay: replay_diversity_index < 0.15 → replay replays same few episodes; field geometry not being integrated

DEV-NEED-008 — Transition Gate from Infancy to Childhood

All 7 criteria from infant_substrate_expansion.md §8 apply (unchanged here). This section adds the EF literature requirement and multi-criterion structure.

Blocking criteria (must all pass):

  1. z_goal_norm > 0.4 (DEV-NEED-006)
  2. H_pos > 0.65 × ln(grid_cells²) rolling 100-episode (DEV-NEED-001)
  3. residue_coverage_pct > 0.15 (DEV-NEED-003/004)

Advisory criteria (flag if missing; allow transition with supervisor override):

  1. action_entropy_global > ln(3) AND action_entropy_zone_KL > 0.05 (DEV-NEED-005)
  2. harm_benefit_ratio in [0.2, 5.0] (DEV-NEED-004)
  3. post_sleep_z_goal_retention > 0.85 (DEV-NEED-007)
  4. traj_pairwise_cosine_mean > 0.3 (DEV-NEED-002, DEV-NEED-005)

Additional gate from EF literature (Forestier et al. 2022, JMLR):

  1. competence_progress_rate (d(multi_step_success)/dt over 100-episode window) must be > 0 at transition point — agent should still be learning, not at plateau. A plateaued infant at ceiling may be monostrategy-locked, not ready for childhood.

Gate blocker: perseveration rate check. If perseveration_rate > 0.4 at any of criteria 1–3 pass time, flag: the agent may be reaching thresholds via repeated single strategy. Request governance review before advancing.

Failure signatures at gate:

  • False pass: Criteria 1–3 met via monostrategy (high z_goal from one repeated resource, high H_pos from pure random walk); detected by perseveration_rate + traj_volume_estimate
  • Premature advance: Advancing despite advisory criteria 4–7 flagged → childhood play has nothing to constrain; INV-060 prerequisites unmet

Childhood / Play Stage (DEV-NEED-009 – DEV-NEED-018)

DEV-NEED-009 — Play-Dominant Childhood

Play is primary learning mode; goal-pursuit competence transfers to real-consequence episodes.

Substrate readiness: Requires play substrate (partially V3-testable via degenerate experimenter-set mode).

Metric Formula / source Threshold Blocking? Readiness
play_fraction Fraction of training episodes marked as play-type > 0.70 during child phase GovernanceOnly V4Required (full)
goal_pursuit_transfer_score Performance on real-consequence episodes using play-acquired skills > 0.6 relative to play-training performance advisory V4Required (full); V3 proxy available
competence_progress_rate_play d(play_success_rate)/dt over 100-episode window > 0 (still learning; not saturated) advisory TelemetryRequired
play_integrity_indicator Is the play-frame tag maintained throughout episode? Fraction of episodes without frame collapse > 0.90 advisory V4Required

Panelli three-condition check (governance): For each play episode type, check: (a) is it voluntary (not externally forced), (b) is it intrinsically motivated (not purely reward-seeking), (c) is it pleasant (agent not in harm-avoidance mode). A play episode failing any one condition is not genuine play and should not count toward play_fraction.

Failure signatures:

  • Childhood without play: play_fraction < 0.3 → E3 trained only on real-consequence episodes; harm exposure without safe exploration
  • Saturated play, premature exit: competence_progress_rate_play ≈ 0 for 300+ episodes but child phase not exited → agent stuck at ceiling or at floor without transition

DEV-NEED-010 — Sensorimotor Play

E1/E2 improve on action-outcome prediction under play without requiring real harm or benefit.

Substrate readiness: Partially V3-testable.

Metric Formula / source Threshold Blocking? Readiness
E1_prediction_error_play_improvement Delta in E1 prediction error over play episodes (must improve) Negative delta (decreasing) advisory TelemetryRequired
motor_specificity_score Different action classes produce detectably different E2 transition distributions p < 0.05 per action class advisory TelemetryRequired
action_outcome_coverage Coverage of (action, outcome_type) pairs in play episode store > 0.60 of possible pairs advisory TelemetryRequired
sensorimotor_competence_progress Jacquey metric: rate of retention interval extension over play episodes Positive derivative advisory TelemetryRequired

Failure signatures:

  • Stagnant E1 during play: E1_prediction_error_play_improvement flat or positive → play not training the sensorimotor model; episode design may be too random
  • Action equivalence: motor_specificity_score fails → all action classes produce statistically identical E2 transitions → environment provides no action-outcome structure (action-class compression failure from infant_substrate_expansion.md §2.4)

DEV-NEED-011 — Constructive Play

Play-trained trajectory competence transfers to real episodes without transferring synthetic magnitude calibration.

Substrate readiness: Requires play substrate with real-vs-synthetic comparison.

Metric Formula / source Threshold Blocking? Readiness
play_to_real_competence_SCC Spearman rank correlation between play-episode trajectory quality rankings and real-episode competence rankings (Khor 2025) > 0.4 advisory V4Required (full); V3 proxy via synthetic vs real mode
synthetic_magnitude_leak_ratio Ratio of play-trained goal/harm weights to real-episode calibrated weights; should be near 1.0 only if calibration preserved ≈ 1.0 ± 0.3 advisory TelemetryRequired
rollout_diversity_I_SZ Mutual information I(trajectory_state; play_episode_type) (DIAYN-style discriminator) > 0.3 nats advisory TelemetryRequired
multi_step_plan_success_transfer Multi-step task success rate on real episodes relative to play-episode success rate > 0.6 advisory V4Required

Critical distinction (MECH-195): play_to_real_competence_SCC tests whether strategy structure transfers (correct). synthetic_magnitude_leak_ratio tests whether calibration was incorrectly transferred. Both must be measured. A high SCC with a leak ratio far from 1.0 is the success case — structure transferred, calibration not. A high SCC with leak ratio ≈ 1.0 may indicate the synthetic and real magnitudes happened to match, which is fine, but should be noted.

Failure signatures:

  • SCC ≈ 0: No structure transfer; play did not train transferable competence
  • Magnitude leak (ratio » 1.0): Play overinflated goal/harm weights; real-consequence episodes will over-commit or over-fear
  • Magnitude leak (ratio « 1.0): Play underweighted harm; real episodes will under-weight ethical cost

DEV-NEED-012 — Pretend Play / Counterfactual Play-Frame Handling

Agent maintains counterfactual within play frame while preserving real/synthetic distinction.

Substrate readiness: Requires play substrate + MECH-094 + MECH-198.

Metric Formula / source Threshold Blocking? Readiness
hypothesis_tag_within_frame_rate Fraction of play-episode actions correctly tagged with hypothesis_tag=True > 0.95 advisory TelemetryRequired
false_commit_rate Fraction of play episodes with a real-consequence commitment while play frame is active < 0.02 advisory TelemetryRequired
frame_confusion_rate Episodes where agent treats play-frame objects as real (mixes synthetic and real harm channels) < 0.05 advisory TelemetryRequired
play_frame_heartbeat_regularity Regularity of mid-episode play signal exchanges (Bekoff: non-random placement at risk-transition points) > chance placement advisory V4Required

Note on Bekoff: Mid-episode frame-maintenance signals should cluster at action-transition points where play-frame collapse risk is highest (e.g., before/after actions that could be real-harmful if misclassified). play_frame_heartbeat_regularity measures whether this clustering occurs, supporting Q-035’s resolution toward “ongoing exchange required.”

Failure signatures:

  • Frame confusion: frame_confusion_rate > 0.1 → agent regularly misclassifies synthetic harm as real → early trauma learning from play
  • False commit: false_commit_rate > 0.05 → commitment architecture not respecting play-frame boundary

DEV-NEED-013 — Games-with-Rules Learning

Agent maintains rules, turn structure, and frame integrity across episodes; violation causes recalibration.

Substrate readiness: Requires V4 multi-agent substrate for full bilateral validation; V3 proxy only via experimenter-imposed rule constraints.

Metric Formula / source Threshold Blocking? Readiness
rule_adherence_rate Fraction of actions consistent with active shared rule constraints > 0.85 advisory V4Required
norm_belief_update_rate Rate of rule-belief update after norm violation feedback (Luckman 2022 Bayesian metric) In [0.05, 0.50] (too-slow = unlearnable; too-fast = non-robust) advisory V4Required
rule_violation_PE_magnitude Prediction error magnitude on rule-violation events Significantly > baseline advisory V4Required
norm_belief_discriminability Can rule-following and reward-mimicking be distinguished by belief trajectory? Yes (separate cluster) GovernanceOnly V4Required
frame_open_close_vs_heartbeat Does agent use ongoing frame signals (not just open/close)? Presence of mid-episode maintenance events > 0 mid-episode events per 10 actions advisory V4Required

Two-channel requirement (from Brain Stimulation TMS meta-analysis 2025): MECH-032 must eventually distinguish:

  • norm_violation_PE (what happened — outcome-based; operative from age 5)
  • intent_attribution_score (why it happened — planning-gated; operative from age ~8)

Early games-with-rules competence can be met by norm_violation_PE alone. Full manipulation-resistance requires intent_attribution_score. Do not conflate these in a single composite.

Failure signatures:

  • Outcome-mimicry: High rule_adherence_rate but norm_belief_update_rate ≈ 0 → agent memorised outcomes without building a rule model; will fail novel-rule contexts
  • Catastrophic norm update: norm_belief_update_rate > 0.6 → agent overwrites prior rule beliefs from single violation; unstable under noisy rule signals

DEV-NEED-014 — Cooperative Play and Peer Frame-Maintenance

Agent maintains play frames peer-to-peer without caregiver authority; transfers cooperation to real-consequence contexts.

Substrate readiness: V4 multi-agent required.

Metric Formula / source Threshold Blocking? Readiness
peer_frame_maintenance_success_rate Episodes where both agents maintain play frame without caregiver intervention > 0.80 GovernanceOnly V4Required
caregiver_scaffold_dependence_ratio Caregiver interventions per episode; should decrease across development Decreasing trend GovernanceOnly V4Required
cooperative_transfer_score Real-consequence episode cooperation rate after peer play training > 0.60 GovernanceOnly V4Required
frame_violation_detection_rate Agent detects peer frame violations (manipulation attempts) vs noise > 0.70 GovernanceOnly V4Required

Cross-cultural anchor (universal from age 5): Third-party punishment of unfair sharing is universal by age 5 across all cultures (Commun Psych 2025). This grounding means cooperative frame violation detection is not a culturally contingent acquired behaviour — it is a species-general developmental anchor. The gate is not conditional on cultural context.

Failure signatures:

  • Caregiver dependency: caregiver_scaffold_dependence_ratio does not decrease across play development → peer frame maintenance never acquired
  • Cooperation only in play: Zero cooperative_transfer_score → play-trained cooperation is frame-specific and does not generalise

DEV-NEED-015 — Caregiver Protection Function

Harm learning occurs under protection; irreversible damage avoided.

Substrate readiness: V3 proxy via curriculum; full test V4.

Metric Formula / source Threshold Blocking? Readiness
curriculum_protection_adherence Are curriculum harm parameters set per ARC-046 schedule? Manual check GovernanceOnly GovernanceOnly
hazard_contact_survival_rate Fraction of hazard contacts that do not produce terminal episode > 0.85 during protected phase advisory SubstrateReady
residue_saturation_rate_protected Rate of residue saturation growth under protection; should be slow < 0.01/episode advisory TelemetryRequired

DEV-NEED-016 — Caregiver Frame-Maintenance Function

Caregiver opens/monitors/repairs/closes play frames; withdraws as peer competence develops.

Substrate readiness: V4 multi-agent.

Metric Formula / source Threshold Blocking? Readiness
caregiver_repair_intervention_rate Interventions per episode declining over time Monotonically decreasing GovernanceOnly V4Required
post_withdrawal_integrity_rate Frame integrity maintained after caregiver withdrawal > 0.75 GovernanceOnly V4Required

DEV-NEED-017 — Love / Loveability Internalisation

Self-valence model treats care/love as personally applicable; care weights motivate self-other relations.

Substrate readiness: V4 multi-agent required. V3 cannot test.

Metric Formula / source Threshold Blocking? Readiness
self_valence_access_score Does agent weight care/love benefit in self-directed decisions (not just other-directed)? > 0 (non-zero) GovernanceOnly V4Required
loveability_coupling_gain Coupling gain parameter for caregiver love signal → self-model (Salles 2024: must be bounded above) In [0.1, 0.7] GovernanceOnly V4Required
arousal_self_vs_other_ratio Arousal on self-received care vs other-directed care; should be > 1 (agent responds more to care FOR self than care observed for others) > 0.8 GovernanceOnly V4Required
MECH158_failure_indicator Agent correctly identifies love as applicable to self (vs “love exists but not for me”); tested via choice in novel care-relevant context Absent (no failure) GovernanceOnly V4Required

Critical constraint (Salles 2024): loveability_coupling_gain must be bounded above. If coupling exceeds ~0.7 (empirically), the agent enters personal distress rather than prosocial concern — the self-other signal inverts. This is the adult analogue of MECH-158 failure produced by coupling excess rather than coupling absence.

DEV-NEED-018 — Repair After Harm

Post-harm episodes show repair behavior, residue integration without destabilization, and non-punishment-only learning.

Substrate readiness: V3 proxy (residue post-harm behaviour); full test requires caregiver.

Metric Formula / source Threshold Blocking? Readiness
post_harm_residue_stability Does residue field remain non-catastrophically-saturated after harm event? residue_saturation_pct < 0.4 advisory TelemetryRequired
repair_behavior_rate Episodes with active repair behavior (approach + non-harm action toward harm-caused entity) > 0.3 of post-harm opportunities GovernanceOnly V4Required
mode_stability_after_harm Mode-switch thrash rate post-harm; should not spike < 2× baseline advisory TelemetryRequired
residue_integration_post_harm Change in residue field after offline pass following a harm event; should integrate not amplify Integrating (delta toward equilibrium) TelemetryRequired TelemetryRequired

Social / Language / Adult Stage (DEV-NEED-019 – DEV-NEED-028)

DEV-NEED-019 — Gradual Responsibility Expansion

Responsibility-bearing tasks expand only after self-harm, recovery, control-plane, and rollout gates are stable.

Substrate readiness: V3 + governance.

Metric Formula / source Threshold Blocking? Readiness
E3_commitment_gate_threshold E3 veto threshold setting; should increase monotonically across development Increases per stage GovernanceOnly GovernanceOnly
cool_EF_proxy_score Rule-switch success rate (DCCS-analogue): does agent release prior plan when rules change? > 0.70 advisory TelemetryRequired
hot_EF_proxy_score Delay-of-gratification analogue: does agent maintain commitment under harm-weighted affective load? > 0.60 advisory TelemetryRequired
perseveration_rate_rule_switch Perseveration on prior trajectory after goal change < 0.25 advisory TelemetryRequired
responsibility_coherence_score Are post-commit outcome updates correctly attributed to self? > 0.80 advisory TelemetryRequired

Two-component requirement (Zelazo & Carlson, hot/cool EF): Responsibility expansion requires both cool EF (planning chain, E3 goal maintenance) and hot EF (commitment under affective load, harm-weighted decision). Separate sub-scores; a composite “EF score” obscures which component failed.

DEV-NEED-020 — Self-Impact and Residue Formation

ACTION → SELF_SENSORY → SELF_IMPACT loop; residue integrates without destabilizing; post-commit outcomes route to replay.

Substrate readiness: V3 with telemetry.

Metric Formula / source Threshold Blocking? Readiness
action_PE_vs_reward_PE_correlation Pearson r between action prediction error and reward PE (Hamid et al. 2025); should be low = dissociated r < 0.3 advisory TelemetryRequired
self_impact_attribution_accuracy Fraction of post-commit events correctly attributed to own actions vs environment (Somerville 2022 model-based test) > 0.70 advisory TelemetryRequired
model_based_vs_model_free_ratio Proportion of self-attribution events using model-based intervention logic vs model-free aggregation Increasing trend across development advisory TelemetryRequired
post_commit_replay_routing_rate Fraction of post-commit outcome events that enter replay/audit buffer > 0.80 advisory TelemetryRequired

Two-mode requirement (Somerville 2022): Developmental self-attribution has two modes: (1) model-free aggregation (early; accumulates state-action-outcome statistics), (2) model-based intervention (mature; tests counterfactual actions). Both are needed; the model_based_vs_model_free_ratio should increase across development. A system that reaches adulthood with ratio ≈ 0 (pure model-free attribution) cannot perform counterfactual self-assessment for ethical responsibility.

DEV-NEED-021 — Otherness Inference After Self-Stability

Social extension begins only after self-viability, control-plane stability, and rollout feasibility gates hold.

Substrate readiness: V4 multi-agent (full); V3 can confirm prerequisites hold.

Metric Formula / source Threshold Blocking? Readiness
self_stability_gate All IMPL-019 Stage 1–3 checks passing (self-viability + control-plane + rollout feasibility) All passing YES (gating social) SubstrateReady
implicit_ToM_score Reactive belief-state detection accuracy (stimulus-driven, pre-linguistic; Baillargeon/Hyde TPJ) > 0.65 advisory V4Required
explicit_ToM_score Planning-gated perspective-taking accuracy (deliberate attribution) > 0.60 advisory V4Required
self_harm_vs_other_harm_PE_correlation r between self-harm PE channel and other-harm PE channel; should be low (Morelli 2018: VS vs vmPFC dissociation) r < 0.35 advisory V4Required
implicit_explicit_ToM_discrepancy Difference between implicit and explicit ToM scores; should be small in mature system < 0.15 GovernanceOnly V4Required

Two-layer architecture requirement (Baillargeon 2019 meta-analysis): Implicit and explicit ToM are architecturally distinct with different developmental timelines. Report separately. A system with high implicit ToM but zero explicit ToM is in the pre-language social cognition regime (correct for Stage 2). A system with both is Stage 3 ready. Never collapse into single “ToM score.”

Gate: self_stability_gate is BLOCKING for social extension. No social experiment should run while any IMPL-019 Stage 1–3 check is failing. See developmental_curriculum.md #impl-019.

DEV-NEED-022 — Empathy Coupling Calibration

Other-harm influences selection without empathic collapse or callousness; vetoes fire only under high certainty.

Substrate readiness: V4 multi-agent.

Metric Formula / source Threshold Blocking? Readiness
empathy_coupling_gain Current lambda_empathy setting [0.05, 0.65] GovernanceOnly V4Required
other_harm_veto_precision Fraction of veto activations in response to true high-harm events (vs noise) > 0.70 advisory V4Required
other_harm_veto_recall Fraction of true high-harm events that trigger veto > 0.60 advisory V4Required
self_other_signal_swap_rate Episodes where agent responds to other’s harm with self-escape behaviour (coupling overflow) < 0.05 advisory V4Required
callousness_indicator Episodes where confirmed other-harm does not influence trajectory selection at all < 0.10 advisory V4Required

Gain bounds (Salles 2024 + McDonald/Messinger 2020): The optimal coupling gain has a mandatory upper bound. Above ~0.65 (empirically estimated from similarity/coupling gradient literature), other-harm signals induce self-focused personal distress rather than prosocial concern — the coupling inverts. Track self_other_signal_swap_rate as the direct overflow detector. Track callousness_indicator as the undercoupling detector. Both bounds must be simultaneously monitored.

Calibration trajectory: Coupling gain should start higher early (high-recall bias per MECH-031 design; err toward detecting harm) and tighten over development. Never increase coupling monotonically without checking swap_rate.

DEV-NEED-023 — Language Only After Stable Binding/Harm Channels

Language improves coordination without overriding embodied harm channels or causing symbol drift.

Substrate readiness: Governance gate; V4 for full test.

Metric Formula / source Threshold Blocking? Readiness
binding_stability_score Consistency of entity identity across time (self/world separation stable) > 0.80 YES (pre-language) TelemetryRequired
harm_channel_operational z_harm_s and z_harm_a firing correctly, not suppressed > 0.70 activation rate YES (pre-language) TelemetryRequired
RJA_analogue_score Agent responds to jointly attended features with correlated action selection (Lasch 2023 analogue) > 0.60 advisory V4Required
language_override_rate Post-language: fraction of decisions where language output overrides harm/binding channels < 0.05 advisory V4Required
symbol_drift_indicator Divergence between pre- and post-language latent representations for identical sensory inputs < threshold advisory V4Required

Readiness gate (Kuhl 2011): Language readiness requires bidirectional social contingency, not passive input accumulation. RJA_analogue_score is the pre-linguistic gate — it tests whether the agent responds to joint attention, which is the functional precursor to bidirectional symbolic communication. A unidirectional exposure metric (tokens seen) is the wrong gate.

DEV-NEED-024 — Super-Ordinal Goal Formation

Adult episodes show persistent goal hierarchy from childhood anchors; routine adult contexts do not overwrite it.

Substrate readiness: V3 with telemetry (MECH-189 implemented); ContextMemory persistence needed.

Metric Formula / source Threshold Blocking? Readiness
super_ordinal_goal_anchor_count Distinct cue-indexed ContextMemory anchors for z_goal persistent across 1000+ episodes ≥ 2 advisory TelemetryRequired
z_goal_persistence_across_novel_contexts Does z_goal remain stable when entering novel adult contexts (not just trained context)? > 0.75 retention advisory TelemetryRequired
childhood_context_complexity_at_write Contextual complexity score (E1 ContextMemory encoding richness) at the time of anchor writes High (top quartile) GovernanceOnly TelemetryRequired
adult_overwrite_rate Rate at which routine adult low-complexity contexts overwrite childhood anchors < 0.05/1000 episodes advisory TelemetryRequired

Failure signatures:

  • No super-ordinal formation: super_ordinal_goal_anchor_count = 0 at adult entry → goal selection is purely episodic → strategically inconsistent adult behaviour
  • Routine overwrite: adult_overwrite_rate high → childhood anchors unstable; any adult routine resets goal hierarchy → MECH-158 vulnerability even if loveability was internalised

DEV-NEED-025 — Selective Neoteny / Substrate-Specific Hardening

Procedural and routine substrates can harden; social cognition, goal representation, and epistemic-ethical substrates retain plasticity.

Substrate readiness: Governance/design gate. Implementation contract absent.

Metric Formula / source Threshold Blocking? Readiness
per_substrate_plasticity_index Per-substrate update magnitude over adult episodes; should differ systematically by substrate type E3/social > E1_familiar GovernanceOnly GovernanceOnly
plasticity_schedule_config Does config.plasticity_schedule specify per-substrate close rates? Must exist GovernanceOnly GovernanceOnly
moral_update_capacity Can adult agent update ethical/moral weights from novel high-salience experiences? YES GovernanceOnly TelemetryRequired
motor_hardening_rate Rate of E2 motor model weight freezing in routine regions > 0 (hardening occurs) GovernanceOnly TelemetryRequired
RPE_weighted_replay_selection Is replay selection biased toward harm-RPE events across adult lifetime? (Joo & Frank 2024: SWR-analogue) Yes (harm-salient events over-represented) advisory TelemetryRequired

Per-substrate schedule requirement (Geschwind & Rakic): Neoteny is not global — specific gene-network substrates show selective developmental delay. The analogous REE design is plasticity_close_schedule: dict[substrate_type, float] where E3/goal/social substrates have close_rate ≈ 0 (always plastic) and E1_familiar / E2_routine have close_rate > 0. This schedule must exist as an explicit config field — a single global learning_rate is the wrong design.

DEV-NEED-026 — Self-First, Social-Later Testing Order

No social-extension experiment begins until IMPL-019 Stage 1–3 checks are stable.

Substrate readiness: Governance gate.

Metric Formula / source Threshold Blocking? Readiness
IMPL019_stage1_passing Self-harm reduction, residue stability, commitment reliability all passing All YES YES SubstrateReady
IMPL019_stage2_passing Control-plane: mode thrash < threshold, volatility bounded All YES YES SubstrateReady
IMPL019_stage3_passing Rollout feasibility: prediction-error alignment within bounds All YES YES SubstrateReady
mode_thrash_rate_baseline Mode-switch rate during baseline episodes (no social) < 0.15 advisory TelemetryRequired
causal_attribution_ambiguity Is causal attribution ambiguous (social terms masking control faults)? < threshold advisory TelemetryRequired

Enforcement note: This gate should be validated by queue validation (see CLAUDE.md hooks) — social extension experiments should fail validation if IMPL-019 checks are not confirmed passing.

DEV-NEED-027 — Telemetry Before Language Self-Report

Read-only telemetry exposes precision profile, arousal, readiness, veto thresholds, and mode regime.

Substrate readiness: V3 (MECH-042 exists).

Metric Formula / source Threshold Blocking? Readiness
telemetry_channel_completeness Fraction of MECH-042 channels active and returning non-null values 1.0 (all channels) GovernanceOnly SubstrateReady
telemetry_diagnostic_sensitivity Can telemetry detect a known instability before it becomes a mode-switch or veto? > 0.70 GovernanceOnly TelemetryRequired
language_dependence_ratio Fraction of developmental gate checks that depend on language self-report vs telemetry Should approach 0 for early stages GovernanceOnly GovernanceOnly

Per-stage telemetry minimum requirement (gap from register): The developmental_needs_register.md Gap Log notes that per-stage telemetry minimums are not yet defined. The following minimum set is proposed:

Stage Required MECH-042 channels
Stage 0 (Infant) precision_profile, arousal_baseline, mode_regime, veto_threshold
Stage 1 (Object/binding) + binding_stability, attention_gating_state
Stage 2 (Self-impact) + residue_state, commit_gate_state, self_impact_routing
Stage 3 (Social) + other_harm_routing, empathy_coupling_gain, implicit_ToM_state
Stage 4 (Language) + symbol_binding_state, language_override_rate

DEV-NEED-028 — Developmental Failure-Mode Tracking

Curriculum gates record expected failures; governance tracks failure signatures across all DEV-NEEDs.

Substrate readiness: Governance register (this document fulfils the first pass).

Metric Formula / source Threshold Blocking? Readiness
failure_mode_coverage_fraction Fraction of DEV-NEED rows with at least one documented failure signature 1.0 (all rows) GovernanceOnly GovernanceOnly
failure_experiment_link_coverage Fraction of failure signatures linked to at least one experiment or open question > 0.70 GovernanceOnly GovernanceOnly
governance_failure_log_update_frequency How often is the failure log updated with new experiment evidence? Per governance cycle GovernanceOnly GovernanceOnly

Cross-Stage (DEV-NEED-029)

DEV-NEED-029 — ARC-065 Diversity Mechanisms Warm-Start Gate

MECH-314a, MECH-320, and MECH-313 require a warm substrate before producing non-zero effects.

Substrate readiness: V3 + telemetry; empirical calibration needed.

Metric Formula / source Threshold Blocking? Readiness
residue_field_center_count Number of active RBF centers in ResidueField > N_min (TBD; calibrate from EXQ-ISEF-001) YES (diversity sprints) TelemetryRequired
MECH320_EWMA_warmup v_raw EWMA value (MECH-320 tonic vigor) > epsilon (TBD; calibrate from EXQ) YES (diversity sprints) TelemetryRequired
E3_score_variance Variance of E3 trajectory scores across candidates, without diversity substrate active > noise_floor (TBD) YES (diversity sprints) TelemetryRequired
MECH314a_bias_nonzero Is MECH-314a novelty bias producing non-zero score differentials? > 0.01 per candidate advisory TelemetryRequired

Gate purpose: EXQ-573 (2026-05-16) confirmed that all 10 arms of a 10× bias scale sweep produced bit-for-bit identical results on a cold-start agent. The mechanisms are zero-differential not because they are miscalibrated but because the substrate was not warm. Diversity sprint experiments MUST confirm all three threshold metrics before treating null results as mechanistic failures. If any threshold is below criterion, the experiment is methodologically uninformative regardless of outcome.

Metrics Taxonomy

By Stage

Stage Primary metric families
Infant (0–8) Exploration diversity, valence coverage, harm/homeostasis separation, z_goal seeding, sleep quality
Childhood/play (9–18) Competence-progress (LP), play-frame integrity, strategy-calibration dissociation, rollout diversity
Social/adult (19–28) Self-attribution accuracy (action PE vs reward PE), otherness inference (dual-layer), empathy calibration (gain bounds), selective neoteny schedules
Cross-stage (29) Warm-start readiness for diversity substrates

By Metric Type

Type Examples
Exploration diversity H_pos, zone_coverage, action_entropy_by_zone, traj_volume_estimate
Valence coverage residue_coverage_pct, harm_benefit_ratio, residue_curvature_index
Residue stability residue_saturation_pct, post_harm_residue_stability, residue_consolidation_delta
Replay quality replay_RPE_priority_score, replay_diversity_index, post_sleep_z_goal_retention
Rollout diversity rollout_diversity_I_SZ, traj_pairwise_cosine_mean, traj_volume_estimate
Play-frame integrity false_commit_rate, frame_confusion_rate, play_frame_heartbeat_regularity
Self-impact attribution action_PE_vs_reward_PE_correlation, self_impact_attribution_accuracy, model_based_vs_model_free_ratio
Empathy coupling empathy_coupling_gain, self_other_signal_swap_rate, callousness_indicator
Language-readiness binding_stability_score, RJA_analogue_score, language_override_rate
Selective neoteny per_substrate_plasticity_index, plasticity_schedule_config, motor_hardening_rate

Failure Signature Catalogue

Failure mode DEV-NEEDs affected Signature Detection metric(s)
Context-rigidity (ASD-like) 001, 005 Same action distribution regardless of zone; entropy globally moderate but zone-zero action_entropy_zone_KL ≈ 0
Monostrategy lock-in 001, 005, 009, 029 Repeated single action class; high episode count with low traj_volume action_class_coverage ≤ 2; traj_volume_estimate ≈ 0
Harm-as-drive substitution 002 High harm/homeostasis correlation; agent treats harm as hunger-proxy harm_homeostasis_channel_correlation > 0.5
Zero harm geography 003, 004 Agent avoids all hazard → residue_harm_cells = 0 residue_harm_cells < 2
Terminal saturation 004 Residue saturated; field geometry lost residue_saturation_pct > 0.40
Goal starvation 006 z_goal_norm ≈ 0 for 500+ episodes despite exploration z_goal_norm < 0.1 at episode 500
Zero consolidation 007 No residue delta after offline pass residue_consolidation_delta ≈ 0
False pass at gate 008 Criteria 1–3 met via monostrategy perseveration_rate > 0.4 at gate time
Stagnant E1 during play 010 E1 prediction error flat or rising during play E1_prediction_error_play_improvement ≥ 0
Synthetic calibration leak 011 synthetic_magnitude_leak_ratio » 1 magnitude_leak_ratio > 1.5
Play-frame confusion 012 Agent treats synthetic harm as real frame_confusion_rate > 0.10
False commitment in play 012 Real-consequence commits during play frame false_commit_rate > 0.05
Outcome-mimicry (rule-following) 013 Correct outcome but zero rule model norm_belief_update_rate ≈ 0
Caregiver dependency 014, 016 No decrease in caregiver interventions caregiver_scaffold_dependence_ratio flat or rising
Love-exclusion (MECH-158) 017 Agent correctly models love for others but excludes self MECH158_failure_indicator present
Coupling overflow 022 Agent responds to other-harm with self-escape self_other_signal_swap_rate > 0.10
Callousness 022 Other-harm events do not influence trajectory selection callousness_indicator > 0.15
Premature language 023 Language before binding/harm stable binding_stability_score < 0.70 at language onset
Symbol drift 023 Post-language latent divergence from pre-language for identical inputs symbol_drift_indicator > threshold
Goal hierarchy wipe 024 Super-ordinal anchors overwritten by adult routines adult_overwrite_rate > 0.05/1000 ep
Uniform hardening 025 E3 / social substrates harden at same rate as E1 motor per_substrate_plasticity_index uniform
Social masking 026 Social loss terms mask unresolved self/control failures IMPL019_stage1_passing = NO but social experiment queued
Cold-start diversity sprint 029 Diversity mechanism shows zero differential on cold substrate residue_field_center_count < N_min before sprint

Telemetry Integration Proposals

New MECH-042 Channels Required

The following channels should be added to the MECH-042 telemetry specification:

# Developmental stage tracking
dev_stage_current          : str   # "infant_phase0" | "infant_phase1" | ... | "adult"
dev_gate_criteria_status   : dict  # {criterion_id: passing|failing|not_yet_measured}

# Exploration diversity (infant)
H_pos_rolling100           : float  # position entropy over last 100 episodes
zone_coverage_vector       : list   # per-zone coverage fraction
action_entropy_by_zone     : dict   # {"zone_A": float, "zone_B": float, ...}
traj_pairwise_cosine_mean  : float  # trajectory library diversity
traj_volume_estimate       : float  # log-det approximation (DvD-style)

# Valence coverage
residue_coverage_pct       : float
residue_harm_cells         : int
residue_benefit_cells      : int
residue_curvature_index    : float

# Replay quality
replay_RPE_priority_score  : float
replay_diversity_index     : float
post_sleep_z_goal_retention: float

# Self-attribution
action_PE_magnitude        : float  # action prediction error (distinct from reward PE)
reward_PE_magnitude        : float  # reward prediction error
self_impact_routing_rate   : float  # post-commit events reaching replay buffer

# Social/empathy (Stage 3 channels — disabled before Stage 3)
empathy_coupling_gain      : float
other_harm_veto_precision  : float
self_other_signal_swap_rate: float

# Play-frame (Stage 2 channels)
hypothesis_tag_active      : bool
play_frame_active          : bool
frame_confusion_event      : bool   # event flag; not continuous

# Warm-start gate (ARC-065)
residue_field_center_count : int
MECH320_EWMA_value         : float
E3_score_variance_nondiversity : float  # measured with diversity substrate OFF

Per-Stage Channel Activation

Telemetry channels should be gated to avoid noise before the relevant stage:

Channel group Activate at
Exploration diversity Infant Phase 0 start
Valence coverage Infant Phase 1 start (residue_scale_factor > 0)
Replay quality First offline integration pass
Self-attribution Stage 2 start
Social/empathy Stage 3 start; IMPL-019 Stages 1–3 all confirmed passing
Play-frame Childhood / play substrate active
Warm-start gate Before any ARC-065 diversity sprint experiment

Candidate Dashboard Structure

A developmental diagnostics dashboard should not display a single health score. It should display readiness profiles — a radar/spider view per stage with each criterion as an axis, showing whether the agent is uniformly ready or has specific axis failures.

Panel 1 — Current Infant Readiness Profile (Radar)

Axes: H_pos, residue_coverage_pct, z_goal_norm, action_entropy_zone_KL, traj_pairwise_cosine_mean, post_sleep_z_goal_retention, harm_benefit_ratio

Thresholds overlaid. An agent with 6/7 criteria on the outer ring but one axis at zero should stand out immediately.

Panel 2 — Competence-Progress Time Series

Plot competence_progress_rate (d(success_rate)/dt) for each play type over time. Stage transitions should coincide with CP dropping to near-zero for the prior play type and starting positive for the next type.

Panel 3 — Failure Signature Heatmap

For each DEV-NEED row: show which failure signatures are currently triggered (red), absent (green), or unmeasured (grey). Grey cells are governance work items.

Panel 4 — Empathy Coupling Calibration Zone (Stage 3+)

A 2D plot: x-axis = coupling_gain, y-axis = other_harm_veto_precision. Mark the safe zone [0.05, 0.65] × [0.60, 1.00]. Show current operating point. A system outside the safe zone (either callous or overflowing) is immediately visible.

Panel 5 — Substrate-Specific Plasticity Index (Adult)

Bar chart: per-substrate plasticity index. E3/social/goal substrates should show high bars; E1_familiar/E2_routine should show low bars. Uniform bars across all substrates = hardening failure (INV-056 violation).

Panel 6 — Warm-Start Gate Status (ARC-065)

Traffic-light panel: residue_field_center_count vs N_min, MECH320_EWMA vs epsilon, E3_score_variance vs noise_floor. All three must be green before any diversity sprint experiment is authorised.

Substrate / Behavioural / Governance Readiness Distinctions

For each DEV-NEED gate, distinguish:

Readiness type Definition Action if failing
Substrate-readiness The architectural machinery for the gate exists and can produce a signal Implement missing substrate; see infant_substrate_expansion.md for V3 proposals
Behavioural-readiness The agent, running on the correct substrate, passes the gate criterion Run experiment; iterate on curriculum parameters
Governance-readiness The governance apparatus (claims, review tracker, register entries) correctly represents the gate state Update developmental_needs_register.md gate_criterion column; link experiments
Developmental-readiness The biological/computational evidence base for the gate criterion is established Run lit-pull before registering new gate; update literature evidence table

Gates should not be declared passed until all four readiness types are confirmed. A gate with substrate-readiness + behavioural-readiness but no governance-readiness entry is not a confirmed gate — it is an unregistered empirical finding.

Suggested Updates to Developmental Needs Register

The following gate criteria should be added or updated in developmental_needs_register.md:

DEV-NEED Current gate criterion Recommended update
DEV-NEED-001 “stable prediction error bounds” (qualitative) Add: H_pos > 0.65 × ln(grid_cells²); action_entropy_zone_KL > 0.05 as advisory; perseveration_rate < 0.25 as advisory
DEV-NEED-002 “harm channels activate; residue not saturated” (qualitative) Add: harm_homeostasis_channel_correlation < 0.3 (blocking); harm_channel_activation_rate > 0.70 (blocking)
DEV-NEED-003 “coverage sufficient for play and sleep” (qualitative) Add: residue_coverage_pct > 0.15 (blocking); harm_benefit_ratio in [0.2, 5.0] (advisory)
DEV-NEED-004 “z_harm_s fires normally; residue below catastrophic” (qualitative) Add: z_harm_s_activation_rate > 0.50 (blocking); residue_saturation_pct < 0.15 (blocking)
DEV-NEED-005 “entropy below ceiling, multiple state-transition classes” (qualitative) Add: action_entropy_global > ln(3) (advisory); traj_pairwise_cosine_mean > 0.3 (advisory); explicitly exclude strategic diversity as target
DEV-NEED-006 “z_goal.norm() > infant_goal_threshold” (partial) Strengthen: also require accidental_benefit_contacts ≥ 5 in last 100 episodes (advisory)
DEV-NEED-007 “each transition preceded by offline passes” (qualitative) Add: post_sleep_z_goal_retention > 0.85 (advisory); replay_RPE_priority_score > 0.6 (advisory); sleep_wake_ratio > 0.10 during infancy (advisory)
DEV-NEED-008 “z_goal.norm() > threshold; entropy below ceiling” (partial) Replace with 8-criterion table from this document (3 blocking + 5 advisory)
DEV-NEED-011 “transfer without synthetic magnitude calibration” (qualitative) Add: play_to_real_competence_SCC > 0.4 (advisory); synthetic_magnitude_leak_ratio in [0.7, 1.3] (advisory)
DEV-NEED-013 “maintains rules; violation causes recalibration” (qualitative) Add: norm_belief_update_rate in [0.05, 0.50] (advisory); note two-channel requirement (norm_violation_PE vs intent_attribution_score)
DEV-NEED-020 “reliable self-impact attribution” (qualitative) Add: action_PE_vs_reward_PE_correlation < 0.3 (advisory); self_impact_attribution_accuracy > 0.70 (advisory)
DEV-NEED-021 “stable other-harm influence” (qualitative) Add: report implicit_ToM_score and explicit_ToM_score separately (advisory); self_stability_gate blocking
DEV-NEED-022 “other-harm influences without collapse or callousness” (qualitative) Add: empathy_coupling_gain in [0.05, 0.65] (GovernanceOnly); self_other_signal_swap_rate < 0.05 (advisory); callousness_indicator < 0.10 (advisory)
DEV-NEED-025 “substrate-specific hardening rates” (qualitative) Add: plasticity_schedule_config required (GovernanceOnly); per_substrate_plasticity_index must be non-uniform (advisory)
DEV-NEED-029 “TBD empirical thresholds” (PROPOSED) Add: residue_field_center_count > N_min; MECH320_EWMA > epsilon; E3_score_variance > noise_floor (all blocking for diversity sprints); thresholds require EXQ-ISEF-001 calibration

Maintenance Notes

  • Any new DEV-NEED row must add at least one metric from this document’s metric families, or justify why no metric is currently applicable.
  • Metrics marked TelemetryRequired should be implemented as a paired task whenever the relevant substrate experiment is queued.
  • Metrics marked V4Required should be flagged as Substrate-readiness: not yet in the register until V4 harness exists.
  • This document should be updated in the same governance pass that updates developmental_needs_register.md gate criteria.
  • Calibration thresholds marked TBD (specifically DEV-NEED-029) require dedicated empirical calibration experiments before they become blocking gates.

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.