Developmental Needs Register

This is the canonical audit and checklist view of REE developmental needs. The narrative specification remains Developmental Curriculum and Staged Training.

Metrics companion: Quantitative metrics, failure signatures, telemetry proposals, and gate thresholds for each DEV-NEED row are in developmental_metrics.md (2026-05-16).

Use this register when planning implementations, reviewing developmental claims, or checking whether a new developmental idea is already represented. The register preserves existing REE terminology and claim IDs; it does not replace docs/claims/claims.yaml.

Register Rules

  • A developmental need is a required curriculum condition, gate, mechanism, or failure-mode check.
  • A row marked by existing claim IDs is already represented by the claims registry, even if the gate is still immature.
  • A row marked PROPOSED is not a new registered claim; it is a traceability marker for a gap that may need a future claim.
  • “Gate criterion” means either a quantitative test, an implementation gate, or a governance check that can eventually be made testable.
  • “Current status” distinguishes narrative commitments from implementation-ready requirements.

Status Vocabulary

  • Specified: stated in canonical architecture documentation, whether or not it has an executable gate.
  • Claimed: represented by one or more registered REE claim IDs.
  • Narrative gate: a stage or readiness gate described conceptually but not yet converted into a thresholded test.
  • Implementation-ready: substrate, instrumentation, and gate shape are concrete enough for an implementation or experiment task.
  • V3 proxy only: single-agent V3 can test a curriculum approximation, but not the full developmental claim.
  • Requires V4: needs multi-agent, caregiver, peer-frame, or social substrate beyond the current V3 substrate.
  • PROPOSED: not a registered claim; used only to preserve traceability for a gap that may need a future claim.

Stage Coverage Summary

Load-bearing single-claim dependencies: INV-055 grounds all 8 infant-stage DEV-NEED rows (001–008). ARC-019 grounds 10+ rows across all stages. INV-041 grounds all 10 childhood-stage rows (009–018). A governance demotion of any of these three claims requires a full register review pass.

Needs Register

Dev Need ID Developmental Need Stage Claim IDs Required substrate / mechanism Gate criterion Failure if absent Current status / maturity Priority / readiness Evidence / file links Open questions
DEV-NEED-032 Phase 0 reward-free babbling epoch Pre-infant (Stage 0 prerequisite) ARC-074, INV-055, MECH-277, INV-073 A reward-free Hebbian babbling epoch must precede any E3 scoring or reward-weighted learning: basic z_self structure, non-trivial residue gradients, and sensorimotor mappings are established before the infant stage’s harm/benefit exposure begins. ARC-074 substrate implemented; use_babbling_phase config flag enabled; z_self.norm() increases during babbling and residue coverage > 0 before first hazard contact. Entire infant stage substrate is cold-start; MECH-313/314/320 diversity mechanisms are inert; DEV-NEED-001 gate cannot be meaningfully tested. Claimed (ARC-074, 2026-05-16); implementation pending; no dedicated experiment yet. Now developmental_curriculum.md#arc-046
infant_substrate_expansion.md §4 Phase 0
Griffin 2026 (Hebbian+dopamine babbling model); Garcia-Guzman 2026 (spontaneous arm dynamics); Doupe & Kuhl 1999 (song learning)		 What is the minimum babbling epoch duration before the infant stage’s harm gradient is activated? Should babbling be strictly reward-free or can incidental benefit contacts be registered?
DEV-NEED-001 Sensorimotor grounding Stage 0 / infant ARC-019, INV-055, MECH-215 WORLD, HOMEOSTASIS, HARM, and SELF_SENSORY streams remain separable; E1/E2 prediction loops have stable basic dynamics. Stable prediction-error bounds, safe harm avoidance, and coherent self/world attribution before later stages. Later ethics, social inference, and language bind to ungrounded streams; self/world confusion contaminates responsibility. Specified as narrative gate; quantitative bounds still immature. Now developmental_curriculum.md#arc-019
developmental_curriculum.md#inv-055		 What exact prediction-error, self/world attribution, and harm-avoidance thresholds define “stable enough” for Stage 0 exit?
DEV-NEED-002 Harm/homeostasis separation Stage 0 / infant ARC-019, ARC-046, INV-055 HARM and HOMEOSTASIS channels are trained as distinct signals; z_harm_s and z_harm_a fire during protected exposure without being collapsed into drive. Harm channels activate under reduced hazard exposure while homeostatic drive remains separately measurable; residue does not saturate. Harm, hunger, fatigue, and survival pressure blur together; E3 learns unsafe avoidance or drive-proxy shortcuts. Specified; needs explicit channel-separation gate. Now developmental_curriculum.md#inv-055
developmental_curriculum.md#arc-046
control_plane.md		 Should the gate be based on channel correlation, ablation sensitivity, or downstream E3 calibration?
DEV-NEED-003 Infant valence-map formation Infant INV-055, ARC-046, ARC-013 Residue field and SD-014-style valence geography are populated by early harm/benefit contact, with sleep consolidating the map. Coverage of harm/benefit geography is sufficient for later play and childhood constraints; residue curvature remains bounded. Childhood has no affective geography to constrain or build on; later goal pursuit is strategically empty or brittle. Specified; gate mostly qualitative. Now developmental_curriculum.md#inv-055
residue_geometry.md
sleep.md		 What minimum coverage metric distinguishes a usable valence map from random exploratory traces?
DEV-NEED-004 Protected harm exposure Infant ARC-046, INV-043, INV-055 Curriculum or caregiver attenuates harm magnitude and residue accumulation while allowing normal nociceptive pathway training. z_harm_s and z_harm_a fire normally; residue accumulation stays below catastrophic saturation; protection is removed gradually. Agent is destroyed or permanently residue-saturated before acquiring avoidance, repertoire, or repair capacity. Claimed; V3 uses curriculum parameters, full caregiver version requires multi-agent substrate. Now developmental_curriculum.md#arc-046
developmental_curriculum.md#inv-043		 What attenuation schedule best preserves harm salience without hiding consequences?
DEV-NEED-005 Behavioural repertoire before childhood Infant to childhood transition INV-055, ARC-019 Novelty-driven exploration trains E1 on diverse transitions and gives E2 enough action-outcome range for structured play. Behavioral entropy is below the pure-random ceiling but broad enough to show multiple viable state-transition classes. Childhood affordance constraints have nothing to constrain; play becomes arbitrary signal injection over a monostrategy. Specified; gate named but not calibrated. After V_s diversity developmental_curriculum.md#inv-055
developmental_curriculum.md#minimal-stage-gates		 Which entropy floor/ceiling and transition-diversity metrics should be canonical?
DEV-NEED-006 z_goal seeding from accidental benefit Infant INV-055, MECH-189, MECH-329, MECH-306 z_goal is seeded by accidental benefit contacts before deliberate planning. MECH-329 (wanting-before-liking sequencing, 2026-05-16) specifies the ordering: wanting (VALENCE_WANTING) write precedes liking (VALENCE_LIKING) write during accidental benefit contact, grounding z_goal in approach motivation before consummatory reinforcement. z_goal.norm() > 0.4 (canonical blocking threshold from developmental_metrics.md) before childhood entry; z_goal writes traceable to benefit encounters; wanting-write precedes liking-write in episode records. Goal channels remain empty; childhood play has no target representation to structure; SD-012-style goals remain blocked. Specified; canonical threshold 0.4 in developmental_metrics.md; wanting-before-liking mechanism registered 2026-05-16. After V_s diversity developmental_curriculum.md#mech-189
developmental_metrics.md DEV-NEED-006 row
play_mode.md#play-as-sd-012-curriculum-arc-050		 Does wanting-before-liking ordering require temporal separation between the two writes, or is co-occurrence acceptable?
DEV-NEED-007 Frequent offline integration during early development Infant and stage transitions INV-055, INV-010, ARC-011 Sleep/offline integration is frequent during infancy and between stages; replay consolidates residue, z_goal, and E1 transition structure. Each developmental transition is preceded by offline passes that improve map stability, goal seeds, and repertoire quality. Waking experience remains unstable; stage transitions amplify noise, residue spikes, or monostrategy collapse. Specified; substrate improving in V3, developmental gate not yet canonical. Now developmental_curriculum.md#offline-integration-between-stages
sleep.md
hippocampal_systems.md#replay-and-offline-reprojection		 What sleep:wake ratio and post-sleep improvement metrics should be required for infancy?
DEV-NEED-008 Transition gate from infancy to childhood Infant to childhood transition INV-055, INV-060, ARC-019, ARC-042, ARC-041, MECH-160 Infant outputs must be present before play-dominant childhood starts: valence map, repertoire, and z_goal seed. z_goal.norm() > infant_goal_threshold and behavioral entropy below ceiling; sensorimotor play can be structured. ARC-042 E1 gate: cosine_sim(hazard-proximate, hazard-distal) < 0.95 (EXQ-181b found 0.9999 untrained – supervised ContextMemory training per MECH-153 required before E3 receives gradient signal). ARC-042 E2 gate: action_bias non-near-zero (E2 trajectory generation is context-differentiated, not context-blind). Childhood starts before substrate exists; play cannot train goal pursuit or frame maintenance meaningfully. Without ARC-042 E1/E2 gates: E3 ethical selection operates as uniform scoring throughout childhood – architecturally present but non-discriminating; near-chance ethical selection despite wired circuitry (EXQ-181b: cosine_sim=0.9999, 0/3 seeds pass). Explicit in curriculum; needs executable gate. After V_s diversity developmental_curriculum.md#inv-055
developmental_curriculum.md#inv-060		 Should the gate require residue coverage, E1 quality, or only z_goal and behavioral entropy?
DEV-NEED-009 Play-dominant childhood Childhood INV-041, INV-060, INV-058, MECH-194, ARC-073 Play is the primary learning mode for goal-pursuit competence; synthetic goal/harm signals train E3/E2/E1 with full learning flow. ARC-073 specifies that the play-to-real transition is triggered by competence saturation (d(PE)/dt < threshold), not by scheduled duration – premature transition leaves policy structure undertrained. Childhood training mix is play-dominant; goal-pursuit competence transfer to real episodes confirmed; transition occurs at competence saturation, not schedule. Strategy acquisition requires real harm or is avoided; adult goal pursuit stays undertrained; premature schedule-based transition causes representational brittleness (Baranes & Oudeyer 2013). Without a committed/constrained childhood phase (INV-041): E3 ethical selection operates on near-random terrain weighting; adult ethical agency is brittle under novel social contexts lacking childhood-era cue-value associations (developmental IGT analog: Bechara et al. 2000). Claimed; ARC-073 registered 2026-05-16; implementation needs staged play harnesses with LPM-based exit criterion. Requires play substrate developmental_curriculum.md#inv-060
play_mode.md#why-play-is-structurally-necessary-inv-058
play_substrate_design.md §5 roadmap		 What is the operationalized LPM threshold (d(PE)/dt < X) that triggers the play-to-real transition?
DEV-NEED-010 Sensorimotor play Early childhood INV-041, INV-060, MECH-197, MECH-327, Q-048 Single-step synthetic goals and binary synthetic harm structure novelty-driven exploration into action-outcome learning. Within play episodes, action selection is steered by the E1 prediction-error (PE) locus: actions targeting high-PE regions of z_world are preferentially chosen over lower-PE alternatives, implementing curiosity-driven probe-action selection within the play frame (MECH-327). E1/E2 improve on action-outcome prediction under play without requiring real harm or real benefit; during sensorimotor play episodes, action selection shows PE-locus bias – actions disproportionately probe high-PE regions of z_world relative to chance (MECH-327 criterion). Infant exploration does not become structured; constructive play prerequisites are unmet; if PE-locus steering is absent, exploration is uniform rather than directed toward regions of maximal learning signal, slowing E1 calibration. Claimed as play progression; MECH-327 registered and added here 2026-05-16; PE-locus gate criterion pending. Requires play substrate developmental_curriculum.md#inv-060
play_mode.md#sensorimotor-play-earliest
play_substrate_design.md		 What E1/E2 competence level is enough to enter constructive play? What PE-locus bias coefficient (fraction of actions targeting top-k PE regions) distinguishes structured from unstructured sensorimotor exploration?
DEV-NEED-011 Constructive play Childhood INV-041, MECH-197, MECH-195, ARC-050, MECH-196 Multi-step synthetic goals train E2 rollout and E3 trajectory selection; policy structure transfers while magnitude calibration is re-anchored later. Play-trained trajectory competence transfers to real episodes without transferring synthetic magnitude calibration. Agent can move but cannot build multi-step strategies; play contaminates real goal/harm thresholds. Claimed; test design implied by play transfer and recalibration. Requires play substrate play_mode.md#constructive-play
play_mode.md#strategycalibration-dissociation-mech-195
play_mode.md#play-as-sd-012-curriculum-arc-050		 Which transfer metric separates “policy structure transferred” from “synthetic calibration leaked”?
DEV-NEED-012 Pretend play / counterfactual play-frame handling Childhood INV-041, MECH-198, ARC-049, MECH-094, Q-049, INV-061, MECH-200, MECH-201, MECH-202 Hypothesis tag and play frame tag cooperate: counterfactual representations are maintained inside an authorized synthetic frame. ARC-049’s three-level signal architecture (L1 discrete open/close frame markers; L2 continuous background signal; L3 explicit repair signals at ambiguity junctures, Bekoff 1995) delimits pretend sequences so both parties can distinguish play content from real commitment targets. Agent can act on counterfactual play content while preserving the real/synthetic distinction and commitment-gate discipline; L1 frame-open and frame-close signals (play-bow equivalent) delimit pretend sequences and are recognisable as distinct from non-play action; commitment gate does not fire on counterfactual content inside a valid L1/L2-maintained frame. Adult commitment architecture is untested in low-stakes mode; frame confusion, false commits, or synthetic-real leakage emerge; frame-boundary collapse (absent, ambiguous, or unilateral L1 signal) allows counterfactual content to leak into the real commitment pathway, producing false commits or real harm discounted as play-harm. Claimed with a testable prediction; needs developmental ablation; Gap Log note re MECH-327 corrected 2026-05-16 (see Gap Log). Requires play substrate developmental_curriculum.md#inv-060
play_mode.md#pretend-play-mech-198
play_substrate_design.md §1		 What minimal pretend-play task exposes false commits and frame confusion without requiring full social V4? What L1 signal criterion (duration, detectability, bilateral confirmation) is sufficient to distinguish a valid pretend frame from unframed action?
DEV-NEED-013 Games-with-rules learning Later childhood INV-041, MECH-197, MECH-328, ARC-049, INV-059, Q-035, Q-050 Shared constraints become part of the play frame; ongoing shared-state monitoring and rule-following are trained. MECH-328 (ongoing bilateral monitoring protocol) is the implementation mechanism for ARC-049’s heartbeat requirement. Agent maintains rules, turn structure, and frame integrity across episodes; rule violation causes recalibration or repair. The agent cannot coordinate under shared constraints; open/close-only frame signaling may hide mid-episode collapse. Claimed as progression; Q-035 resolved 2026-05-16 (ongoing monitoring is primitive; heartbeat required). Requires play substrate play_mode.md#games-with-rules
play_substrate_design.md §1.10/1.11
play_mode.md#frame-integrity-play-vs-manipulation-inv-059		 What minimal implementation of the ARC-049 heartbeat satisfies the games-with-rules frame integrity requirement at V3 proxy level?
DEV-NEED-014 Cooperative play and peer frame-maintenance Late childhood to adulthood INV-041, MECH-197, MECH-199, ARC-049, ARC-047, Q-051 Multi-agent play frames are co-maintained by peers; caregiver scaffolding is withdrawn. Agent maintains play frames peer-to-peer without caregiver authority and transfers cooperation to real-consequence contexts. Adult cannot distinguish safe peer play from manipulation or cannot coordinate shared goals. Conceptually specified; full test is V4 multi-agent. Requires V4 multi-agent play_mode.md#cooperativesocial-play
play_mode.md#caregiver-role-transition-mech-199
social.md#arc-047		 What is the minimal V4 social harness for peer frame-maintenance?
DEV-NEED-015 Caregiver protection function Infant and early childhood INV-041, INV-043, ARC-046 Caregiver or curriculum prevents irreversible harm while allowing imperfect harm-learning. Harm learning occurs under protection; irreversible damage and catastrophic residue are avoided. Ethical development collapses into survival, domination, or damage repair before E3 is ready. Claimed; V3 curriculum proxy only. V3 proxy only developmental_curriculum.md#inv-043
developmental_curriculum.md#arc-046		 Which parts of caregiver protection can be tested in single-agent V3 versus requiring V4 caregiving agents?
DEV-NEED-016 Caregiver frame-maintenance function Childhood INV-041, MECH-199, INV-043, ARC-049 Caregiver opens, monitors, repairs, and gradually hands off play frame maintenance across play types. Caregiver intervention protects frame boundaries early, then withdrawal leaves peer-level frame competence intact. Play becomes unsafe or unilateral; strategy acquisition is confused with manipulation. Claimed; full implementation requires multi-agent frame machinery. Requires V4 multi-agent developmental_curriculum.md#inv-060
play_mode.md#caregiver-role-transition-mech-199
play_mode.md#frame-maintenance-signal-architecture-arc-049		 How should withdrawal be staged so caregiver authority does not become a permanent crutch?
DEV-NEED-017 Love / loveability internalisation Childhood, with lifelong consequences INV-041, INV-043, MECH-158, MECH-052 Caregiver models love and shared valence; the developing agent internalises that love applies to itself, not only to others in the abstract. Self-valence model treats care/love as personally applicable; care weights can motivate self-other relations rather than remaining abstract. “Love exists but not for me”: ethical motivation collapses into survival, domination, or indifference. Claimed; not directly testable in V3 single-agent substrate. Requires V4 multi-agent developmental_curriculum.md#inv-043
developmental_curriculum.md#mech-158
social.md#mech-052		 What behavioral signature distinguishes loveability internalisation from merely predicting caregiver reward?
DEV-NEED-018 Repair after harm Childhood through adulthood INV-041, ARC-019, INV-043, MECH-158 Curriculum includes repair, forgiveness, residue integration, and corrigibility after harm caused before full ethical capacity. Post-harm episodes show repair behavior, residue integration without destabilization, and non-punishment-only learning. Responsibility becomes punishment-only; residue destabilizes; the agent learns shame, domination, or avoidance instead of repair. Narrative commitment under existing claims; no dedicated repair gate. Requires play substrate developmental_curriculum.md#responsibility-precedes-full-ethical-capacity
control_plane.md#justification-gate-invariants-care-vs-other-harm
sleep.md		 Gap: should repair after harm become its own claim or remain part of ARC-019/INV-043?
DEV-NEED-019 Gradual responsibility expansion Infant to adulthood ARC-019, ARC-046, INV-043, IMPL-019 E3 load, commitment thresholds, and consequence exposure increase by stage as attribution and stability improve. Responsibility-bearing tasks expand only after self-harm, recovery, control-plane, and rollout gates are stable. Premature full E3 load creates responsibility collapse, overcommitment, or brittle guilt/punishment dynamics. Specified; testing order claim exists, quantitative schedule missing. Now developmental_curriculum.md#responsibility-precedes-full-ethical-capacity
developmental_curriculum.md#impl-019
e3.md		 What staged thresholds define partial versus full ethical capacity?
DEV-NEED-020 Self-impact and residue formation Stage 2 ARC-019, ARC-013, ARC-015, INV-012 ACTION -> SELF_SENSORY -> SELF_IMPACT loop binds post-commit outcomes to residue geometry and responsibility flow. Reliable self-impact attribution; residue integrates without destabilizing modes; post-commit outcomes route to replay/audit. Over-commitment occurs before responsibility attribution; residue becomes generic reward or noise. Existing architecture; developmental gate is narrative. Now developmental_curriculum.md#minimal-stage-gates
residue_geometry.md
e3.md		 Which self-impact attribution metrics should gate entry into otherness inference?
DEV-NEED-021 Otherness inference after self-stability Stage 3 / social extension ARC-019, ARC-010, MECH-031, MECH-032, IMPL-019 OTHER_SELFLIKE inference reuses self/world/impact machinery after object persistence, self-viability, and control-plane stability are adequate. Social extension begins only after self-viability, control-plane stability, and rollout feasibility gates hold. Social losses mask unresolved self/control faults; empathy and otherness are misattributed to unstable targets. Specified; enforced as testing-order heuristic. Requires V4 multi-agent developmental_curriculum.md#impl-019
social.md#mech-031
social.md#mech-032		 What self-stability gate is sufficient before introducing multi-agent loss terms?
DEV-NEED-022 Empathy coupling calibration Stage 3 / social extension ARC-010, MECH-031, MECH-032, MECH-036 Shadow bundles, lambda_empathy, v_other_veto, and social precision route inferred other states into pruning and ranking. Other-harm influences selection without empathic collapse or callousness; vetoes fire only under high certainty/catastrophic risk. Too-early or overstrong coupling causes overwhelm; undercoupling causes callousness and false-negative other-harm. Existing social architecture; developmental calibration gates immature. Requires V4 multi-agent developmental_curriculum.md#minimal-stage-gates
social.md#mech-031
social.md#mech-036		 How should coupling tighten over development after the high-recall OTHER_SELFLIKE bias?
DEV-NEED-023 Language only after stable binding/harm channels Stage 4 / optional ARC-019, ARC-048, IMPL-019, MECH-042, MECH-308 Language is introduced as high-bandwidth externalization after stable binding, harm, self-impact, and telemetry channels already exist. Language improves coordination without overriding embodied harm channels or causing symbol drift/confabulation. Language bootstraps before referents are stable; symbols override harm and binding channels. Narrative gate; no dedicated developmental language experiment. Governance only developmental_curriculum.md#minimal-stage-gates
social.md
control_plane.md#telemetry-exposure-channels-mech-042		 What gate shows language improves coordination without overriding harm, binding, or telemetry channels?
DEV-NEED-024 Super-ordinal goal formation Childhood MECH-189, INV-056, INV-055 High-salience benefit under high contextual complexity writes persistent cue-indexed ContextMemory anchors that bias adult z_goal seeding. Adult episodes show persistent goal hierarchy from childhood anchors, while routine adult contexts do not overwrite it. Goal selection remains purely episodic; strategic inconsistency and MECH-158 vulnerability increase. Claimed; implementation/gates pending. Requires play substrate developmental_curriculum.md#mech-189
hippocampal_systems.md		 What contextual-complexity threshold triggers a super-ordinal write? Lit verdict 2026-06-10 (directional, not a settle): prefer complexity_mode='external' PE/surprise gate over the monotonic 'novelty' default (biological novelty = comparator-mismatch vs whole world model, not anchor-key distance; encoding is signed/nonmonotonic in PE). See mech_189_evidence_overview.md / targeted_review_connectome_mech_189.
DEV-NEED-025 Selective neoteny / substrate-specific hardening Lifespan INV-056, MECH-159 Procedural and routine motor substrates can harden; social cognition, goal representation, and epistemic-ethical substrates retain plasticity. Plasticity schedules are substrate-specific and preserve moral/social/goal update capacity through adulthood. Uniform hardening creates adult MECH-158-like failure: love can be understood but not newly accessed or updated. Claimed as design principle; implementation contract absent. Governance only developmental_curriculum.md#inv-056
social.md#mech-159		 How are hardening rates represented in substrate configs and governance gates?
DEV-NEED-026 Self-first, social-later testing order Implementation sequencing IMPL-019, ARC-019, ARC-010 Harness readiness, self-viability, control-plane stability, and rollout maturity are tested before social extension and coupling. No social-extension experiment begins until self-harm/recovery, mode-switch, and rollout-feasibility metrics are stable. Social behavior becomes uninterpretable because base faults are masked by social terms. Active governance heuristic; not a new architecture commitment. Governance only developmental_curriculum.md#impl-019
social.md#arc-047		 Should this order be enforced by queue validation or remain a review convention?
DEV-NEED-027 Telemetry before language self-report All early stages, especially pre-language MECH-042, ARC-019 Read-only telemetry exposes precision profile, arousal, readiness, veto thresholds, and mode regime before language can report needs. Developmental diagnostics can detect instability without relying on language output or inducing severe destabilization. Failure modes remain hidden until trauma-level events or confabulatory language reports surface them. Existing MECH-042; not yet integrated into every developmental gate. Now control_plane.md#telemetry-exposure-channels-mech-042
developmental_curriculum.md#arc-019		 Which telemetry channels are required per stage, and which are optional diagnostics?
DEV-NEED-028 Developmental failure-mode tracking Governance / all stages ARC-019, INV-043, MECH-158, MECH-198, INV-059; PROPOSED if converted into a dedicated governance claim Curriculum gates must record absences and failure signatures: love-exclusion, frame confusion, residue saturation, premature responsibility, language drift, empathic collapse, and monostrategy lock-in. Each developmental gate records expected failure if absent and links failed experiments or open questions to the relevant need row. Repeated developmental ideas fragment across prose docs; governance cannot tell whether a failure is new, known, or already represented. This register implements the first source-doc pass; dedicated tracker/claim not registered. Governance only developmental_curriculum.md#curriculum-failure-modes
developmental_curriculum.md#mech-158
play_mode.md#frame-integrity-play-vs-manipulation-inv-059		 Should failure-mode tracking live only in this register, or also become a generated governance index?
DEV-NEED-029 ARC-065 diversity mechanisms warm-start gate Infant / infant-to-childhood transition (cross-stage) ARC-065, MECH-313, MECH-314, MECH-320; PROPOSED gate claim ARC-065 child mechanisms (MECH-314a novelty, MECH-320 tonic vigor, MECH-313 noise floor) require a warm substrate before producing non-zero effects: a populated ResidueField, a built-up EWMA reward-rate, and non-random E3 scores (trained network). The infant stage produces this warm substrate; the mechanisms were designed for a post-infant agent, not a cold-start one. A formal gate criterion is needed to prevent diversity sprint experiments from running at the wrong developmental stage and returning methodologically correct but scientifically uninformative null results. MECH-314a produces non-zero score bias (ResidueField has at least N_centers centers); MECH-320 v_raw EWMA > epsilon after K episodes; E3 score variance > noise floor (measured without diversity substrates). All three thresholds TBD from empirical calibration. Diversity sprint experiments run on cold-start agents, return zero-differential results (confirmed: V3-EXQ-573 all 10 arms bit-for-bit identical), and are misread as calibration failures. Sprint effort spent scaling bias parameters that are zero because the substrate is not ready, not because the parameters are miscalibrated. PROPOSED gap. V3-EXQ-573 (2026-05-16) confirmed zero-differential across 10x bias scale sweep due to cold-start warm-start failure. No dedicated gate criterion yet. See behavioral_diversity_acceptance_criteria.md open questions 4-5. After V_s diversity behavioral_diversity_acceptance_criteria.md open questions 4-6
EXQ-573 null result (2026-05-16)
developmental_curriculum.md#inv-055		 What minimum episodes / ResidueField coverage / EWMA warmup are required before each of MECH-313, MECH-314a, MECH-320 can produce non-zero effects? Should v_t_floor serve as an infant-stage proxy for MECH-320 before EWMA is available?
DEV-NEED-030 Stage-aware replay scheduling Cross-stage (infant, childhood, adult) ARC-011, SD-017, MECH-285; PROPOSED new scheduling claim Replay scheduler must accept a dev_stage parameter and dispatch a stage-indexed policy: infant = coverage-maximising (coverage_floor >= 0.6, RPE_weight <= 0.2); childhood = interleaved play+real (play:real >= 1:1, RPE_weight = 0.4); adult = RPE+staleness priority (RPE_weight = 0.6, staleness_weight = 0.3). See replay_development_analysis.md §6.1–6.4 for parameter tables. replay_scheduler_stage matches current dev_stage; stage-specific metrics pass (infant: replay_zone_coverage_fraction > 0.6; childhood: replay_play_real_interleave_ratio >= 0.5; adult: monostrategy_prevention_score >= 1.0). Infant stage uses adult RPE-priority: replay_low_salience_fraction < 0.1, valence map never consolidated. Adulthood uses infant coverage-priority: MECH-124 option-space contraction is never caught. PROPOSED; not registered as a claim. Analysis and parameter tables in replay_development_analysis.md §6. After infant_substrate_expansion.md substrate features replay_development_analysis.md
sleep_aggregation_cluster.md
developmental_curriculum.md		 What is the stage-detection mechanism — inferred from episode count, explicit curriculum parameter, or gate-passage signal? Should the scheduler fall back to coverage-priority when no stage is set?
DEV-NEED-031 MECH-124 prevention gate All stages (especially adult and post-harm childhood) MECH-124, ARC-011; PROPOSED new monitoring claim monostrategy_prevention_score (post-sleep traj_volume / pre-sleep traj_volume) must be monitored per sleep cycle; MECH-285 staleness correction provides the mechanism. If monostrategy_prevention_score < 1.0 over a rolling window, sleep is amplifying the dominant trajectory rather than expanding the option space. monostrategy_prevention_score >= 1.0 across rolling 5 sleep cycles; option_space_contraction_rate <= 0 (traj_volume_estimate non-declining across sleep). Progressive option-space contraction: adult agent converges to PTSD/depression phenotype (Walker PTSD analog). High-RPE traumatic content replayed preferentially → dominant mode lock-in → viable alternatives excluded from policy. MECH-124 contraction proceeds silently because current telemetry lacks the monostrategy_prevention_score channel. PROPOSED; not registered as a claim. Monitoring design and clinical mapping in replay_development_analysis.md §4, §6.4. With sleep aggregation cluster implementation replay_development_analysis.md
sleep_aggregation_cluster.md
hippocampal_systems.md		 Should MECH-124 prevention gate block stage progression or only trigger an alert? What intervention is available when the gate is red — increase staleness_weight, reduce RPE_weight, or require a coverage-boosting recovery cycle?
DEV-NEED-033 Developmental stage ordering and E1/E2/E3 maturational sequence Cross-stage (self -> objects -> others) ARC-059, INV-064, MECH-276, ARC-081 Training proceeds in three ordered stages (ARC-059: (1) self-as-object action-space discovery, (2) objects-as-patterns schema formation, (3) others-as-special-objects modelling); E1/E2 sensorimotor competence matures before E3 ethical/goal integration completes (INV-064 maturational-sequence necessity); the waking-phase signal driving cross-stage schema aggregation is deliberate scientist-agent intervention, not passive correlation (MECH-276). Stage N substrate is present before Stage N+1 training begins (self-attribution substrate before object-schema formation before other-modelling); E3 integration metrics lag E1/E2 competence rather than leading it. Object and other schemas bind to an ungrounded self; the adult risk-profile / executive-function pathologies that INV-064 unifies (adolescent risk-taking past sensorimotor competence; outsized early-life-disruption effects on executive function) are mis-modelled; aggregation generalises from correlation without counterfactual self-attribution. Claimed (ARC-059 / MECH-276 2026-04-21; INV-064); ordering is a design + governance commitment; no dedicated staged-ablation experiment yet. After V_s diversity developmental_curriculum.md#arc-019
claims.yaml ARC-059 / INV-064 / MECH-276 (scientist-agent cluster, MECH-275–278)		 What metric certifies a Stage-N substrate as “present enough” to begin Stage N+1? Can the three-stage ordering be validated in single-agent V3, or does others-as-special-objects require V4 multi-agent?
DEV-NEED-034 Plasticity-crystallization necessity and critical-period lock Infant through adolescence (lifespan crystallization) INV-075, INV-074, ARC-075, ARC-076 Any signal that is BOTH self-extinguishing AND load-bearing requires either a plasticity LOCK or active maintenance once the bootstrapping signal fades (INV-075 parent invariant); diversity circuits converge to monostrategy under uniform Hebbian-equivalent plasticity unless a time-bounded plasticity asymmetry lets them reach competitive weight (INV-074, the competitive special case); the infant curriculum must implement plasticity-magnitude asymmetry – suppressing F’s effective update weight during the MECH-333 plasticity-open window (ARC-075); the three BG-like commitment loops (ARC-021 / MECH-069 / ARC-023) are calibrated in a bounded juvenile/adolescent window and then crystallized as an INV-075 LOCK-arm instance (ARC-076). The plasticity-open window measurably suppresses dominant-pathway (F) update weight relative to diversity circuits; post-window crystallization is measurable (diversity representations persist rather than being overwritten); commitment-loop calibration completes before the lock engages. Warm-start / infant-curriculum diversity benefit is transient – overwritten by the dominant scoring pathway (the INV-074 monostrategy mechanism, confirmed by V3-EXQ-573 zero-differential); commitment loops never stabilize or never re-open. Claimed (INV-074 / INV-075 / ARC-075 / ARC-076 cluster); MECH-333 substrate-side; no dedicated plasticity-asymmetry experiment yet. After V_s diversity behavioral_diversity_acceptance_criteria.md
developmental_curriculum.md
EXQ-573 null result (2026-05-16)		 What window duration and plasticity-suppression ratio produce lasting diversity benefit? How is the LOCK arm represented in substrate configs and governance gates? (cross-links DEV-NEED-025 selective neoteny, DEV-NEED-029 warm-start gate)
DEV-NEED-035 Caregiver-scaffolded rule population (natural pedagogy / ZPD) Childhood / social MECH-337, ARC-077, ARC-063 A third rule-source pathway, peer to ARC-062 (top-down gated-policy apprehension) and ARC-064 (bottom-up behavioural-pattern extraction) and fed upstream by ARC-065 behavioural diversity (ARC-077): a caregiver/teacher structures the agent’s exposure (Vygotsky ZPD; Wood/Bruner/Ross scaffolding) and marks regularities ostensively (Csibra & Gergely natural pedagogy); marked regularities enter the ARC-063 CandidateRule field (MECH-337 is the “in” face of the three-face model). Ostensively-marked regularities enter the candidate-rule field at higher weight than unmarked co-occurrences; caregiver scaffolding measurably accelerates rule acquisition relative to unscaffolded exposure. Rule acquisition relies solely on bottom-up extraction (slow, error-prone) or top-down imposition (brittle); the social-pedagogy route that grounds shared norms is absent. Claimed (ARC-077 / MECH-337, rule_apprehension_layer.md); full bilateral validation requires V4 multi-agent caregiver substrate. Requires V4 multi-agent rule_apprehension_layer.md
arc_062_rule_apprehension_plan.md (GAP-L)		 What V3 proxy (e.g. a scripted caregiver that ostensively marks regularities) can test the ostensive-weighting prediction before full V4 social agents exist?

Gap Log

Requirements Implied But Not Explicitly Claimed

  • DEV-NEED-018 repair after harm is present in the curriculum narrative under ARC-019 and INV-043, but there is no dedicated repair-after-harm claim or gate.
  • DEV-NEED-023 language timing is implied by ARC-019 Stage 4 and by ARC-048’s language bootstrap framing, but the gate tying language to stable binding/harm channels remains qualitative. Lasch et al. 2023 (Infancy; RJA at 10-11 months predicts vocabulary growth) provides a candidate quantitative pre-linguistic language-readiness gate.
  • DEV-NEED-027 telemetry before language self-report is supported by MECH-042, but the curriculum does not yet define per-stage telemetry minimums.
  • DEV-NEED-028 failure-mode tracking is implemented here as a register practice. It should become a dedicated claim only if REE governance needs a formal failure-mode tracker beyond source documentation.
  • DEV-NEED-029 ARC-065 warm-start gate is a PROPOSED gap surfaced 2026-05-16 by V3-EXQ-573 null result. No claim registered yet. The gate criterion (ResidueField coverage, EWMA warmup threshold, E3 score variance floor) requires empirical calibration from a two-phase warm-start + measurement experiment before a formal Q-claim or SD can be written.
  • DEV-NEED-030 stage-aware replay scheduling is a PROPOSED gap surfaced 2026-05-16 by replay_development_analysis.md. The current replay scheduler has no developmental stage awareness; infant-appropriate coverage-maximising policy and adult RPE+staleness policy are both absent. A stage-indexed scheduler must be implemented before EXQ-IDEV-003 and EXQ-IDEV-004 can produce meaningful signal. No claim registered yet.
  • DEV-NEED-031 MECH-124 prevention gate is a PROPOSED gap surfaced 2026-05-16 by replay_development_analysis.md. Monitoring for monostrategy_prevention_score is absent; option-space contraction can proceed silently across sleep cycles. The gate requires monostrategy_prevention_score telemetry (new MECH-042 channel) and MECH-285 staleness correction. No claim registered yet; priority is with sleep aggregation cluster implementation.
  • DEV-NEED-032 Phase 0 babbling epoch added 2026-05-16 (ARC-074 registered same day). Implementation pending; no experiment queued yet. Precedes DEV-NEED-001 in the infant stage.
  • Additional claims registered 2026-05-16 that need future register integration (not yet full rows; tracked per maintenance rule MA-3 from developmental_governance_review.md): ARC-073 (play-to-real transition via competence saturation) partially integrated into DEV-NEED-009; MECH-329 (wanting-before-liking) partially integrated into DEV-NEED-006; MECH-327 (PE locus probe-action selection, sensorimotor play) RESOLVED – added to DEV-NEED-010 Claim IDs, Mechanism, Metric/Gate, and open question columns 2026-05-16. Note: a prior Gap Log entry mislabeled MECH-327 as “play signal punctuation” – that content (Bekoff 1995 three-level L1/L2/L3 signal architecture) is in ARC-049, already present in DEV-NEED-012; DEV-NEED-012 Metric/Gate and Failure Mode updated 2026-05-16 to surface L1/L3 criteria and the frame-boundary-collapse failure mode explicitly. MECH-328 (goal-space continuity for play-to-real transfer) still needs a DEV-NEED-011/013 row note when play substrate implementation begins.
  • ARC-042 and INV-041 integrated 2026-05-16 per MA-3/MA-4 audit (developmental_governance_review.md Priority 1 missing claims): ARC-042 added to DEV-NEED-008 (Claim IDs + Gate criterion extended with E1 ContextMemory cosine_sim(hazard-proximate, hazard-distal) < 0.95 threshold and E2 action_bias non-near-zero threshold; Failure if absent extended with E3 uniform-scoring consequence). INV-041 added to all 10 childhood-stage rows (DEV-NEED-009 through DEV-NEED-018 Claim IDs). INV-041 elevated to load-bearing SPoF in Stage Coverage Summary (grounds childhood rows 009–018). Failure mode for DEV-NEED-009 extended with adult ethical brittleness consequence: novel social contexts lacking childhood-era cue-value associations (Bechara et al. 2000 IGT analog).

Claims Referenced But Lacking Implementation Substrate

  • INV-043 and MECH-158 require multi-agent caregiver modelling for full tests; V3 only supplies curriculum proxies.
  • ARC-049, INV-059, Q-035, MECH-197, MECH-198, and MECH-199 need play-frame substrate beyond the V3 degenerate experimenter-set play mode for full bilateral validation.
  • INV-056 needs a substrate-level hardening/plasticity schedule before it can be tested as an implementation contract.
  • MECH-189 needs cue-indexed persistent goal-anchor writes and an adult z_goal seeding readout before it can be gated.
  • ARC-048 is now surfaced as a related developmental claim, but the curriculum still lacks a language-readiness gate.

Items Lacking Gate Criteria

Updated 2026-05-16: developmental_metrics.md provides quantitative thresholds for all DEV-NEEDs where V3 substrate is sufficient. Summary of new gates added:

  • DEV-NEED-001: H_pos > 0.65 × ln(grid_cells²) [blocking]; action_entropy_zone_KL > 0.05, perseveration_rate < 0.25 [advisory]
  • DEV-NEED-002: harm_homeostasis_channel_correlation < 0.3 [blocking]; harm_channel_activation_rate > 0.70 [blocking]
  • DEV-NEED-003: residue_coverage_pct > 0.15 [blocking]; harm_benefit_ratio in [0.2, 5.0] [advisory]
  • DEV-NEED-004: z_harm_s_activation_rate > 0.50 [blocking]; residue_saturation_pct < 0.15 [blocking]
  • DEV-NEED-005: action_entropy_global > ln(3) [advisory]; traj_pairwise_cosine_mean > 0.3 [advisory]
  • DEV-NEED-006: z_goal.norm() > 0.4 [blocking]; accidental_benefit_contacts ≥ 5 in last 100 episodes [advisory]
  • DEV-NEED-007: post_sleep_z_goal_retention > 0.85 [advisory]; replay_RPE_priority_score > 0.6 [advisory]
  • DEV-NEED-008: 8-criterion table (3 blocking, 5 advisory) — see developmental_metrics.md
  • DEV-NEED-011: play_to_real_competence_SCC > 0.4 [advisory]; synthetic_magnitude_leak_ratio in [0.7, 1.3] [advisory]
  • DEV-NEED-013: norm_belief_update_rate in [0.05, 0.50] [advisory]; two-channel (outcome vs intent) required
  • DEV-NEED-017: loveability_coupling_gain in [0.1, 0.7] [GovernanceOnly]; MECH158_failure_indicator: absent [GovernanceOnly]
  • DEV-NEED-020: action_PE_vs_reward_PE_correlation < 0.3 [advisory]; self_impact_attribution_accuracy > 0.70 [advisory]
  • DEV-NEED-021: implicit_ToM_score and explicit_ToM_score reported separately; self_stability_gate blocking
  • DEV-NEED-022: empathy_coupling_gain in [0.05, 0.65] [GovernanceOnly]; self_other_signal_swap_rate < 0.05 [advisory]
  • DEV-NEED-025: plasticity_schedule_config required field [GovernanceOnly]; non-uniform per_substrate_plasticity_index [advisory]
  • DEV-NEED-029: residue_field_center_count > N_min; MECH320_EWMA > epsilon; E3_score_variance > noise_floor [all blocking for diversity sprints; thresholds TBD from EXQ-ISEF-001]

Remaining gaps:

  • DEV-NEED-010 through DEV-NEED-014: most gates require V4 multi-agent substrate; V3 proxies specified in developmental_metrics.md for sensorimotor and constructive play gates.
  • Thresholds marked TBD in DEV-NEED-029 require empirical calibration from dedicated experiments before becoming blocking gates.

Cross-ref (ARC-080 object-representation umbrella, 2026-06-04). DEV-NEED-021 (otherness inference after self-stability, which lists object persistence as a prerequisite) and DEV-NEED-033 (the ARC-059 self->objects->others ordering) both presuppose an object-representation layer. That layer is now mapped as a single cross-cutting primitive under ARC-080: the four object capabilities (permanence, self-as-object, tools, others-as-object) are specialisations of one primitive, and ARC-080 records that the object permanence DEV-NEED-021 presupposes is not yet built (only the SD-039 goal-snapshot ghost bank exists, not a token-instance object-file). Object-substrate work is V4 / late-V3 – off the V3-closure path.

Items Lacking Experiments

  • No dedicated developmental ablation currently validates the full infant-to-child gate.
  • No full play-type progression experiment currently validates sensorimotor -> constructive -> pretend -> rules -> cooperative ordering.
  • No V4 caregiver experiment currently validates loveability internalisation or caregiver frame maintenance.
  • No experiment currently distinguishes telemetry-first diagnostics from language self-report diagnostics.
  • No experiment currently tests selective neoteny as substrate-specific hardening rather than a narrative design principle.

Contradictions Or Ambiguous Dependencies

  • Play helps unblock SD-012 by synthetically seeding z_goal, but infant-stage z_goal seeding still requires accidental benefit contact before childhood. The dependency boundary between DEV-NEED-006 and DEV-NEED-009 should remain explicit.
  • Q-035 resolved (2026-05-16 per claims.yaml and play_substrate_design.md §1.10/1.11): continuous ongoing monitoring is the correct framing and is primitive from 2-3 months in infant development (Tronick 1978, Fantasia 2014, Bekoff 1995); what escalates across play types is content complexity, not the monitoring architecture. ARC-049 must specify ongoing heartbeat mechanism (not just open/close tags). DEV-NEED-013 open question updated accordingly.
  • V3 curriculum proxies can test protected harm exposure and some play mechanics, but cannot test INV-043’s core caregiver-love hypothesis. Rows that require caregiver agency should not be promoted from V3-only evidence.
  • Language timing depends on stable binding and harm channels. ARC-048 is cross-linked, but no quantitative gate yet prevents future language work from bypassing developmental readiness checks.

Maintenance Rules

  • Any future developmental claim must be entered in this register in the same change that registers or modifies the claim.
  • Every register row must have at least one claim ID or be marked PROPOSED.
  • Every row should eventually have a gate criterion and failure mode. If either is immature, say so explicitly.
  • If an item is split, superseded, or retired, preserve traceability by marking the row status and linking successor rows; do not silently delete the row.
  • New experiment proposals that test development should cite the relevant DEV-NEED-* row in addition to claim IDs.
  • Generated indexes should not be edited by hand only to reference this register. Update source documentation and let the normal governance/indexing pipeline regenerate derived files when appropriate.

Amendments added 2026-05-16 (from developmental_governance_review.md)

  • MA-1 (Claims.yaml precedence): When a claim referenced in this register changes status in claims.yaml (e.g., a Q-claim is resolved, an open question is closed), the register’s open questions and Gap Log must be updated in the same governance pass. Do not leave the register narrative inconsistent with the canonical claim status.
  • MA-2 (Quarterly evidence-state review): At each full governance cycle, update the “Current status / maturity” column of rows whose primary claims have changed exp_conf or tier since the last review. Mark “Evidence advancing” when exp_conf has increased; “Promoted” or “Demoted” when the claim changed status tier.
  • MA-3 (Developmental-keyword claims require same-day register update): Any claim registered in claims.yaml whose title or description contains developmental language (infant, childhood, play, curriculum, stage, caregiver, babbling, pretend, repertoire, neoteny) must have a register row (or an update to an existing row) in the same commit. If the claim is V4-only and no row exists, add a Requires V4 multi-agent row with a PROPOSED gate.
  • MA-4 (Orphaned claims check at governance cycle): At each full governance cycle, grep claims.yaml for developmental-keyword claims and verify each has a register row. New orphans become PROPOSED rows in the next pass.
  • MA-5 (Single-point-of-failure annotation): Claims appearing in 5 or more rows (currently INV-055 and ARC-019) are noted in the Stage Coverage Summary as load-bearing. A governance demotion of either requires a full register review pass before the demotion recommendation is confirmed.

Amendments added 2026-06-03 (backward-traceability G2 reconciliation)

  • check_backward_traceability.py precision fix. The G2 checker had matched developmental keywords by raw substring, so play matched replay/interplay/preplay (the hippocampal/sleep replay family appears ~300x across claims) and stage matched non-developmental “processing stage” / “NREM stage 2” / “two-stage architecture” / “Stage 1 implementation”. This produced a persistent ~122-claim warning that was ~75% false positives. The checker now matches keywords at a word boundary (\bplay excludes replay) and treats stage as developmental only in developmental constructions (developmental stage, staged development, ordered stages, stage 0, infant/childhood/maturational stage). It also carries an explicit TRACEABILITY_EXEMPT list (MECH-090, MECH-172, SD-049, MECH-191) for claims whose keyword is an incidental mention, each with a one-line rationale. Per MA-3, the canonical keyword set still includes stage and neoteny; the checker’s word-boundary/context handling is a precision refinement, not a narrowing of the policy – a genuinely developmental claim using “stage” in a developmental construction is still caught.
  • 25 genuinely-developmental claims traced. After the precision fix, 25 real developmental claims lacked a register reference. Added to existing rows: DEV-NEED-001 (MECH-215), -006 (MECH-306), -008 (ARC-041, MECH-160), -010 (Q-048), -011 (MECH-196), -012 (Q-049, INV-061, MECH-200, MECH-201, MECH-202), -013 (Q-050), -014 (Q-051), -023 (MECH-308), -032 (MECH-277, INV-073). New cross-stage rows: DEV-NEED-033 (stage ordering + E1/E2/E3 maturation: ARC-059, INV-064, MECH-276), DEV-NEED-034 (plasticity crystallization / critical-period lock: INV-075, INV-074, ARC-075, ARC-076), DEV-NEED-035 (caregiver-scaffolded rule population: MECH-337, ARC-077). G2 check now passes (exit 0).

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