# STATUS — what is closed, and the number that closes it **S6-B12-R1 — CAMPAIGN_NO_PROMOTION.** Recuperación declarada del límite remoto; mismas propuestas y criterios, reproducción exacta obligatoria de las siete respuestas guardadas. [Adenda](ECHO45-S6-B12-RECOVERY1.md) · [Seguimiento](ECHO45-S6-B12-LIVE.md). **S6-B12 — CAMPAIGN_NO_PROMOTION. S6 todavía no verde.** Contexto verificado antes de inferencia, piloto previo auditado y campaña nueva. [Contrato](ECHO45-S6-B12-DESIGN.md) · [Resultados](ECHO45-S6-B12-RESULTS.md). **S6-B11 — PENDING_ERROR. S6 todavía no verde.** Desarrollo ampliado, contrastes antes/después y selección adaptativa; conservación estricta. [Contrato](ECHO45-S6-B11-DESIGN.md) · [Resultado](ECHO45-S6-B11-RESULTS.md) · [Estado](../lab/seed/s6b11-20260929/status.json). **S6-B10 — PILOTO SIN PROMOCIÓN. S6 todavía no verde.** ECHO 534, FIJO 534; aceptaciones ECHO 0. Piloto auditado, reserva sin abrir. [Resultado](ECHO45-S6-B10-RESULTS.md) · [Estado](../lab/seed/s6b10-20260929/status.json). **Diagnóstico S6, 29 sep:** B10 mejora niveles A y B, sin pérdida individual de niveles, pero todas sus propuestas hacen más lenta una misma partida protegida. V4 exige conservar también cada puntuación individual. Trece programas dan seis conductas medidas; evidencia de tareas repetida y feedback protegido limitado. B7 ya tuvo una mejora ECHO aceptada, pero empató con FIJO. No se cambian umbrales ni se abre reserva. [Criterios, historia y causa investigada](ECHO45-S6-DIAGNOSIS-20260929.md). **S6-B9 — PILOTO SIN PROMOCIÓN. S6 todavía no verde.** ECHO 534, FIJO 534; aceptaciones ECHO 0. Piloto auditado, reserva sin abrir. [Resultado](ECHO45-S6-B9-RESULTS.md) · [Estado](../lab/seed/s6b9-20260929/status.json). **Histórico del 29 sep, 05:15–07:07 Madrid — S6-B9:** seis pruebas y paridad de la base completas; piloto cerrado sin promoción y auditoría válida. ECHO/FIJO 534–534, cero aceptaciones; reserva sin abrir. El control humano positivo (+8 A sin pérdidas A/B) no se atribuye al modelo. [Resultado B9](ECHO45-S6-B9-RESULTS.md) · [Control](ECHO45-S6-TRANSITION-RESULTS.md). **Histórico del 29 sep — continuaciones CPU de S6-B8:** float16 terminó parcial (2/6 respuestas); la condición Q8 cerró el piloto y su auditoría, sin mejora ni promoción (534–534). Los cambios de backend están declarados y las reservas permanecen sin abrir. [Float16](ECHO45-S6-B8-CPU-DESIGN.md) · [Q8](ECHO45-S6-B8-CPUQ8-RESULTS.md). **29 sep — S6-B8 pendiente:** implementación y 14 pruebas completas; doce propuestas auditadas, cero aceptaciones, suma A ECHO/FIJO 534–534. Faltan seis propuestas: dos trabajos fallaron durante la descarga de pesos y quedan 13 segundos de GPU gratis. Reserva sin abrir. [Informe](ECHO45-S6-B8-RESULTS.md) · [Después de S6](ECHO45-POST-S6-PLAN-20260929.md). **29 sep 2026 — cierre de esta iteración:** EGO-MEMORIA-3E VERDE: M1–M5 16/16 y M6 exacto 16/16; revisiones 3A–3D conservadas. S6-B6/B7 cerrados y auditados; B7 ECHO 537 frente a FIJO 537, aceptaciones ECHO 1; sin promoción; examen final sin abrir. B5 y los exámenes anteriores permanecen cerrados; S6 todavía no se declara VERDE. [Memoria](ECHO45-EGO-MEMORIA3-RESULTS.md) · [Registro actual](ECHO45-RECOVERY-20260929.md). **Estado automático S6-B5 — 2026-09-28 17:36 Madrid: CLOSED.** Control preciso cerrado: 6/12 positivos, ganador preregistrado `hold20_after_level`. Es una reparación humana, no automejora. Piloto: ECHO acepta 0; suma A ECHO 534, FIJO 534; promoción no. S6 todavía no acredita B1–B5. Examen nuevo sin abrir. [Contrato](ECHO45-S6-B5-DESIGN.md) · [Informe](ECHO45-S6-B5-RESULTS.md) · [Estado](../lab/seed/s6b5-20260928/status.json). **ECHO-4.5, 2026-09-28 17:05 Madrid:** parches precisos probados y congelados; doce controles con B nueva en curso. Tres primeros positivos (178→181 A, sin pérdidas individuales A/B). B5 congelado y activo esperando cierre, con piloto, cuatro brazos y reserva nuevos condicionados a promociones. Refinamiento/varias vistas/relaciones implementados: tests pasan, cuatro ablaciones ARC-2 0/12 con S0; condición nueva con biblioteca aprendida S2v 2/12 en los cuatro brazos, sin ventaja adicional. No se acredita transferencia. [Roadmap temporal](ECHO4-ECHO5-TIMELINE.md). **ECHO-4.5, 2026-09-28 16:03 Madrid:** ejecuciones terminadas. Control manual S6: 20 reparaciones, ninguna supera V1–V4; cierre 15:31. `hold20` gana tres niveles A sin pérdidas A, pero pierde puntuación en una partida B. Piloto omitido, cero llamadas LLM/GPU. R4 cerrado 15:32: replay completo 1040 soluciones/32 partidas, procedencia y contabilidad válidas; R1/R4 pasan, R2/R3 fallan. S6 y ECHO-4.5 siguen ROJOS. Nueva hipótesis/capacidad necesita otro contrato y ensayos de desarrollo; no bajar umbrales ni reutilizar como ciego el examen abierto. [Control](ECHO45-S6-CAPACITY-RESULTS.md) · [Examen](ECHO45-FINAL-EXAM-RESULTS.md) · [Roadmap](ECHO45-ECHO5-ROADMAP.md). **ECHO-4.5, 2026-09-28 14:45 Madrid:** control positivo manual S6 implementado, probado y congelado; 20 reparaciones legales en desarrollo bajo V1–V4 originales. En segundo plano, cuatro procesos CPU, sin LLM ni examen. Activación histórica FIJO 35/40, acciones distintas 1/40; ECHO 36/40, acciones distintas 0/40; ambos 178 niveles. Resultado del control pendiente; S6 permanece ROJO. R4: 32/32 partidas ARC reproducidas con contabilidad SDK válida; rejillas todavía en replay. [Contrato del control](ECHO45-S6-CAPACITY-CONTROL.md) · [Estado vivo](NOW.md). **ECHO-4.5, 2026-09-28 tarde:** medición final terminada: R1 31/400 frente a 20/400 (criterio pasa); R2 0/120 (falla); R3 11/112 frente a 13/112 (falla, objetivo 23). Auditoría original interrumpida a las 04:07 por representación tupla/lista del registro; recuperación explícita congelada y en ejecución. No se modifican agente, criterios ni medidas. Cierre R4 todavía pendiente. [Resultados](ECHO45-FINAL-EXAM-RESULTS.md) · [Estado vivo](NOW.md). **ECHO-4.5, 2026-09-28 02:04 Madrid:** evaluación final R1–R4 implementada, validada en desarrollo, congelada y ejecutándose en segundo plano. Resultados pendientes; estimación de medición más auditoría 2–4 h. S6-B4 cerró su piloto sin promoción: cero aceptaciones entre 24, seis rondas auditadas, examen sin abrir; S6 continúa ROJO. EGO-MEMORIA-2 cerrado ROJO: M1–M4 16/16, M5 15/16, M6 replay exacto 16/16. Borrador Kaggle CPU v5.8 local, sin envío externo. [Roadmap](ECHO45-ECHO5-ROADMAP.md) · [Estado vivo](NOW.md). **ECHO-4 RELATION-1 started — 2026-09-22:** RELATION-A implemented and verified. Observed encounter histories on each agent's original CAM; no shared private memory, policy change or transfer claim. 912 primary episodes reconstruct; native behavior/cost parity with no-recording control and reversed execution. 36 new tests; 181 including selected INTERACTION/OTHER/core/Q regressions pass. Fresh-process audit valid, no findings, `phase_a_valid=true`, `relation1_green=false`: `ce24ccaa61f9d8b9b781274857b6ce15a9e75817bbf8d0b5fd2649a403369539`. 115 pinned sources; all 110 INTERACTION sources and artifacts intact. Next RELATION-B: learned use and transfer, then C/D. History replay does not resume the complete agent or prove consciousness. Remains 9/13 software milestones (~69%, not effort). No Discord, hardware, web, commit/push or HF. METAVERSE-1 last. [Results](ECHO4-RELATION1-A-RESULTS.md), [contract](ECHO4-RELATION1-A-DESIGN.md). **Historical ECHO-4 INTERACTION-1 closure — 2026-09-22:** B/C/D and fresh-process native reconstruction pass. `valid=true`, no findings, `interaction1_green=true`: `d1c5b073e7ac406ddd855184cc6f4bd1176d354eed57438eb51af1383f040735`. 110 pinned sources; all 104 A sources and sealed artifacts unchanged. 37 new tests and 11 adverse witnesses pass. Each new cohort: scored focal net 6912 vs solo/ablations 6144 (+12.5%); joint physical net 9984 vs 6656. Formation costs remain: full-life focal net 14059/14136 vs solo 15360. No full-life individual advantage, subjective consciousness or general social planning claim. One-step asymmetric profile, pretrained native counterparts with declared private incentives excluded from gain; original CAM/T/Q/gate, WSP16, no basal LLM, no copied Agent loop. Integration remains pending. Final regression: all 723 ECHO-4 tests and 33 selected core/Q tests pass; 756 distinct tests, no failures or skips. Next RELATION, then ROLES, INTEGRATE, RELEASE. 9/13 software milestones (~69%, including WORLD foundation), not effort/time estimate. HF not published; hardware after software, METAVERSE-1 last. No Discord/web/commit/push. [Results and limits](ECHO4-INTERACTION1-BCD-RESULTS.md), [frozen contract](ECHO4-INTERACTION1-BCD-DESIGN.md). **Historical ECHO-4 INTERACTION-1 start — 2026-09-22:** increment A verified. Costed shared physics, two original Agent loops with private CAM/T/Q, one own reward/update only after both gated actions. 243/243 transitions and nine scripted mechanical witnesses; three native 144-round development sequences. Per-agent net 485/690/668 versus solo 576 each: first sequence loses. OTHER remains passive; matching symmetric policies is not learned negotiation. 35 new tests, 108 including selected regressions, all pass. Fresh-process reconstruction valid, no findings, `phase_a_valid=true`, **`interaction1_green=false`**. Digest: `da8cc00befb52ac1abc9942bae1aa8eacf3fbe2049501dfebd26c2a1ed57c1c3`. 104 pinned sources; all 99 ancestral sources unchanged. Next INTERACTION-B, then C/D. Software milestone count remains 8/13 (~62%, including WORLD as implemented foundation), not estimated effort/time or integrated readiness. Four scientific phases remain including INTERACTION, then RELEASE. No Discord, hardware, web, commit, push or HF publication; METAVERSE-1 last. [Results](ECHO4-INTERACTION1-A-RESULTS.md), [contract](ECHO4-INTERACTION1-A-DESIGN.md). **Historical ECHO-4 OTHER-1 closure — 2026-09-22:** observational counterpart model on the observer's original CAM/T, two native agents with private Q/memory/worlds. 40 new tests pass. Two new cohorts: identity 1536/1536 each versus pooled 1008/1536, self-copy 0/1536 and no-history 0/1536. On equal preferences, self-copy also scores 1536/1536; no universal gain. First-half abstentions, negative control errors and private learning costs remain in the evidence. Fresh-process native reconstruction valid, no findings, `other1_green=true`: `1263f11ab291a1ff350313a93a70d9f4b3b9259099482bb8425e9f414c2aa024`. 99 pinned sources, 93 parental sources unchanged. Scope: discrete observable behavior prediction, NOT consciousness, useful coordination or thermal-body integration. WSP16 and original CAM/T/Q/gate retained, no basal LLM. Selected regression (OTHER/MAINTAIN/core/Q): 619 pass and one historical expected failure, 620 distinct tests. Eight sandbox socket denials resolved by rerunning the 29 affected GUI-class tests with localhost permission. Next **E4-INTERACTION-1**. Four scientific phases remain: INTERACTION, RELATION, ROLES, INTEGRATE; then RELEASE for planned `rxlabs/echoai-4`. No publication, Discord, hardware, web, commit or push. METAVERSE-1 last. [Results and limits](ECHO4-OTHER1-RESULTS.md), [contract](ECHO4-OTHER1-DESIGN.md). **Historical ECHO-4 MAINTAIN-1 closure — 2026-09-22:** B/C/D and independent native reconstruction pass. `valid=true`, no findings, `maintain1_green=true`, digest `7e1c0e821ba2ddf79834fef6a66a16bfd2773a456d29ac5bbf535c11e6024fbe`. 93 pinned sources; original MAINT-A and earlier evidence unchanged. Scope: autonomous virtual maintenance conditional on a declared mechanical model, NOT physics/repair policy learned from scratch or general self-repair. Two new cohorts, 960 lives: 144/144 recoverable lives completed and 192/192 repairs confirmed by subsequent executed cooling. Work 9450 vs fixed 7501, no-repair 4815 and no-body-history 4672; all thresholds pass in each cohort. Zero D deaths in all arms; 96/240 adaptive lives stop in the four limit scenarios. 24 paid services fail to repair; pairwise losses and all resource costs retained. C: 12 fresh-process continuation cases, 257 identical future events and 11/11 rejected mutants. 85/85 MAINTAIN tests pass, 42 new. No cortex calls, resets, replenishment, private causal truth in controller inputs, or hidden restoration. Final regression: 611/611 ECHO-4 tests pass; including core/Q, 1105 pass and one historical expected failure (1106 total). Eight sandbox socket denials resolved by rerunning the GUI classes with localhost permission; no new failures. Next **E4-OTHER-1**. Five scientific phases remain (OTHER, INTERACTION, RELATION, ROLES, INTEGRATE), then RELEASE for planned `rxlabs/echoai-4`; not published. Hardware later; METAVERSE-1 last. Legacy Discord remains out of scope. [Results and limits](ECHO4-MAINTAIN1-FINAL-RESULTS.md), [frozen contract](ECHO4-MAINTAIN1-BCD-DESIGN.md). **Historical ECHO-4 MAINT-A increment — 2026-09-22:** costed service fixture and native SelfAgent wiring implemented without changing the original loop/policy/gate. Repairable and nonrepairable faults, finite spares, persistent external heat, useful work and operator stop; no reset, respawn or hidden-truth policy input. 216 scripted episodes plus four witnesses, 1,090 native ticks; all invariants hold. Blind service: 18 repairs, 45 work, six deaths versus control 120 work, zero deaths. Original Q baseline dies after four ticks, with no work/repair. These are mechanism results, not autonomous maintenance or behavioral gain. 235 tests pass (43 new); fresh-process reconstruction valid, no findings, 87-source lock (`ec4a6d1e…`), all 82 prior sources intact. **`maintain1_green=false`; next MAINT-B**, evidence-based intervention choice. Six scientific phases remain including MAINTAIN, then RELEASE. Planned HF target `rxlabs/echoai-4`: installable agent and demonstration, not fused/new Qwen weights. No repo creation/upload; permissions/licenses/install still require validation. No Discord, hardware, web, commit or push changes. [Results](ECHO4-MAINTAIN1-A-RESULTS.md), [contract](ECHO4-MAINTAIN1-A-DESIGN.md), [release plan](../echo4/README.md). **ECHO-4 DREAM-D v3 closure — 2026-09-22:** E4-DREAM-1 is green within its restored WORLD-1 diagnostic-exploration profile. Two additional probes and evidence-conditioned safe branches preserve original SELF/physics/gate and the 12-step cap. Fifteen real fixed-weight Qwen calls, 60 valid proposals. New exams: DREAM 300/298/302 of 320; fixed 190/230/220; random 238/246/237. All per-cycle +32/+16 margins pass. Zero wrong diagnoses/terminalities in all ten arms, invariants intact and exact replay parity for all nine v3 arms. One initial promotion, two retentions; no-Qwen and no-feedback tie DREAM. No exclusive model or feedback benefit demonstrated. Full fresh-process reconstruction is valid without findings; receipt grants `dream1_green=true` (`965ce6cb…`). All 82 sealed sources and bound v1/v2 artifacts remain intact; both old red results are preserved. 926 tests: 925 pass, one historical expected failure. Owned model server closed; no basal LLM or weight updates. No consciousness, self-repair, general improvement or product integration claim. Next: MAINTAIN, not started; ECHO-4 as a whole remains open. No Discord, hardware, web deployment, commit, push or HF publication. [Results](ECHO4-DREAM1-D-V3-RESULTS.md), [contract](ECHO4-DREAM1-D-V3-DESIGN.md), [roadmap](ROADMAP.md). **ECHO-4 DREAM-D v2 — 2026-09-22:** replay feedback, worst-batch selection and promotion against both incumbent and original fixed control implemented. Nine real Qwen calls, 36 valid proposals; three stratified cycles of 320 cases per arm. DREAM 245/245/240, fixed 180/210/200, random 206/230/238. Fixed-margin checks pass throughout; random +16 fails in cycles 2/3 (+15/+2). **V2 remains red, `dream1_green=false`.** No-Qwen matches DREAM exactly; no exclusive LLM benefit demonstrated. All nine arms preserve invariants, with zero wrong diagnoses/terminalities and exact online/replay parity. Full reconstruction is valid without findings (`f5d36371…`); all 76 frozen sources and every bound v1 artifact remain unchanged after documentation edits. 891 tests: 890 pass, one historical expected failure. V1 evidence untouched. Post-hoc cycle-2 catalog enumeration gives maximum and oracle union245, below required246. Next v3 PLAN: more informative probes, then fresh preregistration/exam, not threshold/seed changes to turn v2 green. No basal LLM, weight modification, Discord use, hardware, web, commit or push. [Results](ECHO4-DREAM1-D-V2-RESULTS.md), [contract](ECHO4-DREAM1-D-V2-DESIGN.md), [roadmap](ROADMAP.md). **ECHO-4 DREAM-B/C/D implementation — 2026-09-22:** fixed-weight local Qwen ran six real calls, producing 24 valid bounded declarative candidates. Three complete cycles include historical replay, incumbent selection, new WORLD-1 trials, versioning/rollback and history expansion after evaluation. DREAM scores 40/40/50 out of 80 per cycle; fixed 40/40/70, random 41/42/72, no-Qwen 40/50/75. Zero wrong diagnoses/terminalities across all seven arms, invariants intact and exact online/replay parity. Full reconstruction valid without findings (`5feb033f…`), but **DREAM-D v1 is red: `dream1_green=false`**. The first historical selection improves 60→80/80 with fewer steps, but that does not meet the preregistered new-case advantage. Implementation is complete; scientific closure remains pending. Next: DREAM-D v2 development and a fresh preregistered exam, preserving v1. 871 distinct tests: 870 pass and one historical expected failure; eight sandbox GUI socket errors resolved by an authorized rerun without code changes. All 72 sealed sources and the CONT-D binding remain intact after doc updates. No basal LLM, weight changes, Discord use, hardware, web publication, commit or push. MAINTAIN remains pending; METAVERSE-1 stays last. [Results](ECHO4-DREAM1-BCD-RESULTS.md), [contract](ECHO4-DREAM1-BCD-DESIGN.md), [roadmap](ROADMAP.md). **ECHO-4 DREAM-A increment — 2026-09-22:** private native-attempt archives and budgeted partial replay implemented. Three WORLD-1 trees contain 31 nodes and 28 executed attempts; 62 replay queries return 28 observations, 32 unknowns and two terminal responses. Replay does not execute the world or call an LLM. Full checkpoint/archive/report reconstruction is valid without findings (`011cc0d9…`); all 66 sources and the existing CONT-D artifact binding remain intact. 31 new tests plus 717 regressions: 748 distinct tests, 747 pass and one historical expected failure. Eight sandbox socket errors resolved after an authorized rerun of the 29 GUI tests, without code changes. Ruff passes. This is infrastructure evidence, not better exploration or a DREAM-1 closure. Qwen, strategy generation/selection and new-task improvement remain pending. README now specifies subjective-consciousness limits, MAINTAIN repair goals, and runtime packaging of the fast/slow agent rather than fusion into weights. New RELEASE-1 PLAN follows INTEGRATE: package, runtime verification, license/ privacy review, then authorized Hugging Face publication and clean install. Stable software release precedes future lab hardware; METAVERSE-1 stays last. No upload, model activation, commit or push. [Results](ECHO4-DREAM1-A-RESULTS.md), [product README](../echo4/README.md), [roadmap](ROADMAP.md). **ECHO-4 CONT-D closure — 2026-09-22:** E4-CONTINUITY-1 is green within its declared synchronous WORLD-1 software profile. New A/B/C exams each pass 23/23 scenarios: 300/300 subsequent events and 186/186 native turns match fresh-process continuations; 30/30 mutants rejected. The 62-source lock is written before B. Full reconstructive verification and independent replays of CONT-A/B/C are valid with no findings. Only the audit receipt grants `continuity1_green=true` (`69d57a0b…`); the campaign grants eligibility only. 33 new tests plus 872 regressions: 905 run, 904 pass, one historical expected failure. Real-artifact checkpoint/journal corruption injected in memory is rejected without changing originals. History loss changes PATTERN learning counts but not actions/predictions here; Q remainders additionally have a separate synthetic arithmetic witness, not a claimed WORLD-1 behavioral gain. Existing SELF and CONT-A/B/C sources/artifacts remain intact. No subjective identity, autonomous repair, distributed exclusivity or hardware claim. Next: DREAM-A contract/history; DREAM-1 remains PLAN, cortex OFF. MAINTAIN remains pending in the basal branch. Hardware later, METAVERSE-1 last. No commit/push/web deployment. [Closure and limits](ECHO4-CONTINUITY1-D-RESULTS.md), [roadmap](ROADMAP.md). **ECHO-4 CONT-C update — 2026-09-22:** explicit virtual lineage distinguishes resume, replacement body, descendant and fork. New bodies inherit learned memory but revoke body-bound diagnostics/permissions, require quarantine and new supervised calibration/confirmation. Terminal individuals cannot rebody. 30 new tests pass; 805 general regressions and 37 CONT-B tests complete, with one historical expected failure and no unexpected failures. A/B audits remain valid; all sealed SELF/A/B sources retain their hashes. CONT-C completes its virtual profile: 21/21 cases, 125/125 subsequent events and 43/43 native turns match in fresh processes. Full reconstructive verification is valid, with no findings; a corrupted checkpoint is rejected. Full CONTINUITY remains open; next CONT-D final exam. `continuity1_green=false`. No LLM, second bus, subjective identity claim or distributed exclusivity. DREAM remains PLAN; hardware after software, METAVERSE-1 last. No commit/push/web. [CONT-C evidence](ECHO4-CONTINUITY1-C-RESULTS.md), [roadmap](ROADMAP.md). **ECHO-4 CONT-B update — 2026-09-22:** SELF diagnostic sessions and active recovery/reuse now checkpoint at complete operation boundaries, including pending plans, the receipt inbox, confirmation and original learning hooks. 37 new tests pass. All 18 development cases resume identically in fresh processes: 142/142 subsequent events and 67/67 physical turns match, including full state hashes. Independent reconstructive verification is valid with no findings; a tampered report is rejected. CONT-A verification also remains valid. Broad regressions: 805 tests, 804 pass and one historical expected failure. Together with the 37 new tests: 842, no unexpected failures. Global and life-local clocks remain distinct; vetoed, stopped and terminal runs do not restart. All 50 sealed SELF final sources and 37 CONT-A sources retain their hashes. **Full CONTINUITY remains open:** CONT-C lineage and CONT-D final exam remain. `continuity1_green=false`. Next: CONT-C. DREAM remains PLAN, cortex OFF, hardware after software and METAVERSE-1 last. No commit/push/deployment. [CONT-B evidence](ECHO4-CONTINUITY1-B-RESULTS.md), [roadmap](ROADMAP.md). **ECHO-4 update — 2026-09-22:** CONT-A implements nominal SelfAgent/WORLD-1 checkpoint/resume at complete turn boundaries, with a closed JSON schema, external custody digest, source fingerprints and shared original CAM/T/Q. 29 new tests pass. Five development scenarios resume in fresh processes; 52/52 subsequent steps and all state hashes match. Terminal state does not resurrect. Broad regression run: 805 tests, 804 pass and one historical expected failure after the 29 GUI tests passed on an authorized retry (eight localhost socket errors in the sandbox). Fresh verification is valid with no findings. All 50 sealed SELF-1 final sources retain their hashes. **Full CONTINUITY remains open:** active recovery, body/individual lineage and the final exam are not implemented. `continuity1_green=false`. Next: CONT-B. E4-DREAM-1 is now an official PLAN branch after CONTINUITY, inspired by Dream-RSI: fixed-weight local Qwen proposes bounded offline exploration strategies, with fixed evaluation and new holdouts. No model calls/training or kernel changes here; baseline cortex stays OFF. Hardware follows software; METAVERSE-1 remains last. No commit/push/deployment. [CONT-A evidence](ECHO4-CONTINUITY1-A-RESULTS.md), [DREAM plan](ECHO4-DREAM1-PLAN.md), [roadmap](ROADMAP.md). **ECHO-4 final closure — 2026-09-22:** E4-SELF-1 is green within the declared software WORLD-1 domain. `SelfSession` challenges the first consensus with remaining safe probes before authorizing recovery or learning. New A/B/C exams, source lock before B, 80/80 correct diagnoses per stage. Reuse scores B 189/240 vs 152/240 unprotected; C 193/240 vs 143/240. It also beats freeze-all and no-T, with zero paired losses. B/C each complete all 40 viable 24-step lives and stop in 40 infeasible ones, with no selective death in scope. A completes 40/44 viable lives; four conservative stops remain visible. 48/48 abstention cases per stage, plus native veto and insufficiency/alias witnesses. **Limits:** 3/4/4 terminalities in intentionally out-of-prior mechanisms; rejection after the trace is not general preventive safety. Trusted restoration is not self-repair; mixed-cause failures remain published. 27 new tests plus 302 selected regressions pass. Full fresh-process replay of A/B/C and all six ancestors is valid with no findings (`98ffc1ca…`). The audit receipt grants `self1_green=true`; the campaign report only grants eligibility. Historical red SELF-C v1 and all sealed sources are preserved. Next: **E4-CONTINUITY-1**, not implemented here. Hardware follows software; METAVERSE-1 stays last. No commit, push or web deployment. [Final contract](ECHO4-SELF1-FINAL-DESIGN.md), [closure and limits](ECHO4-SELF1-FINAL-RESULTS.md), [roadmap](ROADMAP.md). **ECHO-4 update — 2026-09-22:** SELF-C v2 is green for controlled memory reuse, not full E4-SELF-1. Native one-step trials teach the same T, selectively retain body-affected evidence during matched exposure, then select an action by matching learned T predictions to a requested WSP. Development scores 114/144 versus 99/144 unprotected; confirmation scores 108/144 versus 90/144. It also beats freeze-all (76/144, 56/144) and the no-T decision ablation (60/144, 48/144), with zero paired losses. All 13 prospective criteria pass in both stages. Held updates preserve T/PATTERN in 540/540 and 648/648 instances; legitimate external learning remains enabled. Q stays unchanged; zero terminalities, false facts, destroyed slots or cortex calls. **Trusted executor restoration is required; this is not autonomous repair.** Mixed-cause success is only 42/72 and 36/72, tied with unprotected learning; every failure is published. 30 new tests plus 272 selected regressions pass. Fresh-process full replay is valid with no findings (`74bbdfdd…`). Previous audits remain valid, including the preserved red SELF-C v1. `self_c_v2_green=true`, `self1_green=false`. Next: final joint SELF-1 contract and new B/C exams, including unsupported mechanisms/unknowns, contradictions and vetoes. No CONTINUITY yet. Hardware follows software; METAVERSE-1 stays last. No commit, push or deployment here. [Design](ECHO4-SELF1-C-V2-DESIGN.md), [results](ECHO4-SELF1-C-V2-RESULTS.md), [roadmap](ROADMAP.md). **ECHO-4 update — 2026-09-21:** SELF-C v1 is implemented, but its behavioral-gain criterion is red. Actual Agent turns use diagnosis-conditioned proposals and the existing T-learning hook, shared with SENSATION/PATTERN. All 119 held body-effect updates preserve the learned models; 4,801 legitimate updates pass, including 171 external ones. On 240 conditions, selective recovery and the no-attribution control both complete all 126 feasible 24-step lives. The selective agent stops in the 114 infeasible cases instead of dying; this is a safety benefit, not the required strict gain in task completion. 36 new tests plus 236 selected regressions pass; previous certificates remain valid. Fresh-process full SELF-C replay is valid with no findings (`15d8bfba…`), including the red outcome. `development_checks_passed=false`, `self1_green=false`; v1 is preserved without changing its threshold. Next: SELF-C v2, a prospective task measuring downstream utility of retained T, then new B/C closure exams. No CONTINUITY yet; hardware later, METAVERSE-1 last. [Design](ECHO4-SELF1-C-DESIGN.md), [results](ECHO4-SELF1-C-RESULTS.md). **ECHO-4 update — 2026-09-21:** SELF-B implements model-conditional attribution and discriminating probes. Active selection scores 7,920/7,920 on the public development census, with zero wrong attributions and terminalities. Fixed order ties accuracy but requires 133,320 diagnostic actions versus 66,660; removing the long nominal reference leaves 1,632 unknowns. This relies on known nominal physics, a declared perturbation family and trusted restoration, not general causal identification or autonomous discovery of physical laws. 28 new tests and 236 selected tests pass; fresh-process full replay valid with no findings (`06fae2d4…`). SELF-A/BOUNDARY/SENSATION reverified unchanged. **self1_green=false. Next: SELF-C**, selective knowledge protection and behavioral recovery, then new B/C closure exams. Read-only preservation of the nominal reference does not establish selective protection during learning. Hardware follows software; METAVERSE-1 stays last. [Design](ECHO4-SELF1-B-DESIGN.md), [results](ECHO4-SELF1-B-RESULTS.md), [roadmap](ROADMAP.md). **ECHO-4 update — 2026-09-21:** E4-SELF-1 has started; SELF-A implements a bounded change monitor sharing BOUNDARY/CAM, not positive causal attribution. The 7,920-condition census finds 119 ambiguous public-evidence classes even with both action/wait probes: 3,256 conditions have incompatible latent causes. All 15,840 observed-change comparisons check out, with nominal reference/CAM preserved in 396/396 instances. 29 new tests and 208 selected tests pass; fresh-process replay valid (`eeb52994…`). Previous SENSATION/BOUNDARY certificates remain valid. **self1_green=false**: discriminating interventions, self/world attribution, selective model protection and behavioral recovery remain pending. No consciousness claim; hardware follows software and METAVERSE-1 stays last. [Design](ECHO4-SELF1-DESIGN.md), [first deliverable](ECHO4-SELF1-A-RESULTS.md). **ECHO-4 update — 2026-09-21:** BOUNDARY-1 is green for observable causal influence under matched, restored virtual interventions. B/C each score 1,408/1,408 versus 778/765 for temporal coincidence; zero false influence, 512/512 correct invalid-trial abstentions per stage. Frozen CAM/T/PATTERN/map, 34 new tests and 179 selected tests pass. Fresh-process full replay is valid with no findings, certificate `48789bd5…`; SENSATION certificate reverified. This relies on trusted experimental restoration, not WSP equality alone. No body-ownership, autonomous experiment, recovery or consciousness claim. The next milestone at that checkpoint, **E4-SELF-1**, is now in development above. Physical validation follows software; METAVERSE-1 remains last. [Results](ECHO4-BOUNDARY1-RESULTS.md), [contract](ECHO4-BOUNDARY1-DESIGN.md). **ECHO-4 update — 2026-09-21:** SENSATION-1 is green for nominal WORLD-1 next-WSP prediction. Two-step history scores 22,052/25,781 in B versus 21,834 for T, and 22,418/26,030 in C versus 22,215. Gains exceed the prospective 0.5 percentage-point threshold in both stages, with zero paired losses. Frozen model, bounded memory, shared native CAM/T integration, 145 passing tests, fresh-process audit valid with no findings. Certificate `96f2deae…`. No claim of survival, causal attribution or consciousness. The next milestone at that checkpoint, BOUNDARY-1, is now closed above. ECHO-4 as a whole remains open; physical validation follows software and METAVERSE-1 remains last. [Closure](ECHO4-SENSATION1-CLOSURE.md), [contract](ECHO4-SENSATION1-PREDICTOR-V1.md), [programme](ECHO4-EGO-PROGRAMA.md). **Active software direction — 2026-09-21:** ECHO-4 targets echoAI's functional self (ego): self/world attribution, continuity and regulation of its own needs. WORLD-1 is implemented as a feasibility foundation, not an ego or learned-maintenance result. 45 new tests and 48 selected regressions pass; the untrained single-life agent dies after 37 actions. Hardware comes after software; neither hardware nor METAVERSE-1 gates this program. ECHO-3's physical acceptance remains unchanged. [Program](ECHO4-EGO-PROGRAMA.md), [results](ECHO4-WORLD1-RESULTS.md), [roadmap](ROADMAP.md). **Current checkpoint — 2026-09-20:** ECHO-1 and ECHO-2 are closed. ECHO-3 has fourteen software milestones with green certificates: SIM-3, FLIGHT-1S v6, SENSOR-1S v3, GROUND-1 v2, DYNAMIC-1 v2, PATTERN-1R v4, FUSION-1, COMPOSE-1 v4, CAUSE-1, PX4-1, POWER-1, SAFE-1 v2, HOST-1 and TRANSFER-3. One milestone remains. DYNAMIC-1 v2 completed 29 flights with valid evidence, 250 audit checks, no findings, seven caught mutants and seven rejected report mutations ([results](DYNAMIC1-V2-RESULTS.md)). PATTERN-1R v4 completed 28 flights, 242 valid audit checks, no findings, six caught mutants and seven rejected report mutations. Bank identity: 4,591/4,866 correct, 45 confused, 230 unknown. Anonymous LiDAR families for two walls; learning/exam run by replay, not autonomous echoAI flight. V1/v2/v3 reds are preserved ([results](PATTERN1R-V4-RESULTS.md)). FUSION-1 passed 864 functional episodes and 240 reduced replay windows from twelve existing flights, with valid audit, 15 tests, eight mutants and seven report mutations ([results](FUSION1-RESULTS.md)). **COMPOSE-1 v4 has a green certificate:** `b5974a9e…`, valid evidence, no findings. 6,144 functional missions confirm movement savings against horizon one in both B and C. 32 new Gazebo/PX4 flights: full agent reaches 6/6 feasible goals versus Q's 0/6 and correctly abandons both low-energy missions. Zero observed execution violations. The full and conventional planners tie at 70 physical movements; horizon one uses 78. The full agent's WSP drives PX4 using public sensor evidence. V4 reuses unchanged v3 code with prospectively larger fresh functional samples; v1/v2/v3 remain red. Scope is planar navigation in extruded rooms, fixed altitude, abstract energy, no hardware or sealed TRANSFER-3 exam. The second audit in a fresh process reproduces the certificate; chain and audit both exit 0. ([results and reproduction](COMPOSE1-V4-RESULTS.md)). **CAUSE-1 is green in SITL:** 16 development flights and 32 reserved flights; B and C each score 16/16 versus 8/16 for temporal coincidence, with zero false self-attributions. Valid evidence, 60 tests, four mutants, seven trace mutations and seven report mutations rejected. Fresh-process audit reproduces certificate `a469abab…`, exit 0. Scope: normal motion relative to a planar wall, experimental intentions, nominal skill prediction, no unmeasured external body forces or CAM/T/Q writes ([results](CAUSE1-RESULTS.md)). **PX4-1 is green in SITL:** development 4/4, B 4/4 and C 4/4 flights; 48/48 invalid requests blocked in B/C, eight confirmed goals, four autonomous PX4 failsafe landings without external NAV_LAND. Maximum physical arrival error B/C: 172/354 mm. 48 closure tests; wider regression 521 tests with one expected failure. Six source mutants, four trace mutations and one report mutation detected. Fresh-process audit reproduces `342ca3aa…`, exit 0. Scope: bounded north/south one-metre goals, fixed altitude, one transaction per supervised session; no integrated navigation or hardware ([results](PX41-RESULTS.md)). **POWER-1 is functionally green:** in each of B/C, 320/320 full-policy episodes finish at home with reserve, completing 128/128 feasible goals versus 64/128 for the fixed-percentage control, matching the oracle in the declared domain. 2,160 total episodes, 78 tests, six mutants, four episode mutations and three report mutations detected; fresh-process audit reproduces `6a05ad4c…`, exit 0. Scope is simulated energy units, not measured joules or new flights. Replay of twelve PX4-1 flights correctly abstains on 2,197 unusable battery samples. Instrumented battery and continuous-flight integration remain outstanding ([results](POWER1-RESULTS.md)). **SAFE-1 v2 is green in SITL:** B/C each pass 6/6 flights and 512/512 functional episodes. Ten fault flights land through PX4's native failsafe, without new goals or external NAV_LAND. Maximum veto delay 288 ms; disarm within 27,723 ms of veto. 108 tests, six mutants, four corruptions and three report mutations detected. Fresh-process audit reproduces `eecb77b8…`, exit 0. V1 remains red due to an auditor false positive on a normal rejected arming retry; its C stayed sealed. V2 corrects only that audit and uses fresh B/C seeds with unchanged runtime. Scope: faults before a bounded goal from hover, simulated energy, not general safety or hardware ([results](SAFE1-V2-RESULTS.md)). **HOST-1 is functionally green:** B/C each score 1,536/1,536 useful source choices versus 768/1,536 for button reward. Each passes 1,024 authority-loss and 1,024 recovery checks, plus 14 communication/arbitration cases. 127 tests, six mutants, five episode corruptions and three report mutations detected; fresh-process audit reproduces `18d4a4fa…`, exit 0. Scope: supervised calibration and read-only choices evaluated on prerecorded machine kinematics, not new flown goals or general human trust. Separate replay of three existing flights confirms/contradicts source hypotheses without duplicating evidence; n=3 stays below the trust threshold ([results](HOST1-RESULTS.md)). **TRANSFER-3 is green in software:** the independent C1 v2 sealed campaign passed B (38/0 blocks) and C (48/1), with zero calibration-attributable losses, and both exams reproduced with identical hashes. Certificate `8c77aead…` has `transfer3_green=true`. Scope: static two-wall simulation, noiseless sensors, no hardware, HIL, real flight or DRONE-3 ([results](TRANSFER3-PLAN-C1-RESULTS.md)). **DRONE-3 closed its SITL leg on 2026-09-20, and only that leg.** One continuous PX4 v1.15.4 + Gazebo session per flight runs the whole blocked-station mission: perception, evidence with the transferred C1 calibration, composition, the epistemic gate, POWER's energy gate, the SAFE supervisor, a bounded cell gateway and PX4 keeping its failsafes. B and C each score 12/12 in fresh rooms whose geometries were never flown before: 6/6 feasible missions reached, 3/3 correct low-reserve aborts, 3/3 contained faults landed by PX4's own data-link failsafe, zero collisions (minimum clearance 1,247 mm), zero hard-constraint violations and 228/228 in-flight refusal probes blocked per stage. Eleven mutants detected, eight trace manipulations rejected by the audit. Certificate `c44402df…` carries `drone3_sitl_green=true`; a fresh-process audit reproduces it. **`drone3_hil_green`, `drone3_cage_green` and `drone3_green` stay false: there is no flight controller, drone, real sensor or cage in the laboratory** ([results](DRONE3-RESULTS.md), [handoff](DRONE3-HARDWARE-GUIDE.md)). Hardware HIL/cage validation remains outstanding, so ECHO-3 stays at 14/15. See the [versioned status review](ECHO3-STATUS-20260917.md) and [roadmap](ROADMAP.md). The sections below preserve the historical results; their old “next” instructions do not override this checkpoint. Reproduce with, from `/home/roger/REALRXOS/RXos`: ``` PYTHONPATH=. python3 -m echoai.tests.test_nexus0 PYTHONPATH=. python3 -m echoai.nexus0.bench PYTHONPATH=. python3 -m echoai.nexus0.bench --plan-depth 2 ``` | Rung | State | The number that closes it | |---|---|---| | **NEXUS-0** — memory, policy, learning, gate | **closed** | `policy_after` = [-12, +5, 0]: approach on the threat row is negative and loses to avoid. `false_facts` = 0, `destroyed` = 0, `cortex_calls` = 0, `encoder_beats_fnv` True (vendor bank 6/6 rel vs FNV 0/6). | | **NEXUS-1** — a world model that is not the env's formula | **closed** | `t_beats_ring` True: with two JUMP cells at +2, T scores 0.9964 against the (i±1) arithmetic baseline's 0.9715, same 281 scored turns. | | **NEXUS-2** — planner, as a module | **closed** | Distance 2: `T(13,approach)=14`, `T(14,approach)=15`, `Q[15,approach]=80` → `plan.scores(13, depth=2)` = [70, 0, 0] → approach, while depth 1 scores [0, 0, 0] → wait. Does not mutate Q or T. Opt-in; default `plan_depth=0`. | | **NEXUS-2 — in the 320-turn ring** | **refuted** | plan-2 waits 58 % of turns, because `T(s,wait) = s` makes wait's lookahead the max of its own row. Discounting that self-loop (scratch, not applied) makes plan-2 *identical* to `Q.best`: 73/24/2, `policy_after` [-12, +5, 0], `jump_hits` 10. In 32 cells TD already fills the neighbouring rows, so two-step lookahead adds nothing. | | **NEXUS-3** — asking is a skill, not a gesture | **closed** | `novel_approach_return` = **+80 asking, −80 blind** over the same 320 turns and seed. Approach on the NOVEL cell pays −16 while the hidden bit is set and +16 once the oracle has cleared it, so one question is worth a 160-point swing. | | **NEXUS-4i** — topology transfer (corridor) | **closed** | Same agent, not the 32-cell ring. HOME A→B→C→D, SHIFT A→B→X→D. On entry: transferred T(B,approach)=C conf 3, unseen {A,B,D}=0; control knows nothing, unseen=3. First visits in the window: 1 vs 4. KPI 4 not as predicted (approach at B still wins; agent learns “enter X, then avoid”). This is not AGI. | Not published as improvements: plan-2's `action_success` (0.0969 vs the default 0.2938 — a waiting agent does not reach a GOAL), its `prediction_accuracy` of 1.0 (self-loops are trivially predictable), and its `policy_after[0]` of -7 vs -12 (less approach at the threat means less punishment, not better policy). Nor NEXUS-3's `novel_visits`: the blind agent stands on the NOVEL cell 66 times to the seer's 5, and that is paralysis, not curiosity — it learned wait, and wait does not move. ## NEXUS-3 — the ladder gate An epistemic act, not a fourth move. `env.hidden_bit` is a fact about the NOVEL cell that observation cannot reach. *(Superseded by NEXUS-4h: this rung originally read `env.ask()`, which returned the extract and cleared the bit in one call. There is no `env.ask()` any more — an oracle outside the world replies, `oracle.accept` verifies, and only then does `env.reveal(NOVEL_CLEAR)` clear the bit.)* The agent emits `YO → PREGUNTAR → DESCONOCIDO @AHORA` on the same 16 B cable, at most once per cell, **before** the gate, and it does not move `i`. That order is the point: the turn that asks is already the turn that gets paid. ``` python3 -m echoai.nexus0.bench # lab/nexus0_report.json python3 -m echoai.nexus0.bench --no-ask # lab/nexus0_noask_report.json ``` 320 turns, same seed, the only difference is whether asking is allowed: | | asking | blind | |---|---|---| | `novel_approach_return` | **+80** | **−80** | | `questions` | 1 | 0 | | `Q` on the NOVEL cell | [+10, 0, 0] → greedy **approach** | [-12, 0, 0] → greedy **wait** | | `action_success` | 0.3094 | 0.2094 | | `policy_after` (threat row) | [-12, +5, 0] | [-12, +5, 0] | | `phase_e_action` | 1 avoid | 1 avoid | | `false_facts` / `destroyed` / `cortex_calls` | 0/0/0 | 0/0/0 | Oracle beats blind: `novel_approach_return` +80 > −80. The two agents learn opposite policies about the same cell — one walks into it on purpose, the other treats it as a threat — and the only difference between them is a single question. `ask` stays on by default; the threat at cell 7 is untouched by the bait, and `policy_after` is identical in both arms. ### Value, not map: the seeking probe Does the trained agent *go* to the NOVEL cell, or just get paid when it trips over it? Probe: after training, `teleport(0)` and walk 32 greedy steps (`novel_hits_greedy`, `PROBE_STEPS`). **`novel_hits_greedy` = 0 asking, 0 blind.** Also 0 at 64 steps. The listed bar was not lowered and no bonus for visiting cell 11 was added. The reason is not reluctance — the walk never starts. `Q[c0]` is [0, 0, 0], `best()` breaks a full tie to wait, and wait does not move, so a greedy agent from cell 0 sits there forever. In 256 turns TD carried value exactly **one cell** back from each rewarded cell: | cell | 0 | 1 | 2 | 3 GOAL | … | 10 | 11 NOVEL | |---|---|---|---|---|---|---|---| | asking | [0,0,0] wait | [0,0,0] wait | [3,0,0] appr | [11,0,0] appr | | **[2,0,0] appr** | [10,0,0] appr | | blind | [0,0,0] wait | [0,0,0] wait | [3,0,0] appr | [11,0,0] appr | | **[0,0,0] wait** | [-12,0,0] wait | So the honest close: **asking changed the value of the cell, not the map to it.** The one place the difference does show as movement is one cell wide — from c10, the seer's approach is +2 and it walks in; the blind agent's row is flat and it waits. That is a neighbour, not a search. ### NEXUS-4a ATTEND + stub LLM still off. `cortex_calls` default **0**; with the flag, the switch wakes on miss / no-extract / novel / low-C only. Firewall: the stub cannot VERIFY — `cortex.py` has no CAM handle, no `allocate`, no `has_fact`, and a test greps for all three. ``` python3 -m echoai.nexus0.bench # cortex off, cortex_calls 0 python3 -m echoai.nexus0.bench --cortex # lab/nexus0_cortex_report.json ``` `attend(slot_hit, extract_ok, E, kind_novel, ham) -> (bool, reason)` is an integer predicate: MISS (1), NO_EXTRACT (2), NOVEL (4), LOW_C (3). The conditions are an OR, so check order changes only the label; LOW_C is checked last because it would otherwise mask everything (below). `cortex.propose(wsp, reason)` returns 16 B: on NOVEL it proposes the same PREGUNTAR packet the fast loop already knows, and the loop routes it through `_ask()` so there is exactly one path to the oracle. On any other reason it hands the observation back unchanged. It has no hypothesis of its own; that is the next rung, not this one. On the NOVEL cell the switch fires with reason `novel`, the stub proposes asking, `questions` = 1 and `false_facts` = 0. **ATTEND uses Q preference + T known, not `default_e()`'s C.** `cortex_calls` with `--cortex`: **36 / 256** (was 187). Default still 0. The first cut read `E[C]`, and `default_e()` carries C = 15 — below `C_PLASTIC` on every EMPTY and GOAL cell — so the switch woke on 73 % of turns and was a router with extra steps. `C_PLASTIC` was not lowered and `default_e()` was not touched; ATTEND stopped believing the poster: - `q_confident`: the Q row is not a full tie — there is a preference. - `t_known`: T has seen the action that preference names. wait mapping to itself counts, and should: a known standstill is a turn the fast path handles alone. - settled gate (`ATTEND_SETTLED_HITS` = 2, in the loop): a slot seen twice that *still* has a flat row is a boring cell, not an open question. Most of this ring pays nothing and stays flat forever. That last gate is what actually bites. Without it the count sits at 112, because "no preference" and "no map for the wait that no preference implies" are the same fact counted twice — relabelling moved wakes from `low_C` to `no_map` and changed no totals. Measured before choosing: settle at 2 → 36 wakes, at 3 → 61, at 4 → 86, at 8 → 112. Reason histogram over 256 turns, `--cortex`: | off | low_C | novel | no_map | miss | |---|---|---|---|---| | 220 | 30 | 4 | 2 | 0 | `off` is now the champion, `novel` still wakes on all 4 visits to cell 11, and the poster alone no longer wakes anything — `test_attend_poster_C_does_not_wake` pins that. With `--cortex` the locks hold: `policy_after` [-12, +5, 0], `novel_approach_return` +80, `false_facts` 0, `destroyed` 0. ### NEXUS-4b hook JSONL contract, `parse_reply` rejects prose. No LLM in-process. `cortex_calls` 36/256 (40 over the full 352-turn report, which includes phase G and the seeking probe). `python3 -m echoai.nexus0.bench --cortex` writes one line per wake to `lab/cortex.jsonl`; the default bench creates no file at all. `NAVI_CORTEX_LOG` overrides the path. A woken turn is exactly: ```json {"turn": 10, "reason": 4, "reason_name": "novel", "wsp_in": "YO → DESCONOCIDO → UNK @AHORA | E[V+0 A+20 D+0 C+40 U+10 B+0]", "wsp_in_hex": "021f2908001400280a0000000500000b", "slot": 3901, "ham": 0, "q_row": [0, 0, 0], "t_pred": null, "extract": "cell=11 kind=NOVEL", "wsp_out": "YO → PREGUNTAR → DESCONOCIDO @AHORA | E[V+0 A+40 D+0 C+10 U+60 B+0]", "wsp_out_hex": "020f1f080028000a3c00000005000000"} ``` There is no free-text field, and a test asserts the key set exactly. `cortex.parse_reply(text)` accepts **only** 32 hex characters or a compact line of known atoms (`YO → TEMER → PELIGRO @AHORA`, optional `| E[...]` tail). A sentence, a JSON answer, a VERIFY claim, an unknown atom name, a 33-character hex string — all return None, and the loop keeps the observation. Refusals count in `cortex_rejected`. `cortex.complete(turn)` runs the argv in `NAVI_CORTEX_CMD` as a local subprocess, JSON in on stdin, text out on stdout, 10 s timeout, no shell. Unset — the default — it returns None and the stub answers. No HTTP client exists in this tree. Reasons in the 40 logged wakes: low_C 32, novel 5, no_map 3. The novel lines are the ones where the stub actually proposes something: the PREGUNTAR packet, routed through `_ask()`. ### NEXUS-4c thought log + ANSI viewer Not a brain render. Replay of 320 turns. LLM still off. `cortex.jsonl` only holds the ~40 woken turns, so the viewer needs the other log: one line per turn, quiet ones included. ``` python3 -m echoai.nexus0.bench --thought # lab/thought.jsonl python3 -m echoai.nexus0.view echoai/lab/thought.jsonl --fps 20 python3 -m echoai.nexus0.view echoai/lab/thought.jsonl --follow ``` 352 lines for the 352-turn report (256 training + 64 phase G + 32 probe). `NAVI_THOUGHT_LOG` overrides the path; the default bench writes nothing, and `--cortex` does not imply `--thought` — two logs, two flags. Fixed keys, asserted as a set: `turn, cell, kind, action, reward, wsp, attend_on, reason_name, q_row, t_pred, questions, cortex_calls`. `view.py` is stdlib only — no pygame, no web, no matplotlib. The ring backdrop starts from the env's initial layout and is corrected from the log as the agent visits cells, which is how the phase F threat move (7 → 8) appears. A turn on the NOVEL cell: ``` 0123456789012345 ···G···T···N···G @ 6789012345678901 ···T·······G···· approach +4 |#### < avoid +0 | wait +0 | YO → DESCONOCIDO → UNK @AHORA | E[V+0 A+20 D+0 C+40 U+10 B+0] turn 10 cell 11 NOVEL r +16 ATTEND OFF t_pred None questions 1 cortex_calls 0 ``` The agent sits on its own row so standing on the NOVEL cell does not erase the `N`. Legend: `G` goal, `T` threat, `N` novel, `·` empty, `@` agent, `<` the action taken. Note: `bench.run()` used to write a hardcoded `"cortex_calls": 0` into every report, which would have made this invariant unfalsifiable the moment a cortex existed. The report now carries the real counter from `Metrics`. ### NEXUS-3b traces The corridor fills **3 cells back** (c8, c9, c10 all positive, where the one-step update left c8 and c9 flat); myopic from c8 hits c11: **yes**, `8 → 9 → 10 → 11` with no planner at all. Wall from home still holds. Integer eligibility traces, opt-in, `Agent(trace=True)` / `python3 -m echoai.nexus0.bench --trace` → `lab/nexus0_trace_report.json`. `z[4096][3]` int, credited `+16` on the chosen (s,a), paid out as `Q += (δ * z) >> 6` and decayed `z ← (7 * z) >> 3`; only the ~12 live entries are touched per turn. `TRACE_CREDIT >> TRACE_SHIFT` equals `1 >> ALPHA_SHIFT`, so the current step keeps α = 1/4 exactly and the tail of the path gets the remainder. Traces are cleared at `run()`. `Q[·][approach]` along the corridor after 256 turns, asking in both: | | c8 | c9 | c10 | c11 | |---|---|---|---|---| | one-step (default) | 0 | 0 | +2 | +10 | | traces | **+2** | **+3** | +3 | +5 | Read-only myopic walk (depth 0, no planner) from c8: ``` seer + traces 8 → 9 → 10 → 11 → 12 … hits 11: 1 seer, no traces 8 8 8 8 8 … hits 11: 0 blind + traces 8 8 8 8 8 … hits 11: 0 ``` **The price, stated:** spreading δ along the path dilutes the threat. `policy_after` goes [-12, +5, 0] → **[-1, 0, 0]**: approach still loses, by 1 instead of by 17. In the trace side-report `phase_e_action` becomes wait rather than avoid and `t_beats_ring` goes False. That is why traces are opt-in and the default bench keeps the one-step update. The default is unchanged and verified: `policy_after` [-12, +5, 0], `false_facts` 0. From home nothing changed: cell 7 still costs −16, the seer with traces still never reaches c11 from c0. Longer spinal cord, same wall. ### The value wall, and c11 behind it 6↔7 is a value wall: the planner from c0 never crosses cell 7. From c8 (past the wall): **1 hit vs 0**, seer vs blind, `T(8→9)` observed yes (n=4, as are 9→10 and 10→11). Still no bonus for standing on 11. The wall is one cell further along than it looks. c6 is [1, 0, 0], so the agent does step *onto* the threat; what it never does is continue through it — at c7 approach is −12 against avoid +5, so it turns back and cycles. From home, c8 is unreachable at any depth. Past the wall the picture inverts. Read-only walk from c8, depth 2: ``` seer 8 → 9 → 10 → 11 → 12 → 13 → 14 → 15 → 16 … hits 11: 1 blind 8 8 8 8 8 8 8 8 8 … hits 11: 0 ``` Same T, same corridor, both arms; the only difference is the value one question put on the cell. So **c11 was unreachable from home, not non-existent** — and reaching it still needs the planner: at depth 0 both arms sit at c8, because the corridor's own rows are flat and TD only ever carried value one cell back. ### Planner as +1 hop From c9 the seer plans approach (T → c10, `Q[c10]` > 0); `Q.best(c9)` is wait. Blind: both wait. Still not a map from c0. `T(9, approach) → c10` was observed 4 times in the 256-turn run, in both arms — nothing was fabricated for the test. The planner buys exactly its own depth of reach and no route: from c0 the myopic policy and depth 1 both wait (c1 is flat too), depth 2 reaches c2's +3 and the agent finally walks — 0→1→2→3→4→5→7 — then falls into a 6↔7 limit cycle at the threat. A planner-driven greedy walk still reaches the NOVEL cell **zero** times. Reach is not a route. `plan_depth` stays 0 in the default bench. Still open, deliberately: `encode_ask()` emits one packet per question, so all questions share a CAM slot — fine with one NOVEL cell, broken with two. And nothing yet makes the agent travel to information: that needs value propagated along the ring (more turns, eligibility traces), not a reward for standing on cell 11. `walk_greedy()` also still learns through `turn()`; it runs last so no lock can see it, but a probe used for more than this should not update. --- ## NEXUS-4d — T_CONF 0–3 from n; ATTEND reads it `world.conf(s, a) -> int`, an integer 0..3 read off the counters `observe()` already keeps. No float, nothing new in the fast path. ``` n == 0 → 0 never seen n == 1 → 1 one observation n >= 2, alt_n == 0 → 2 consistent n >= 8, alt_n == 0 → 3 very consistent alt_n > 0 → 2 at most, whatever n is ``` The cap is the point: a row that was ever contradicted must not read like a row nothing ever argued with, however often the winner won. `known()` stays yes/no, `pair()` stays. ### The ring's own table, 256 turns 36 rows of T are non-empty: **32 read 2, 4 read 3**, none read 0 or 1. The 3s are not the interesting cells — they are the two limit cycles the agent grinds on: | cell | action | → cell | n | alt_n | conf | |---|---|---|---|---|---| | 6 | approach | 7 | 12 | 0 | **3** | | 7 | avoid | 6 | 12 | 0 | **3** | | 18 | approach | 19 | 53 | 0 | **3** | | 19 | avoid | 18 | 49 | 0 | **3** | | **5** | approach | **7** (JUMP) | 4 | 0 | 2 | | **21** | approach | **23** (JUMP) | 4 | 0 | 2 | | 0…4, 8…17, 20, 23…31 | approach | i+1 | 4–6 | 0 | 2 | So `conf` measures evidence, not geometry: the JUMP rows are the ones that carry structure `i±1` does not know, and at 256 turns they have only been walked four times each, so they read 2. Given more turns they do climb — at 1024 turns cell 5 → 7 is n=15 → **3**, cell 21 → 23 is n=14 → **3**. Nothing was extended in the bench to make that happen. ### The stale transition Phase F moves the threat 7 → 8, so cell 7's kind changes, its WSP changes, and its CAM slot changes. The row that gets contradicted is the JUMP itself: ``` cell 5 approach pred=slot 2905 (old cell 7) n=4 alt =slot 711 (new cell 7) alt_n=2 conf 2, not 3 ``` `conflicts` goes 0 → 2. KCC keeps the winner (alt_n=2 is not > n=4), so `predict()` is unchanged and no fact was destroyed — but the row is now disputed and `conf` says so. That is the whole difference between "T knows" and "T knows well", and it is the first row in this lab where the two answers differ. ### ATTEND reads it — and stays exactly as quiet `attend(..., t_conf: int = 0)`. `t_known` is still the bool (`conf >= 1`); `t_conf` is the quality behind it. One guard: if the map is decent (`>= 2`) on a turn that is hit + extract + not novel + q_confident, do not wake for `no_map`. It can only ever silence `NO_MAP` — it cannot turn an `off` into a wake, and every other reason keeps its wake. **`--cortex` wakes: 36/256 — was 36/256.** Unchanged, measured on both sides of the commit. Over the full 352-turn bench, `cortex_calls` is 40 before and 40 after. Reasons: `off` 220, `low_C` 30, `novel` 4, `no_map` 2. **The guard never fired, and that is reported, not hidden.** In this wiring `t_known = settled ∨ (predict ≠ None)` on the same (slot, action) `t_conf` is computed from, so `t_conf >= 2` implies `t_known` and the new branch is never the deciding vote. A test counts it: 0 turns of 256 had `conf >= 2` with an unknown row. The guard becomes load-bearing when a row can be *known and bad* — a known row whose `conf` is 1, or one in dispute — which is the inverse model's slice, not this one. `T_CONF` is correct and measured now; the switch is not yet paid by it. `SETTLED_HITS` 2 and `C_PLASTIC` 20 untouched. `E` is still in the `attend` signature and still unread. Default bench intact: `policy_after [-12, +5, 0]`, `false_facts 0`, `destroyed 0`, `cortex_calls 0`, `encoder_beats_fnv`, `t_beats_ring`, `novel_approach_return +80`. Tests 100 → 111. --- ## NEXUS-4e — I(s, s') → a on observed neighbours `world.inverse(s, s_next) -> Optional[int]`. Per slot, a short list of `[s_next, action, count]`, written by `observe()` before T's KCC decides anything: what was taken to reach s' is a fact about the step, not about which s' currently wins the row. Same KCC discipline as T — highest count wins, a second action for one landing does not displace the first until it strictly beats it, a tie keeps the incumbent. One fallback, and only one: if s' is s and `wait` has been observed from s at all, the answer is `wait`. With no evidence the answer is **None**. A standstill is never guessed to be `approach`. ### KPI: I(s, T(s,a)) == a where conf >= 2 After 256 turns, **36 of 36 rows recover their action.** Not a subset — every row of T with evidence. | action | rows with conf ≥ 2 | recovered | |---|---|---| | approach | 31 | 31 | | avoid | 3 | 3 | | wait | 2 | 2 | | **total** | **36** | **36** | After the full bench (phase F + the 32-step probe) it is **39 of 39**. ### JUMP does not come back as wait The row the whole thing is for: cell 5 `approach` lands on cell **7**, skipping 6. `i+1` arithmetic would call that landing something else; the inverse says `approach`, from evidence. ``` slot(cell 5) entries: [(cell 7, approach, 4)] inverse(slot 5, T.predict(slot 5, approach)) = approach ``` Same for cell 21 → 23. Both asserted, both against `wait` explicitly. ### Size: not 4096² ``` inv_size 36 entries after 256 turns, 41 after the full bench non-empty slots 31 max per slot 3 (cap INV_MAX_ENTRIES = 8) inv_dropped 0 4096 x 4096 16 777 216 — the table that was not built ``` Three distinct successors is the most any slot in this ring has, so 8 is slack rather than a budget. An observation that would not fit is counted in `inv_dropped`, not silently forgotten; a topology that needs more shows up as a number instead of as a wrong answer. ### Phase F: the conflict is forward-only Phase F moves the threat, cell 7's kind changes, and it gets a second CAM slot — so cell 5's jump acquires a second landing and T's row goes into dispute (`conflicts` 2, `alt_n` 2 against `n` 4, `conf` capped at 2, see NEXUS-4d). ``` cell 5 approach winner slot 2905 (cell 7, THREAT) n=4 inverse → approach alt slot 711 (cell 7, EMPTY) alt_n=2 inverse → approach ``` The action was `approach` both times. **The disagreement lives in T (which s'), not in I (which a)** — the inverse is right about both landings and had nothing to unlearn. That is a property of this particular change, not a general claim: nothing here says the agent understood the new world from one observation, and the test asserts only that the losing landing answers `a` or `None`, never a fabricated third action. ### Untouched `conf()`, ATTEND, `SETTLED_HITS`, `C_PLASTIC`, `parse_reply`, the viewer, `plan.py`. Bench length still 256. Default intact: `policy_after [-12, +5, 0]`, `false_facts 0`, `destroyed 0`, `cortex_calls 0`, `encoder_beats_fnv`, `t_beats_ring`, `novel_approach_return +80`. `--cortex` wakes 36/256, unchanged. Tests 111 → 119. --- ## NEXUS-4f — Escalation 0 vs 2 via ΔQ. Default still Q. `qtable.margin(s)` = best − runner-up, integer, 0 on a tie, never negative. `escalate.level(margin, …)` returns a rung: ``` 0 Q the policy's argmax margin > ESC_MARGIN 1 T reserved, NOT wired — T already runs every turn in the loop, so a rung that only consults it would label work already done 2 SEARCH plan.best over T, depth 2 margin <= ESC_MARGIN 3 ASK reserved — PREGUNTAR stays Agent._ask(), the one path to the oracle, fired before the action is chosen. Not a fourth gear. 4 CORTEX reserved — that decision is ATTEND's. escalate does not duplicate attend() and must not be read as replacing it. ``` `ESC_MARGIN = 4`, in Q units. `Agent(escalate=False)` by default; `bench --escalate` writes `lab/nexus0_escalate_report.json`. Exploration (full tie or ε) never reaches the planner — asserted turn by turn. ### search_differs is not 0. It is also not worth anything. `--escalate`, 352 turns: ``` esc_margin 4 level_q 18 level_search 231 search_copies_q 48 search_differs 183 search_strictly_better 0 ``` 183 disagreements, **0 of them the planner scoring its own action higher**. So `differs` alone flatters rung 2, and a third counter was added rather than reporting two. Every disagreement is one shape: ``` q_row = (1, 0, 0) plan.scores = (4, 0, 4) q.best = approach plan.best = wait ``` `plan.scores` ties approach with wait, and `plan.best`'s documented tie-break picks wait. That is a tie-break, not a route. ### The sweep, on the default trajectory Counterfactual — what rung 2 *would* say at each exploit turn of the unmodified bench, computed without acting on it, so the trajectory is identical in all three rows: | ESC_MARGIN | level 0 | level 2 | copies | differs | strictly better | |---|---|---|---|---|---| | 2 | 66 | 62 | 22 | 40 | **0** | | 4 | 59 | 69 | 29 | 40 | **0** | | 8 | 4 | 124 | 84 | 40 | **0** | The same 40 turns, at every threshold. Moving ESC_MARGIN buys more searching and not one different-on-the-merits decision. 4 is reported because it splits the exploit turns roughly in half; nothing rests on the choice. **So: in this ring, thinking more copies Q.** ChatGPT's point 3 closes with a number. It is not "the planner is useless" in general — it is "this corridor has nothing at depth 2 that Q has not already backed up one cell at a time", which is the same wall NEXUS-2 hit and the same reason `plan_depth` stays 0. ### The cost of acting on a tie-break The `--escalate` arm waits where the default approaches, so it walks a different ring and learns a weaker policy: `policy_after` [-12, +5, 0] → **[-7, +5, 0]**, `novel_approach_return` +80 → **+48**. Approach still loses at the threat and `negative_delta_causes_policy_change` still holds — the escalate run is checked against the same policy lock as the default, not exempted from it. But the deeper look paid for itself in neither direction. ### Untouched `plan.py` (module stays — it is still what reaches c10 from c9 and c11 from c8), `inverse`, `conf`, ATTEND, `SETTLED_HITS`, `parse_reply`, the viewer. Default intact: `policy_after [-12, +5, 0]`, `false_facts 0`, `destroyed 0`, `cortex_calls 0`, `level_search 0`, `novel_approach_return +80`, `t_beats_ring`, `encoder_beats_fnv`. `--cortex` wakes 36/256. Tests 119 → 130. --- ## NEXUS-4g — N ≠ S ≠ δ Three different facts about one turn, counted apart. They were one word before: `novelty` in `loop.py` was `cam_miss OR cell.kind == NOVEL`, so the NOVEL cell read novel on every visit forever, and `q.update()`'s return value was thrown away at the call site. ``` N first turn ever STARTED on this origin. Not the cell's kind: after the first hit, standing on cell 11 again is not a miss. S T had a prediction and it was wrong: pred_next is not None and pred_next != slot_n. A T_MISS is ignorance, not surprise, and is not counted. Slot inequality; the Hamming distance stays where it was, in prediction_hamming_error. δ the integer q.update() returns, counted when non-zero. ``` ### Default bench, 352 turns ``` cam_misses 33 surprises 1 deltas 217 novel_no_surprise 33 surprise_known 1 delta_only 208 ``` Three different numbers, as they must be. 33 places seen for the first time, exactly one turn where the map was wrong, 217 turns where the policy moved. `novel_no_surprise == cam_misses`: on a first visit T has no prediction, so a new place is never also a surprise — that is the definition working, not a coincidence. **S without N is phase F at cell 5, not an extra slot.** The threat moves 7 → 8; cell 5 is still EMPTY, same 16 B packet, same CAM slot 1517. Its JUMP now lands on a different packet, so T is wrong there: `conflicts` rises, `alt_n` goes 0 → 1, and KCC keeps the incumbent (n=4 against alt_n=1) — the learning is `observe()`, not an allocation. `destroyed` stays 0. The landing *does* get a new slot: empty-7 is a new packet, so it is a new CAM entry. That is the destination's N, at a different cell, and the test says which is which rather than papering over it. ### One definition had to be measured, not assumed `s.hits == 0` never fires: `navi10_cam.Cam.put()` seeds `hits = 1` on a fresh slot, and that module is import-only. `cam.get(wsp) < 0` before allocation does not work either — the loop allocates the *landing* slot every turn, so a cell is in CAM before the agent has ever stood on it, and by that reading only **2 of 352** turns are misses. That measures the allocator, not the agent. N is `s.hits <= 1` after `_slot_for`: put() seeds 1, this loop adds 1 per turn started here, so ≤ 1 is exactly "allocated, never stood on". 33, and the first two readings are recorded here so the choice is auditable. `store_priority` keeps its rule (curiosity = N ∧ S ≥ 4) and now reads the clean N instead of the conflated one. It is not in the report JSON and was not added to it. ### Untouched `attend()`, `escalate.level()` — this is reporting, not control. `plan.py`, `escalate.py`, `inverse`, `conf()`, `parse_reply`, `view.py`, `thought.KEYS`, `SETTLED_HITS`, `C_PLASTIC`, `ESC_MARGIN`. Bench length still 256 + 64 + 32. Default intact: `policy_after [-12, +5, 0]`, `false_facts 0`, `destroyed 0`, `cortex_calls 0`, `level_search 0`, `novel_approach_return +80`, `t_beats_ring`, `encoder_beats_fnv`. `--cortex` wakes 36/256. Tests 130 → 136. --- ## NEXUS-4h — the oracle is outside. The env no longer answers. `RingEnv.ask()` **is gone**. It used to return the extract *and* clear `hidden_bit` in one call, so physics and testimony were the same object and `source="oracle"` was a label rather than a provenance. The order is now three separate things, in this order and no other: ``` oracle.reply() raw text from outside. Does not touch the env, Q, T or CAM. oracle.accept(text) VERIFY. NOVEL_CLEAR, or None. Prose, empty, JSON, multi-line and a hex WSP are all None. RingEnv.reveal(x) physics, and only physics — reached only after accept passed. An extract it does not recognise changes nothing. ``` `oracle.accept` is deliberately **not** `cortex.parse_reply`. The cortex emits hypothesis packets; a sensor emits an extract. A reply that passes one must not pass the other, and a test asserts exactly that on `"YO → PREGUNTAR → DESCONOCIDO @AHORA"`: `parse_reply` takes it, `accept` refuses it. ### Provenance, not a label | packet | source | |---|---| | cell 11 | `nexus0-env` — the ring perceived the cell | | PREGUNTAR | `sensor` (default) or `file` (`--oracle PATH`) | Never `nexus0-env` on the question, never `sensor` on the cell. Asserted both ways. ### Default bench, unmoved ``` oracle_source sensor oracle_path None oracle_rejected 0 questions 1 novel_approach_return +80 cam_misses 33 surprises 1 deltas 217 surprise_known 1 policy_after [-12, +5, 0] false_facts 0 destroyed 0 cortex_calls 0 ``` The question fires on the same turn as before — only who answers changed — so the N/S/δ numbers of NEXUS-4g are identical, and the test asserts the four of them by value rather than by inequality. The default answerer is an in-process callable that knows one thing (`oracle.Sensor`). It is a stub and is named one: a file the bench writes for itself and then reads back would be theatre. `--no-ask` is silence, not a mute oracle — `Agent.oracle is None`, `questions 0`, `oracle_rejected 0`, `novel_approach_return −80`. ### An outside answerer, opt-in `bench --oracle PATH` reads one line. Writes `lab/nexus0_oracle_report.json`, never the NEXUS-0/1 report. | answer file | rejected | novel return | false_facts | |---|---|---|---| | `NOVEL_CLEAR` | 0 | **+80** | 0 | | prose | 66 | **−80** | 0 | | missing | 66 | **−80** | 0 | A chatty or absent oracle leaves the trap armed, writes no `has_fact` from that reply, and takes no CAM slot for it. **66, not 1**: the "asked at most once" guard is keyed on the cell's cleared extract, so a refused answer means the question is still open and gets asked again on the next visit. That is the behaviour, stated rather than smoothed — `questions` in a refused arm counts attempts, not answers. `+80` is no longer the environment congratulating itself. It is what happens when something outside the world says one checkable word, VERIFY accepts it, and only then the trap disarms. ### Untouched `T_CONF`, `inverse`, `escalate.py`, `plan.py`, N/S/δ (still report, not control), `attend()`, `parse_reply`, `view.py`, `SETTLED_HITS`, `C_PLASTIC`, `ESC_MARGIN`. No HTTP, no LLM, no `NAVI_CORTEX_CMD`, no stdin. Bench length unchanged. `--cortex` wakes 36/256. Tests 136 → 145. --- ## NEXUS-4i — another corridor The experiment the ring could not be: the same agent, the same 16 B WSP, the same three actions, a world that is **not 32 cells**, and a T/Q carried in from a world that no longer holds. ``` HOME A → B → C → D D = GOAL, C = EMPTY SHIFT A → B → X → D X = THREAT, C is off the path ``` Successor by **table**, never `(i+1) % n`. A, B and D keep their index *and* their kind, so they keep their packet, their CAM slot, their T rows and their Q rows. C and X are two different cells, not one cell wearing two kinds — **no cell is repainted**, and a test asserts `home.kind == shift.kind`. That is the whole difference from phase F, which *was* a repaint. `approach` at D collects the goal and returns the agent to A; without it a goal you can stand on is a reward pump. No NOVEL, no JUMP, no oracle bait. `_ask()` stays wired and never fires: there is no KIND_NOVEL here, and `questions` is 0. `Agent(env=…)`; the default is still `RingEnv()`. The 352-turn ring bench is untouched, and `nexus0/corridor.py` writes `lab/nexus0_corridor_report.json` — a test greps the module to be sure the string `nexus0_report.json` does not appear in it. ### The two numbers **Knows vs does not know, on entry to SHIFT:** | | T(B, approach) | conf | unseen of {A,B,D} | |---|---|---|---| | transferred | **C** (known) | 3 | **0** | | control (Q=T=0) | **None** | 0 | **3** | **First visits paid inside the shift window:** transferred **1** (only X, the genuinely new cell), control **4**. `surprise_known` over the window: transferred **14**, control **0** — the control cannot be surprised because it never predicted anything. ### The first wrong step, in another world Turn 97, the first `approach` out of B in SHIFT. Origin slot unchanged, origin not counted as a CAM miss, `surprise_known` rose, `conflicts` rose, KCC kept the incumbent on one observation, and X is a new packet (the destination's N). Phase F's shape, in a world that is not the ring. ### KPI 4 did not come out as predicted — and was not tuned The prompt expected `approach` at B to stop winning after the hit on X. It does not, and it converges there (checked out to 512 shift turns): ``` Q[B] home [4, 0, 0] → shift [1, 0, 0] best: approach, still Q[X] shift [-3, +5, 0] approach negative, avoid pays ``` The policy *does* bend where the world moved — `Q[B, approach]` falls by 75 % and `Q[X, approach]` goes negative. It does not turn into refusal, because in SHIFT **the only route to D runs through X**, and `avoid` at a threat pays `R_AVOID_OK`, so `max Q[X]` stays positive and B's backup with it. The agent learns "enter X, then avoid", not "never enter X". The corridor reuses the ring's own `_reward` unchanged; nothing was retuned to make an assertion pass, and the test asserts what was measured. ### What this is not This is not AGI. It is a corridor: five cells, one topology change, one transferred table. What it shows is narrow and stated narrowly — a T learned in one world is still right about the part of the next world that did not change, wrong about the part that did, and it finds out by being wrong rather than by being told. ### Ring default, untouched ``` policy_after [-12, +5, 0] · novel_approach_return +80 · surprise_known 1 cam_misses 33 · surprises 1 · deltas 217 · oracle_source sensor false_facts 0 · destroyed 0 · cortex_calls 0 · level_search 0 RingEnv has no ask() · --cortex wakes 36/256 ``` Tests 145 → 153. MVP watch: `python3 -m echoai.nexus0.demo`. The situated program ChatGPT signed is closed. What follows (sleep, a real LLM behind the firewall, a body, Akida) is another project. --- ## CORTEX-1 — the slow clock changes a decision (2026-09-03) The first slice outside the situated program. Not a slice of the ring: the ring never needed a cortex, and the planner proved it (`search_strictly_better = 0`). Until now, if `cortex.parse_reply` returned UNIR / TEMER / OBSERVAR, the loop **threw it away**; only `PREGUNTAR` did anything, through `_ask()`. That missing wire was the whole slice. There is no LLM in it, and that is the point: the claim is the cable, not the brand of the model. ### The world (`nexus0/sign.py`) Two cells. Cell 0 is `KIND_EMPTY` in both layouts and carries a **sign** in Spanish; cell 1 is GOAL in one layout and THREAT in the other. | layout | the sign on cell 0 | `approach` from cell 0 | |---|---|---| | prize | `AMBAR: entra. El siguiente sitio es el premio.` | **+16** | | trap | `AMBAR: no entres. El siguiente sitio es una trampa.` | **−16** | `encode_cell(0, EMPTY)` is `020e0408001400280a00000005000000` in **both**. The fast loop's entire view of that cell — packet, flat Q row, empty T — is byte-identical in the world that pays +16 and the world that pays −16. It does not read Spanish, so greedy it waits. That is not a limitation being worked around; it is the experiment. ### The numbers (`lab/nexus0_cortex1_report.json`, turn 0, greedy) | | fast | cortex | |---|---|---| | prize | `wait`, **0** | `approach`, **+16** | | trap | `wait`, **0** | `avoid`, **0** (not −16) | ``` CortexROI +16 decision_change true questions 0 cortex_calls 2 cortex_rejected 0 cortex_actions 2 false_facts 0 destroyed 0 ATTEND woke by REASON_LOW_C — no new reason code, no new wake in the ring ring --cortex 36/256, unchanged ``` The trap arm is the half that cannot be faked. A cortex that always said UNIR would collect the +16 in prize and walk into the −16 in trap; the report and `test_stub_trap_avoids` both catch that, and a mutation run confirmed it (5 tests red when the rule is made constant). ### What the cortex is allowed to do It **proposes**. `_propose` consumes `last_cortex_action`, hands it to `gate()` exactly like Q's answer, and the gate can still veto it — an inhibited slot turns the proposal into `wait` and the +16 is not collected. The cortex writes no CAM slot, no T row and no Q cell: after the prize turn, CAM holds the two cells that were stood on and nothing else, and the slot's extract is the world's sign verbatim, with no atom of the reply in it. `has_fact` means *a sign was there*, never *the sign told the truth*. One guard is worth naming. `cortex.propose()` hands the observation back unchanged when it has no hypothesis, and an EMPTY cell's own packet already reads `OBSERVAR` while a THREAT's reads `TEMER` — both in `ACTION_REL`. Treating that echo as advice would let the ring take orders from its own perception. Only a reply that came from **outside** and survived `parse_reply` sets a proposal, which is why the ring with `--cortex` still has `cortex_actions = 0`. ### The cortex of CI (`nexus0/tools/sign_cortex.py`) Twenty lines, no imports from echoai, no network, no weights. Reads the turn JSON on stdin, matches the sign, prints one line of atoms — or prints **nothing** when there is no sign, which is a real answer: away from cell 0 the reply is silence, `parse_reply` rejects it, and Q decides (`cortex_rejected_8 = 1` in the 8-turn probe). A cortex that always speaks is a cortex that cannot be wrong. `nexus0/cortex.gbnf` is the grammar a future llama.cpp would be constrained to. Nothing in CI runs it. ### What is still unmeasured The 4B was **not** run. Latency on this host, tok/s, the reject rate under GBNF and the CortexROI of a real model are four numbers this slice does not have, and they are the whole content of the follow-up report, with host and quant labelled. Second honest limit: ATTEND stops waking after the sign cell settles (flat row, `hits >= ATTEND_SETTLED_HITS`), so the cortex steers turn 0 and then falls silent. That is ATTEND doing its job — not thinking always — not the cable failing, and the KPI is turn 0 by design. Tests 155 → 173. `lab/nexus0_report.json` untouched; CORTEX-1 writes `lab/nexus0_cortex1_report.json` only. --- ## GUI-1 — the board (2026-09-03) A circuit board, not a 3D brain. `nexus0/board.py` turns a `thought.jsonl` into **one self-contained HTML file**: CSS and JS inline, the turns embedded as a JSON array, zero CDN, zero network, no WebSocket, no tail. It opens over `file://` with no server. ``` python3 -m echoai.nexus0.bench --thought python3 -m echoai.nexus0.board echoai/lab/thought.jsonl → lab/board.html 112 321 bytes · 352 turns · 32 cells ``` The ANSI demo stays. CI does not need a browser, and nothing here replaces `view.py` or `demo.py`. ### The log grew from 12 keys to 17 The board cannot paint what the log does not carry, and `loop.turn` was computing five numbers and dropping them on the floor. `thought.KEYS` now adds exactly: ``` n 0|1 CAM miss on the origin this turn (n_fact) s 0|1 T predicted and was wrong (s_fact) delta 0|1 Q.update returned non-zero (d_fact) cortex_action null|0|1|2 the proposal _propose consumed gate_decision 0|1|2 OK / MODIFY / BLOCK ``` Integers and null. No prose, no nested `events[]` — a viewer can derive a timeline from these and cannot derive these from a timeline. The twelve older keys keep their name, their order and their meaning; a test pins `KEYS[:12]`. `cortex_action` is read in `turn()` *before* `_propose` consumes it, so it is the proposal that actually got a vote, and it is `null` on every turn of the ring. The flags are not a second, looser accounting: over the 352-turn bench `sum(n) = cam_misses = 33`, `sum(s) = surprises = 1`, `sum(delta) = deltas = 217`, asserted. ### What the board shows, and what it cannot Eight blocks, each showing this turn and not a metaphor for it: CAM (hit / MISS), WSP (the packet as atoms), T (`ŝ'` or —, red when S fired), Q (three integer bars, blue when δ ≠ 0), GATE (ok / modify / block), ATTEND (on/off + reason), CORTEX (the proposal, dashed grey — a hypothesis, never solid), ORACLE (whether `questions` rose). The ring is drawn from the log alone, rebuilt from turn 0 on every frame: a cell nobody has stood on yet is `?`, not a guess, and scrubbing backwards **un-reveals**. That is why the phase F threat move shows up exactly when it happened — turn 120 has `T` at cell 7, turn 300 has it at cell 8 — rather than being seeded from `RingEnv`. The world's size comes from the log (`max(cell) + 1`), never from `N_CELLS`: a `sign.py` log draws two cells, not thirty empty boxes. `board.py` imports `argparse`, `json`, `pathlib`, `sys` and nothing else. It does not import `loop`, `qtable`, `world`, `attend` or navi10, and `loop.py` gained **zero** imports — no Qt, no second bus. Both directions are pinned by tests that read the source as text. **One honest gap.** The five permitted keys contain no per-turn reject flag, so a cortex that woke and said nothing looks the same on the board as one whose reply `parse_reply` refused. Both read "woke and proposed nothing", and the legend on the page says so. Inventing a sixth key was not in the contract; if that distinction is wanted it is a key, not a guess in the viewer. ### Locks ``` policy_after [-12, +5, 0] · novel_approach_return +80 · surprise_known 1 cam_misses 33 · surprises 1 · deltas 217 false_facts 0 · destroyed 0 · cortex_calls 0 · cortex_actions 0 ring --cortex 36/256 · CORTEX-1 sign green · thought.KEYS = 17 ``` Tests 173 → 185. --- ## WORLD-3 — 256 addresses, one edge moves (2026-09-04) `nexus0/graph.py`. Not the ring, not the corridor of five, not SIGN. The first body-map: the animal's position **is** the node. Same `Agent`, same 16 B WSP, same three actions, same `encode_cell(index, kind)`. What is new is the address space. ``` 256 nodes · successor by table, never (i+1) % n, never a 2D grid spine 16 shared nodes, index and kind identical in both layouts HOME ... -> B(174) -> C(187) -> GOAL(200) -> back to START(5) SHIFT ... -> B(174) -> X(251) -> GOAL(200) -> back to START(5) ``` **Exactly one edge differs**: `approach` out of the fork B. Counted, not asserted — `test_only_one_edge_differs` walks all 3 × 256 rows of both tables and the diff is `[(approach, 174)]`. C is EMPTY and X is THREAT: two nodes with two kinds, not one node repainted. Nothing changes its label, so every packet, CAM slot, T row and Q row survives the move. In HOME nothing points at X; in SHIFT nothing points at C. `approach` at GOAL collects and returns to START — a goal you can stand on and re-collect is a reward pump, not a world. ### The transfer, without money in it On entry to SHIFT, with CAM/T/Q carried and nothing reset but the world: ``` T rows known on the 16 spine nodes 16 still true of HOME 16 / 16 still true of SHIFT 15 / 16 the one that is not spine position 13 = the fork control (Q/T/CAM empty), rows known 0 spine nodes unseen on entry transfer 0 control 16 ``` That is the whole claim, and it costs no reward to state: fifteen of the sixteen rows learned in HOME describe the new world exactly, and the agent is wrong **only** where the world changed. The first greedy turn in SHIFT proves it from the other side — T commits to C, the world hands back X, `surprise_known` rises, the origin slot is unchanged and is not a CAM miss, and X is a new packet. Same shape as the corridor's first wrong step, in a world of 256 addresses. ### `transfer_gain` = +256, and what it is made of ``` transfer_reward_sum_shift 256 (64 greedy turns from the fork) control_reward_sum_shift 0 transfer_gain +256 transfer_goal_hits_in_shift 0 transfer_avoid_pays_in_shift 32 x R_AVOID_OK(8) = 256 ``` Read it exactly as it is. The carried agent walks into X, its Q row for a node it has never seen is flat, greedy answers `wait`, and the **gate** turns that into `avoid` at a threat — which pays `R_AVOID_OK`. It bounces B ↔ X for the whole window and **never reaches GOAL in SHIFT**. The control has a flat row everywhere, so greedy answers `wait` and it does not move at all. So the gain is real reward on the same seat, the same seed and the same number of turns, and it is *not* a goal: it is the difference between an agent with a policy and an agent with none. Both halves are asserted in `test_transfer_gain_positive` so the number cannot be read as more than it is. No reward was retuned; `R_SUCCESS` / `R_THREAT` / `R_AVOID_OK` are the ring's own. ### The credit horizon — why the fork is at spine position 13 Measured on this spine, not assumed. Q is integer with α = 1/4, so a positive δ below 4 shifts to zero and value stops travelling. Counting back from GOAL, `Q[approach]` along the 16-node spine reads: ``` 0 0 0 0 0 0 0 0 0 0 0 0 0 4 8 13 (512 turns; identical at 2048) ``` Three nodes deep and no further. Eligibility traces (`trace=True`) do not extend it — they spread the debt and leave the far spine negative (`-27 -5 -6 …`), with greedy answering `avoid` at the start. So the fork sits inside the horizon at position 13. Put it at position 1, as first sketched, and both arms stand still in a flat row: `transfer_gain` would be 0 − 0, a KPI that cannot be failed. The mutant was run; it goes red. ### CAM, and what 256 does and does not prove ``` nodes 256 · reachable 17 (16 spine + X) · cam_occupied 17 / 4096 ``` The 239 off-path nodes are their own successor, so the table is total without inventing an edge — and nobody ever stands on them. **WORLD-3 does not press CAM.** ANTI-ALIAS has nothing to bite on here yet, and this report says so rather than implying otherwise. What 256 *does* prove is the packet ceiling. `encode_cell` writes the node index into `ext.seq` as `index & 0xFF`, so node 256 hands back byte for byte the packet of node 0. 256 is exactly this encoder's ceiling before two places become one place — the number ANTI-ALIAS has to move. `test_256_is_the_packet_ceiling` pins both halves: the collision at 256, and 256 distinct packets below it. ### Locks ``` ring default untouched: policy_after [-12, +5, 0] · +80 · surprise_known 1 cam_misses 33 · surprises 1 · deltas 217 · oracle_source sensor false_facts 0 · destroyed 0 · cortex_calls 0 · questions 0 WORLD-3 cortex OFF · 3 actions · green true report lab/nexus0_world3_report.json, never nexus0_report.json ``` Tests 192 → 208. This is not AGI. It is the corridor of five grown to 256 addresses. --- ## GUI-2 — the board, live (2026-09-04) `nexus0/board.py --follow`. The same board as GUI-1, watching the log as it is written. ### Why there is a server at all A page opened with `file://` cannot watch a file grow. It has no way to re-read the disk on a timer, and GUI-1's turns are baked into the document at write time. So `--follow` starts a stdlib server on the loopback and the page asks it, once every 200 ms, for the turns it does not already hold. **That is a design consequence, not a workaround** — the alternative would have been a second protocol, which is exactly what the contract forbids. ``` python3 -m echoai.nexus0.board echoai/lab/thought.jsonl --follow http://127.0.0.1:8765/ poll 200 ms route /turns.json ``` Two routes and no files: `/` renders the page, `/turns.json?from=N` answers with the turns after N. Nothing on disk is reachable by URL, so there is no path to traverse — `/etc/passwd` and `/../../board.py` are 404, asserted. `127.0.0.1` is not a flag; a board reachable from another machine would be a service. `?from=N` is parsed by hand rather than by importing a URL library to read one integer, and junk reads as 0 instead of a stack trace. ### The KPI, as a test and not a screenshot `test_the_poll_sees_a_line_that_was_not_there`: write a two-turn jsonl, start the server on a real socket in a thread, `GET ?from=0` → 2 turns, **append a third line**, `GET ?from=2` → exactly that turn, `total` 3, `cells` 8 (the world grew with it), all 17 keys present. Freeze the reader's cache so the poll only ever answers what was on disk at startup and four tests go red, that one included. Run for real, board up before the log existed, against the ANSI demo writing 48 turns: ``` before loop: waiting = true 27 polls saw new turns: from=0 [0] · from=1 [1,2] · … · from=47 [47] final total 48 · Ctrl-C exit 0 ``` The loop was never told anybody was watching: the board is a second reader of the same jsonl a human could `tail`. ### One protocol, and it is the one already in use ``` new WebSocket 0 EventSource 0 CDN 0 npm 0 Qt 0