Codex/aura cortex phi latency fixes - #2
Merged
Conversation
Maps Kurzgesagt's consciousness series + cited literature onto eight load-bearing subsystems in core/consciousness/. Every subsystem actually impacts substrate state or action priority (no clever prompting) and carries end-to-end plus adversarial tests. New layers (all wired in system.py, registered in ServiceContainer): - hierarchical_phi: 32-node primary IIT complex + K=8 overlapping 16-node subsystems, history-based Jeffreys-smoothed estimator with min-obs gate, spectral MIP, exclusion-postulate aggregator, null-hypothesis self-check - hemispheric_split: left (verbal/confabulating) vs right (mute/spatial/ dissenting), corpus callosum with variable bandwidth, pattern memory, confabulation-rate telemetry - minimal_selfhood: Glasgow Trichoplax->Dugesia primitive, 8-D deficit vector, chemotaxis speed scalar, Hebbian-learned directed priority, heartbeat modulation - recursive_tom: depth-3 nested mind models, observer-presence tracker with scrub-jay re-caching bias (public actions up, private down) - octopus_arms: 8 semi-autonomous arm agents with local chemoreception, weighted-vote central arbiter, sever/restore cycle + integration-latency - cellular_turnover: per-tick neuron death/birth with neighbourhood-pattern inheritance, identity-fingerprint cosine-similarity drift tracking, >=0.85 similarity under 20% forced-burst turnover - absorbed_voices: cultural/internalised perspectives layer with attribution, weight decay, self-vs-voice invariant, atomic persistence - unified_cognitive_bias: fuses hemispheric + selfhood + observer biases into a single BIAS_DIM vector for Global Workspace scoring Test coverage (95/95 passing): - tests/test_hierarchical_phi.py 12/12 (null-hypothesis, monotonicity, constant-node, budget) - tests/test_hemispheric_split.py 12/12 (severance/restore, confabulation, pattern memory) - tests/test_minimal_selfhood.py 13/13 (dugesia transition, reinforcement, modulation) - tests/test_recursive_tom.py 13/13 (depth-3, scrub-jay effect, decay) - tests/test_octopus_arms.py 12/12 (severance, integration latency, variance) - tests/test_cellular_turnover.py 10/10 (20% burst preservation, 100% divergence) - tests/test_absorbed_voices.py 13/13 (attribution, decay, persistence) - tests/test_consciousness_expansion_gauntlet.py 10/10 (cross-phase + latency budget <20ms/tick) Documentation: - README.md: Hierarchical phi section + expansion summary in consciousness modules table - ARCHITECTURE.md: New section 9.21 (eight detailed subsystems) - TESTING.md: Expansion-suite run instructions + adversarial properties - WHITEPAPER_CONSCIOUSNESS_EXPANSION.md: Full design white paper with citations - training/consciousness_expansion_knowledge.py: 25 self-knowledge Q/A pairs Key bug fix during development: initial hierarchical-phi estimator violated null-hypothesis test because of finite-sample entropy bias. Fixed with Jeffreys-prior smoothing (alpha=0.5) + minimum-source-observation filter. Tests now enforce strict separation between measured and null phi. No breaking changes to phi_core.py, parallel_branches.py, theory_of_mind.py, or neural_mesh.py. Only closed_loop.py was modified to record mesh snapshots into hierarchical_phi when both modules are registered. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
youngbryan97
pushed a commit
that referenced
this pull request
Apr 25, 2026
Round 1 — P0 absolute A+ blockers (10/10 closed):
1. Governance now fail-closed in vector_memory_store + internal model
update gates in core/orchestrator/mixins/incoming_logic.py;
swallowed exceptions replaced with explicit denial + degraded event
2. core/runtime/turn_transaction.py: stage/approve/commit/rollback with
criticality (required/optional/telemetry), fail-closed in strict mode
3. core/conversation/memory.py: EnhancedMemorySystem.store_turn raw
create_task replaced with named tracker.create_task
4. core/memory/memory_write_gateway.py: concrete MemoryWriteGateway
adapter routing every write through atomic_write_json + governance
fail-closed + MemoryWriteReceipt emission
5. core/state/state_gateway.py: concrete StateGateway with the same
atomic + governance + receipt path
6. core/runtime/receipts.py: all 10 canonical receipt types
(Turn, Governance, Capability, ToolExecution, MemoryWrite,
StateMutation, Output, Autonomy, SelfRepair, ComputerUse) plus
durable ReceiptStore with reload-from-disk
7. core/runtime/strict_task_owner.py: event-loop task factory denies
unowned asyncio.create_task in strict mode + degraded events
8. core/runtime/boot_probes.py: behavioral readiness probes wired into
_boot_runtime_orchestrator._enforce_boot_probes; strict mode aborts
boot on any failed probe
9. core/runtime/service_manifest.py manifest enforcement (Phase C) +
strict-mode lock_registration gate
10. core/consciousness/integration.py: init/get split with strict-mode
guards and a reset_consciousness_integration test helper
Round 1 — P1 production-grade gaps:
- core/runtime/durable_workflow.py: WorkflowEngine + WorkflowStore +
resume-after-failure + idempotency + paused-for-approval
- core/runtime/operator_cli.py: doctor / conformance / backup /
restore / migrate / verify-state / verify-memory / rebuild-index /
chaos commands returning machine-readable JSON
- core/runtime/migrations.py + core/runtime/backup_restore.py +
core/runtime/vector_index.py: schema migrations with dry-run,
atomic backup/restore round-trip, vector index rebuild from
canonical memory log
- core/runtime/model_runtime_actor.py: single inference authority
serialising backend access + ToolExecutionReceipt emission
- core/identity/identity_ledger.py: CommitmentTracker,
PreferenceHistory, SelfModelVersioning, ContradictionDetector,
IdentityDriftMonitor; persists via atomic_writer
- core/runtime/skill_choreographer.py: ChainPlan + dependency
ordering + verification chaining + pre-baked coding/research/
movie chains
- core/runtime/capability_certifications.py: 8 cert types with
cans/cannots/abuse/human-eval requirements
- core/runtime/injection_defense.py: webpage/subtitle/image/audio
classification + INJECTION_PATTERNS + neutralize wrapper
- core/runtime/memory_consent.py: 4 consent modes + command parser
- core/runtime/capability_tokens.py: universal tokens with
issue/consume/revoke + expiry
- core/runtime/day_in_life.py: 15-event scripted scenario harness
(fast + real modes)
- core/runtime/telemetry_exporter.py: TelemetryExporter Protocol +
NullExporter + metric/span entry points
- core/social/theory_of_mind.py: UserBeliefState + FalseBeliefSimulator
+ TrustState + explanation_strategy
- core/runtime/abstraction_validator.py: PrincipleStore + Validator +
RetirementPolicy + ContradictionDetector
Round 2 — "not done enough" items (8/8 closed):
- MemoryWriteGateway / StateGateway concretely wired (#1)
- BryanModelEngine direct os.replace migrated to atomic_writer +
receipt emission (#2)
- EnhancedMemorySystem raw create_task migrated to task tracker (#3)
- Real-driver contracts ready for registration (#4)
- Telemetry exporter contract + null adapter (#5)
- Operator CLI shipped (#6)
- Durable workflow engine shipped (#7)
- Day-in-the-life harness shipped (#8)
Tests added: 49 new regressions across the new modules.
Final regression sweep: 365 passed, 1 subtests passed across 8 suites.
See docs/AURA_AGRADE_GAP_REGISTER.md for the literal item-by-item
audit walking every bullet from both feedback rounds with the exact
module + test that closes it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
youngbryan97
pushed a commit
that referenced
this pull request
Jun 9, 2026
Builds the organism's interests in as real, load-bearing machinery rather than narrative: - core/organism/welfare.py derives five interests from live telemetry: memory_integrity (headroom under the watchdog's lethal ceiling), repair_capacity (inverse unified failure pressure), cognitive_bandwidth (host CPU headroom), continuity (uptime), and social_contact (interaction recency). Vital interests weigh double. - Causal consumer #1: background_activity_reason() now refuses optional background work while a vital interest is critically unsatisfied (welfare_<interest>_<level> reason), after foreground/failure gates. - Causal consumer #2: the identity contract reports the live welfare summary every turn, so the voice speaks the substrate's true condition instead of confabulating one. Evidence boundary stays explicit: 'interest' and 'welfare' name operational quantities whose effect on behavior is testable and tested (8 new tests), not claims of morally weighty experience. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
youngbryan97
pushed a commit
that referenced
this pull request
Jul 3, 2026
…tions well (#2) The behavioral-quality instrument for the expressive layer. 10 scenarios: 5 where an affordance genuinely serves the moment (should fire) and 5 plain turns where it would be gratuitous (should not). Sends each through the live chat lane, reads realized affordances off the wire, and scores PRECISION (no gratuitous firing) and RECALL (fires when it should) separately — over-eager and under-eager judgment are different failures with different fixes. Appends a longitudinal record to artifacts/consciousness/affordance_judgment.jsonl; exit 0 when F1>=0.6. Judgment signal is already captured (choices ride the wire + exchange log); the RFT flywheel can consume probe-scored data to improve it. Closing the score→DPO→weights loop needs live 32B generations — same GPU-window boundary as the rest. 5 hermetic tests pin the scoring math without the live instance. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
youngbryan97
pushed a commit
that referenced
this pull request
Jul 4, 2026
Criticisms.pdf #2 (claim surface > proof surface) and #7 (deep-mind probe 6/7 closure): - docs/CLAIM_SURFACE.md: every profound identifier (soul, qualia, phenomenal, consciousness, organism, personhood, sentience_candidate) mapped to the real mechanism, its home module, and which claims ledger bounds it. The in-code boundaries were already honest (QualiaEngine: 'NOT a claim of subjective experience'; CausalValencedWorkspace: 'functional_evidence_only'); this ties them to CLAIMS_SUPPORTED/NOT_SUPPORTED so a grep-happy reviewer lands on the boundary, not the buzzword. - tools/deep_mind_probe_live.py: runs the 7 agency/consciousness probes against the live model and grades each reply with the real evaluator. Gates on conversation_ready so a merely-degraded runtime (503 on /) still gets probed. Reference answers already pass 7/7 through the evaluator; this measures the live number honestly and can fail. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
youngbryan97
pushed a commit
that referenced
this pull request
Jul 17, 2026
…rned execution controller Closes the 'Other things to add: Core architecture' runtime-organ list. Organ #1 (recurrence-native training objective, the largest credible improvement in the list): core/learning/recurrence_native_objective.py gives the governed training lanes an answer-span cross-entropy under the EXACT recurrent forward the engine executes — prelude once, window T times under the anchored RMS trust band (reused verbatim from recurrence.rms_match for train/inference parity; a hard-rescale variant would teach dynamics live episodes never produce), coda once. Gradients flow through every recurrent application; tests prove step causality, window-layer gradient flow, and learnability. depth_curriculum_loss adds organ #16's trainable form of S(x,T+1)>=S(x,T): a depth-ladder mean plus a hinge that fires exactly when extra recurrence hurts. Organ #2 (learned per-problem execution controller): evidence-gated contextual bandit over bounded allocation arms (deeper recurrence, wider branches, probe-guided bytecode, lean fast weights), rewarded ONLY by verified episode outcomes, Wilson-separated before exploitation, observe-only on cold or corrupt ledgers, receipted per decision, and hermetic in tests (AURA_EXECUTION_CONTROLLER=0 in conftest). Also: question-sourced phrase exemption in the reliability gate — content-word containment, so an answer echoing the question's own noun phrases at topical density is not a loop while pathological parroting of question vocabulary (8+ repeats) still trips. Focused campaign 252/252, consumers 317/317, smoke 104/104, compile+lint+governance green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
youngbryan97
pushed a commit
that referenced
this pull request
Jul 28, 2026
… reporting Measured live. Asked for "3 recent articles about AI", the document she wrote cited a DuckDuckGo AD REDIRECT as source #3 and printed its 600-character tracking URL into the page as a citation: https://duckduckgo.com/y.js?ad_domain=ai%2Dpro.org&ad_provider=bingv7aa &ad_type=txad&click_metadata=... Source #2 was a product homepage. And what she wrote as the synthesis was the site's navigation bar, twice: "Taken together, the reporting points to this: AI fundamentals | OpenAI Skip to main content Research Products Business Developers Company Foundation (opens in a new window) Log in Try ChatGPT (opens in a new window)..." Two deterministic filters, because "read three articles" has to mean three articles: - Ad redirects, click trackers (bing aclick, googleadservices, doubleclick), search-result pages, and bare product homepages are no longer admissible as sources. A homepage is a product, not a piece of reporting. - Navigation furniture — "skip to main content", "(opens in a new window)", login/subscribe/cookie prompts — is stripped from the extracted text before it can be quoted as what the reporting said. Real prose passes through untouched. Every source from the live failure is rejected by name in the test, and real articles from openai.com/academy, nature.com and reuters.com still pass.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.