feat(orchestrator): self-regulated thinking — graph sensorium + temperature + NARS auto-heal - #23
Merged
Merged
Conversation
- deepnsm.rs: lightweight NSM (Natural Semantic Metalanguage) module for the cockpit notebook system. 74 universal semantic primes, 113-word vocabulary, nsm_decompose(), cosine similarity, legality analysis (primes ratio, molecules ratio, circularity detection). Zero external dependencies. 13 tests passing. Transcoded from Python DeepNSM (AdaWorldAPI/DeepNSM). https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7
osint_audit.rs (notebook-query):
- OsintRegistry: global singleton with atomic counters for 12 pipeline stages
(extraction, refinement, planning, classification, deduction, contradiction,
revision, episodic_store, episodic_retrieve, graph_bfs, spatial_path, xai_api)
- OsintGraphHealth: triplet count, truth distribution, contradictions,
episodic saturation, NARS inference stats
- XaiStatus: ADA_XAI env var presence, call counts, failure rate
- run_osint_audit(): full health report with prioritized recommendations
- 7 tests
cockpit-server:
- New route: GET /api/debug/osint → osint_audit_handler
- Real-time AriGraph health monitoring alongside neural-debug strategy checks
The /api/debug/osint endpoint enables reading "brain" activation and plasticity
via verbose debug — every NARS deduction, contradiction detection, and evidence
revision is counted atomically. Combined with neural-debug's static scanner,
this gives both compile-time (dead/stub/NaN) and runtime (call count, latency,
success rate) visibility into the full OSINT pipeline.
https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7
… reasoning
notebook-query/mri.rs (500+ lines):
- BrainRegion model: 4 regions (perception, reasoning, memory, action)
each with sub-regions mapped from OSINT pipeline stages
- PlasticityState: Hot/Warm/Frozen/Conflicted per entity
(maps to CausalEdge64 bits 49-51)
- ThinkingStyleActivation: per-style call count, quality, NARS effectiveness truth
- ReasoningChain: traced NARS inference steps (deduction/abduction/induction)
- ScanMode: Structural (topology), Functional (activation), Full (DTI with chains)
- run_brain_mri(): collects all data into single BrainMri response
- Health score: region activation minus conflict penalty
- Findings: auto-detected hot plasticity, frozen regions, contradictions
- 6 tests
cockpit-server routes:
- GET /mri → standalone HTML page with live visualization
Auto-refreshes every 5s. Color-coded regions (hot=red, frozen=blue,
active=green, conflicted=amber). Bar charts for activation levels.
- GET /api/mri/scan → JSON API (full scan)
- GET /api/mri/scan/:mode → JSON API (structural/functional/full)
The MRI page shows:
1. Brain regions with activation bars (perception/reasoning/memory/action)
2. Plasticity map: which entities are Hot (learning) vs Frozen vs Conflicted
3. NARS reasoning chains: active deduction/abduction/induction traces
4. Findings: auto-generated health assessment
https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7
orchestrator.rs (550+ lines, 11 tests):
MetaOrchestrator: self-monitoring agent loop with two transparent modes:
1. ADAPTIVE (default): NARS topology learns which thinking style
sequences produce good outcomes. 4×4 = 16 directed edges, each
with a TruthValue (frequency=success rate, confidence=evidence
strength). Selection: exploit highest expected quality (85%) or
explore least-observed edge (15%) for information gain.
2. HARDCODED FALLBACK: When rolling efficiency drops below 0.35,
transparently switches to plan→act→explore→reflex sequence.
When fallback efficiency exceeds 0.55, re-enables adaptive
mode with an exploration burst.
The meta-awareness layer monitors its OWN efficiency:
- Rolling window of last 20 outcome qualities
- NARS revision on every (style_from → style_to) transition
- Automatic mode switching with full event log
- Every mode switch recorded with reason + efficiency at switch time
This IS "thinking about thinking": the orchestrator observes which
cognitive styles are effective, learns optimal sequences via NARS
evidence accumulation, detects when learning isn't working (low
efficiency), and falls back to a known-good baseline. The /mri
endpoint shows the full topology — which style transitions have
high confidence, which are being explored, whether the system is
in adaptive or fallback mode and why.
cockpit-server routes:
- GET /api/orchestrator/status → full snapshot (mode, topology, efficiency)
- POST /api/orchestrator/step → execute one step + optional quality feedback
https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7
…ompass
The Meta-Uncertainty Layer now modulates every orchestrator decision:
Dunning-Kruger position (4 states):
MountStupid → ForceSandbox (Reflex only, block all autonomous action)
ValleyOfDespair → ForceExplore + 2× exploration rate (you're learning)
SlopeOfEnlightenment → Normal operation, 1× exploration
PlateauOfMastery → 0.5× exploration (exploit, you've earned it)
Trust texture (3 levels):
Crystalline → trust topology weights fully (free_will × 1.0)
Fibrous → moderate discount (free_will × 0.7)
Fuzzy → heavy discount (free_will × 0.4), distrust learned edges
Flow state (4 modes):
Flow → normal patience, normal thresholds
Boredom → 1.5× patience, wider thresholds, more exploration
Anxiety → 0.5× patience, tighter fallback (trigger sooner)
Apathy → 0.25× patience, fast fallback
Compass override:
ForceSandbox → Mount Stupid detected, return Reflex only
ForceExplore → Valley + very low competence, override topology
free_will_modifier = DK_humility × trust_factor × flow_patience
Scales topology expected_quality — low free_will = distrust your own learning.
MulAssessment auto-derives from rolling efficiency (demonstrated competence)
vs topology confidence (felt competence). This closes the self-assessment
loop: the system's belief about its own ability is compared against
actual measured performance, and the gap determines DK position.
New StepReason::MulOverride variant for compass/DK overrides.
Every StepResult now carries the full MulAssessment that drove it.
Mode switch reasons now include DK position and flow state.
Snapshot includes MUL for /mri visualization.
6 new tests for MUL integration.
https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7
…rature + NARS auto-heal
Three new self-awareness mechanisms:
GraphSensorium — real-time signals from the knowledge graph:
- contradiction_rate: contradictions / active_triplets
- truth_entropy: Shannon entropy of confidence distribution
- revision_velocity: revisions/step (learning rate)
- plasticity_flux: fraction of Hot entities (environment change rate)
- deduction_yield: inference success rate
- episodic_saturation: memory fullness
→ suggested_bias(): Resolve/Explore/Exploit/Adapt/Stagnant/Balanced
Temperature — LLM-style noise injection for stale thinking:
- 0.0 = deterministic (normal greedy topology selection)
- 1.0 = maximum randomness (break out of local optima)
- Auto-increases on GraphBias::Stagnant (thinking is stuck)
- Auto-decreases on GraphBias::Exploit (graph is consistent)
- Injected as deterministic noise into topology expected_quality scores
NARS Auto-Heal Contingency — the immune system:
- BootstrapTruth: uninitialized truth values → set from_evidence(1,0)
- ResolveContradictions: high contradiction rate → reduce conflicting confidence
- InferMissingLinks: consistent but sparse → run deduction to fill gaps
- CompactDeleted: high episodic saturation → garbage collect
- NormalizeTruth: possible confidence inflation → re-scale
- ResetTopology: orchestrator learning poisoned → wipe NARS edges,
warm restart with temperature=0.5, hardcoded fallback
The self-regulation loop:
graph mutations → GraphSensorium::compute()
→ update_sensorium() adjusts temperature
→ auto_heal() diagnoses + prescribes healing actions
→ select_next() uses from_graph_signals() for MUL assessment
→ DK position derived from graph consistency (demonstrated) vs topology confidence (felt)
→ style selection modulated by temperature + MUL free_will
→ execution → graph mutations → loop
12 new tests for sensorium, temperature, auto-heal, graph bias.
https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7
AdaWorldAPI
pushed a commit
that referenced
this pull request
Aug 11, 2026
The decode half of the .chains codec (osm-soa-bake PR #23): the bake already computed every way's z=32 vertex chain and dropped it after mean_cell — the sidecar un-discards it, and this side reads it back. - osm_features: GET /api/osm/geometry/:idx — open the .chains sidecar once (OnceLock), REFUSE it unless its slab_digest matches the mapped slab (cross-bake geometry against another bake's identities is the drift the pin exists to make loud), resolve row -> identity ordinal -> chain -> lon/lat points. 404 when no chain is stored (nodes, relations) — never 200-with-empty. - osm page: an SVG shape layer inside #tiles (inherits the map transform); classFor(tags) fills water/building/wood/green rings and strokes highways; showShape wired into the existing click detail. - osm_slab_hydrate: ARTIFACTS grows to berlin.chains — a deploy now hydrates all three artifacts (~1.42 GB cold; volume >= 2 GB). The bucket's SHA256SUMS gained the third line additively, so binaries reading only soa+books still verify. Verified in a real browser against the real Berlin bake: a genuine click on a harbour dot resolved "Westhafen I" (natural=water, water=harbour) and drew its shore ring filled; building and landuse rings and a highway polyline drawn via the page's own showFeature; a node correctly 404s; zero page errors. Gaps the POC now makes visible (by design — the POC is the falsifier): shapes are click-only rather than a base fill layer with a node/street overlay; multipolygon relations are unassembled; small rings are sub-pixel at overview zoom. Recorded in the plan's Phase 8. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NMeiLmtDKhomJNSo2ecbJw
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
2 participants
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.
Three new self-awareness mechanisms:
GraphSensorium — real-time signals from the knowledge graph:
→ suggested_bias(): Resolve/Explore/Exploit/Adapt/Stagnant/Balanced
Temperature — LLM-style noise injection for stale thinking:
NARS Auto-Heal Contingency — the immune system:
warm restart with temperature=0.5, hardcoded fallback
The self-regulation loop:
graph mutations → GraphSensorium::compute()
→ update_sensorium() adjusts temperature
→ auto_heal() diagnoses + prescribes healing actions
→ select_next() uses from_graph_signals() for MUL assessment
→ DK position derived from graph consistency (demonstrated) vs topology confidence (felt)
→ style selection modulated by temperature + MUL free_will
→ execution → graph mutations → loop
12 new tests for sensorium, temperature, auto-heal, graph bias.
https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7