Intrikata Stack guide · package 1.7.1
Candidate necessary R&D
Candidate necessary R&D for singularity-class takeoff
This is a conditional causal map across four stated takeoff pathways. It is not a forecast, not a recipe, not an occurrence detector, and not proof of universal necessity. Completing every item would be insufficient to cause or confirm takeoff.
Claim token: ranked-public-proxy-hypotheses-not-occurrence-forecast Ranker: rdgap-v2.1-readiness-and-observability-2026-08-01 Machine receipt: /demo/rdgap-necessity.json Live sibling: areweinthesingularity.com · awits.intrikata.com Live probe receipt: /demo/awits-rdgap-capability-impl.json (37/37 public-proxy probes; not frontier completion)
What the score means
candidate necessity =
0.60 * explicit pathway coverage
+ 0.25 * declared dependency centrality
+ 0.15 * evidence confidencePathway coverage uses the frozen status weights: required=1.00, likely_required=0.75, enabling=0.35, and not_required=0.00.
Markets, paper counts, theme heat, channel pressure, open-model velocity, and direct readiness have zero candidate-necessity weight. They must not change the order below. Direct readiness is unmeasured for all nine gaps; no public proxy is promoted into a fabricated readiness number.
Four declared pathways
| Pathway | Conditional mechanism | Catalog count |
|---|---|---|
| Recursive software R&D | AI improves algorithms, training, inference, and the next AI-R&D cycle until feedback gain exceeds one. | 3 required · 2 likely required |
| Scale + scaffold compounding | Compute, data, post-training, inference search, tools, and human orchestration continue compounding without a new architecture. | 2 required · 2 likely required |
| Automated science + hardware | AI closes algorithm, chip, materials, and experiment loops that expand its successors' resource and method frontier. | 3 required · 4 likely required |
| Distributed agent economy | Reliable agents compound digital R&D and economic capacity through parallelism, specialization, and fast replication. | 3 required · 3 likely required |
Ranked candidates
| # | Candidate R&D gap | Necessity | Pathway coverage | Centrality | Confidence | Observability |
|---|---|---|---|---|---|---|
| 1 | Test-time compute & reasoning generalization | 93.6 | 100.0 | 91 | 72 | Phase 2 · public partial |
| 2 | Reliable long-horizon agency & tool use | 92.1 | 93.8 | 94 | 82 | Phase 1 · public now |
| 3 | Compute, energy & training infrastructure | 83.9 | 81.3 | 89 | 86 | Phase 1 · public now |
| 4 | Automated / recursive R&D loops | 75.5 | 71.3 | 96 | 58 | Phase 3 · lab opaque |
| 5 | Training-data supply & synthetic-data validity | 69.1 | 65.0 | 82 | 64 | Phase 2 · public partial |
| 6 | Continual learning & sample-efficient adaptation | 53.9 | 45.0 | 75 | 54 | Phase 1 · public now |
| 7 | Automated scientific discovery & formal reasoning | 43.1 | 25.0 | 78 | 57 | Phase 3 · needs a new sensor |
| 8 | Grounded world models & multimodal planning | 41.2 | 27.5 | 70 | 48 | Phase 1 · public now |
| 9 | Safe multi-agent & economy-scale coordination | 36.7 | 25.0 | 64 | 38 | Phase 3 · not yet observable |
The top three are the highest candidate necessities across the declared pathways. This ranking says nothing about current completion and is not a singularity probability.
Complete candidate workstream manifest (37)
These rows are candidate research workstreams, not completed frontier capabilities. No row claims completion, deployment, frontier attainment, or measured readiness. Direct readiness is null / unmeasured; direct-readiness and public-attention necessity weights are both 0.0.
Live AWITS v3.2 implements a public-proxy probe for every row (probesPassed=37/37). Probe class mix: 8 decidable, 17 semi-decidable, 12 Tarski-named. Passing a probe is not completing the frontier capability; frontierCapabilityClaim stays not-claimed and workstream readiness stays unmeasured.
The normalized machine manifest at /demo/rdgap-capabilities.json is also embedded under capability_manifest in /demo/rdgap-necessity.json. Each workstream has exactly one parent gap. Evidence, the demotion condition, and observability are declared on that parent and inherited by its workstreams; this avoids manufacturing workstream-level evidence that the catalog does not contain.
| Capability | Parent gap | Candidate research workstream |
|---|---|---|
rd-capability:test-time-compute-reasoning-generalization:01 | 1 · Test-time compute & reasoning generalization | Inference-compute scaling laws with held-out, uncontaminated tasks |
rd-capability:test-time-compute-reasoning-generalization:02 | 1 · Test-time compute & reasoning generalization | Transfer of verifiable-reward training to non-verifiable domains |
rd-capability:test-time-compute-reasoning-generalization:03 | 1 · Test-time compute & reasoning generalization | Process supervision and search over reasoning traces at frontier scale |
rd-capability:test-time-compute-reasoning-generalization:04 | 1 · Test-time compute & reasoning generalization | Cost-aware reasoning: accuracy per token, not accuracy at any price |
rd-capability:reliable-long-horizon-agency-tool-use:01 | 2 · Reliable long-horizon agency & tool use | Long-horizon planning, memory, and recovery under tool failure |
rd-capability:reliable-long-horizon-agency-tool-use:02 | 2 · Reliable long-horizon agency & tool use | Computer-use, browser, and IDE agents with verifiable success rates |
rd-capability:reliable-long-horizon-agency-tool-use:03 | 2 · Reliable long-horizon agency & tool use | Safe tool APIs, sandboxing, and permission hierarchies |
rd-capability:reliable-long-horizon-agency-tool-use:04 | 2 · Reliable long-horizon agency & tool use | Multi-step evaluation suites with anti-overfit holdouts |
rd-capability:compute-energy-training-infrastructure:01 | 3 · Compute, energy & training infrastructure | Power delivery and cooling for multi-GW AI campuses |
rd-capability:compute-energy-training-infrastructure:02 | 3 · Compute, energy & training infrastructure | Next-generation accelerators, HBM, and optical interconnect |
rd-capability:compute-energy-training-infrastructure:03 | 3 · Compute, energy & training infrastructure | Training-system fault tolerance, utilization, and data pipelines |
rd-capability:compute-energy-training-infrastructure:04 | 3 · Compute, energy & training infrastructure | Algorithmic efficiency that multiplies effective compute |
rd-capability:compute-energy-training-infrastructure:05 | 3 · Compute, energy & training infrastructure | Hardware-software co-design within the compute-energy class |
rd-capability:automated-recursive-rd-loops:01 | 4 · Automated / recursive R&D loops | AutoML and architecture search that improves frontier training recipes |
rd-capability:automated-recursive-rd-loops:02 | 4 · Automated / recursive R&D loops | AI co-scientists for experiment design, code generation, and result critique |
rd-capability:automated-recursive-rd-loops:03 | 4 · Automated / recursive R&D loops | Closed-loop training-data generation with quality filters |
rd-capability:automated-recursive-rd-loops:04 | 4 · Automated / recursive R&D loops | Metrics for R&D productivity lift rather than loss curves alone |
rd-capability:training-data-supply-synthetic-validity:01 | 5 · Training-data supply & synthetic-data validity | Synthetic-data filters that remain valid across successive training generations |
rd-capability:training-data-supply-synthetic-validity:02 | 5 · Training-data supply & synthetic-data validity | Resolution of model-collapse versus accumulate-don't-replace results |
rd-capability:training-data-supply-synthetic-validity:03 | 5 · Training-data supply & synthetic-data validity | Licensing, acquisition, and multimodal corpus expansion beyond public text |
rd-capability:training-data-supply-synthetic-validity:04 | 5 · Training-data supply & synthetic-data validity | More capability per training token |
rd-capability:continual-learning-sample-efficient-adaptation:01 | 6 · Continual learning & sample-efficient adaptation | Stable continual and lifelong learning at foundation-model scale |
rd-capability:continual-learning-sample-efficient-adaptation:02 | 6 · Continual learning & sample-efficient adaptation | Efficient fine-tuning and modular adapters with transfer guarantees |
rd-capability:continual-learning-sample-efficient-adaptation:03 | 6 · Continual learning & sample-efficient adaptation | Memory architectures that retain skills under non-stationary tasks |
rd-capability:continual-learning-sample-efficient-adaptation:04 | 6 · Continual learning & sample-efficient adaptation | Data-efficient post-training from sparse human feedback |
rd-capability:automated-scientific-discovery-formal-reasoning:01 | 7 · Automated scientific discovery & formal reasoning | Autoformalization and theorem proving at research scale |
rd-capability:automated-scientific-discovery-formal-reasoning:02 | 7 · Automated scientific discovery & formal reasoning | Closed-loop wet-lab, materials, and drug-discovery agents |
rd-capability:automated-scientific-discovery-formal-reasoning:03 | 7 · Automated scientific discovery & formal reasoning | Verified synthesis of novel algorithms and training methods |
rd-capability:automated-scientific-discovery-formal-reasoning:04 | 7 · Automated scientific discovery & formal reasoning | Cross-domain knowledge graphs that propose testable experiments |
rd-capability:grounded-world-models-multimodal-planning:01 | 8 · Grounded world models & multimodal planning | Unified multimodal world models spanning vision, language, action, and physics priors |
rd-capability:grounded-world-models-multimodal-planning:02 | 8 · Grounded world models & multimodal planning | Sim-to-real and digital-twin planning loops |
rd-capability:grounded-world-models-multimodal-planning:03 | 8 · Grounded world models & multimodal planning | Causal and counterfactual evaluation of plans |
rd-capability:grounded-world-models-multimodal-planning:04 | 8 · Grounded world models & multimodal planning | Embodied and robotics transfer benchmarks with public suites |
rd-capability:safe-multi-agent-economy-scale-coordination:01 | 9 · Safe multi-agent & economy-scale coordination | Game-theoretic protocol design for agent societies |
rd-capability:safe-multi-agent-economy-scale-coordination:02 | 9 · Safe multi-agent & economy-scale coordination | Collusion, deception, and cascade-failure benchmarks |
rd-capability:safe-multi-agent-economy-scale-coordination:03 | 9 · Safe multi-agent & economy-scale coordination | Economic simulation environments with scarce resources |
rd-capability:safe-multi-agent-economy-scale-coordination:04 | 9 · Safe multi-agent & economy-scale coordination | Identity, reputation, and permission systems for agent networks |
Pathway matrix
| Gap | Recursive software R&D | Scale + scaffold | Automated science + hardware | Distributed agent economy |
|---|---|---|---|---|
| Reasoning generalization | required | required | required | required |
| Long-horizon agency | required | likely required | required | required |
| Compute and energy | likely required | required | likely required | likely required |
| Recursive R&D loops | required | enabling | likely required | likely required |
| Data and synthetic validity | likely required | likely required | likely required | enabling |
| Continual learning | enabling | enabling | enabling | likely required |
| Scientific automation | not required | not required | required | not required |
| Grounded world models | not required | enabling | likely required | not required |
| Multi-agent coordination | not required | not required | not required | required |
Direct readiness tests
Every direct-readiness value is null until its declared test is actually assembled and run.
- Reasoning generalization: held-out transfer curves from verifiable to open-ended domains, reporting inference cost and a human baseline.
- Long-horizon agency: contamination-resistant task-horizon evaluation at high reliability with messy objectives, recovery, and unchanged scaffolding.
- Compute and energy: capability per joule and dollar, accelerator supply, interconnect, and inference capacity tracked against demand.
- Recursive R&D: a prospective loop-gain ledger across at least three iterations, including AI-attributable gain, human input, compute, replication, and next-cycle throughput.
- Synthetic-data validity: successive-generation held-out performance and diversity under predominantly synthetic or self-generated mixtures.
- Continual learning: non-stationary tasks comparing continual updates with frozen-model retrieval and external-memory baselines.
- Scientific automation: a prospective registry of AI-originated hypotheses, independent replications, deployment, cycle time, and later capability contribution.
- World models: held-out causal, counterfactual, and long-horizon planning comparisons between explicit and implicit world-model objectives.
- Multi-agent coordination: economic-scale populations reporting throughput, coordination cost, correlated failures, and single-agent baselines.
Phase 1 means a public instrument can be assembled now. Phase 2 means public pieces exist but assembly is incomplete. Phase 3 means the measurement is lab-opaque, needs a new sensor, or requires a phenomenon that does not yet exist at benchmark scale. Phase describes observability, not readiness.
Minimum breakthroughs and substitutes
| Gap | Minimum breakthrough | Route-around or substitute |
|---|---|---|
| Reasoning generalization | Transfer from verifiable training domains to novel, open-ended research decisions. | Scale, brute-force search, specialized tools, or large agent populations. |
| Long-horizon agency | Reliable multi-day execution with verification, recovery, and tools on held-out R&D tasks. | Human orchestration and decomposition into shorter tasks. |
| Compute and energy | Successive generations without power, fabrication, networking, or cost becoming binding. | Efficiency, smaller models, distillation, or existing spare capacity. |
| Recursive R&D | Repeated AI-originated improvements that increase the next R&D cycle's throughput after accounting for labor and compute. | Scale-only or distributed-agent transformation. |
| Synthetic-data validity | Successive systems improve from generated data without collapse or hidden contamination. | Strong priors, online interaction, private corpora, or more efficient learning. |
| Continual learning | Persistent safe adaptation without catastrophic forgetting. | Retrieval, external memory, long context, periodic retraining, or human-maintained state. |
| Scientific automation | Replicable AI-originated discoveries in loop-closing algorithms, chips, energy, or methods. | Software-only R&D or scale-and-scaffold pathways. |
| World models | Causal and counterfactual planning under distribution shift. | Adequate implicit world models from scale and reinforcement learning. |
| Multi-agent coordination | Large populations specialize without correlated failure, conflict, or overhead erasing the gain. | A single strong agent or centrally orchestrated worker pool. |
Gates kept outside physical onset
- Scalable alignment & control is a survivable-start gate, not a physical onset requirement. Model-weight security is a sub-item of control.
- Mechanistic interpretability & oversight interfaces are a survivable-start gate when they add predictive value beyond behavioral evaluation and control protocols.
- Valid capability measurement & evaluation science is a detection and management gate. Bad measurement can hide takeoff; it does not physically prevent it.
Explicit exclusions
- Robotics, embodiment, and sim-to-real are downstream for digital initiation and already live inside world-model and scientific-automation work.
- Economic deployment friction is definition-dependent and already represented by diffusion telemetry (
T-DEFINITIONAL). - Governance and coordination modulate deployment rate rather than supply enabling R&D (
T-DEFINITIONAL). - Model-weight security changes control and proliferation; it is folded into alignment-control (
T-LAB-OPAQUE). - Hardware-software co-design is retained inside compute-energy to avoid double-counting infrastructure.
Residuals and diagnostics
T-COUNTERFACTUAL-NECESSITY: unobserved substitute pathways prevent proof of universal necessity.T-FIXEDPOINT-PLATEAU: more AI assistance may fail to produce more R&D progress.T-LAB-OPAQUE: the needed attribution or composition data are private.T-DEFINITIONAL: membership depends on the chosen definition of singularity or takeoff.R-RDGAP-BAND-CEILING: 4 of 48 gap-band cells are unreachable under their sensor ceilings. Report them; never widen a clamp to manufacture closure.
The separate attention tape uses 0.55*structural prior + 0.45*live pressure, but carries zero necessity weight. Its prior occupies an 18-point band while live pressure spans 31 points, so the prior realizes only 42% of ranking spread. Priors are not stretched to force that diagnostic to match the nominal coefficient.
Topology operations
From the intrikata-topology repository with its virtual environment:
$env:OP_CONTEXT = '{"action":"validate"}'
& '.\.venv\Scripts\python.exe' 'ops\RdGapCatalog_v1\main.py'
$env:OP_CONTEXT = '{"action":"graph"}'
& '.\.venv\Scripts\python.exe' 'ops\RdGapCatalog_v1\main.py'
$env:OP_CONTEXT = '{"action":"expand"}'
& '.\.venv\Scripts\python.exe' 'ops\RdGapCatalog_v1\main.py'validate and graph are offline. expand writes idempotent nodes and edges to the configured Intrikata Topology graph. It materializes pathways, candidates, direct readiness tests, gates, exclusions, and residuals while keeping the attention tape outside the necessity chain.