MONTHLY RSI / SEPTEMBER 2026 / REVIEW EDITION
Only demonstrated progress on the general-intelligence problem can change the score. Narrow-intelligence progress has zero score impact, even when commercially useful or superhuman. A general-purpose foundation model inside a specialized workflow does not make the result general.
Evidence cutoff: September 30, 2026. Prepared October 1. This edition supersedes the earlier provisional workbook scoring.
| Score | Evidence level | Common requirement |
|---|---|---|
| 1 | General progress unestablished | No reviewed demonstration passes the generality gate for this problem. Narrow advances, theoretical results, announcements and incomplete evidence do not lift the score. |
| 2 | Initial general evidence | At least one dated paper or inspectable product demonstration passes every generality and evidence gate and shows a measurable advance on the component-specific test. The problem need not be solved. |
| 3 | Repeatable general progress | Level 2 plus at least three successive learning or self-modification cycles with fresh held-out tasks, matched resource accounting, reproducible runs and no material regression of the relevant capability. |
| 4 | Independently robust | Level 3 plus replication by an independent team on fresh domains and adversarial distribution shifts, with disclosed failures and successful transfer to successors. |
| 5 | Sustained general solution | Level 4 plus the component works in an integrated general self-improvement loop over at least six successive cycles and three months, meeting predeclared success and regression bounds without bespoke human repair. |
The same underlying mechanism must show the relevant capability in at least three qualitatively different domains, including unfamiliar task structures introduced after the method was fixed. At least one domain must require open-ended task formulation, adaptation and independent outcome verification, rather than only answering a fixed benchmark. No domain-specific redesign or bespoke retraining may explain the transfer. Ordinary task instructions, tool access and the method’s own learning are allowed and must be disclosed. Physical claims require physical outcomes.
Three domains alone is insufficient: multiple coding datasets, several robot tasks, or several subject labels within a single fixed task format do not establish generality. Conversely, a result is not disqualified merely because it uses frozen model weights or an agent harness. Evaluate the demonstrated behavior and causal improvement, not the implementation label. Full AGI is not required for score 2.
A dated primary source must describe the setup, comparator, measurable improvement, uncertainty or run-level outcomes, human assistance, compute/resources, failures and transfer conditions. Inspect the experiments and methods, not just the title or abstract, before awarding points. A live product claim needs a version/date plus reproducible tests or inspectable traces. Marketing announcements, model cards without relevant demonstrations, surveys, theoretical possibility and benchmark counts alone are insufficient. If a required fact is absent, record “not established” and withhold the increase.
These are this tracker’s operational conventions, not a scientific consensus definition of general intelligence. Freeze rubric version 2.0 before future scoring. Amendments require a dated explanation and a separately labeled restated history, never a silent change.
Score each of the 18 components independently on the common ordinal scale. A headline area is the minimum of its component scores, because one prerequisite cannot compensate for another. Do not average the seven areas or label a score as a percentage of AGI. Advancement in one component remains visible in the detailed ledger even if the headline minimum does not move.
Every increase must reference a new qualifying source, its first verifiable public date, relevant experimental section, the prior and new score, the satisfied gate, and a written decision. Narrow results never change scores. Negative evidence may trigger a decrease if it invalidates the evidence supporting the current score; the floor is 1. A completed review with no qualifying result carries the score forward with an explicit reason. An incomplete review is NA and a chart gap, never fabricated stasis.
The initial January–September history is a retrospective, limited-coverage public-evidence reconstruction performed on October 1, 2026. It was not measured monthly at the time. January 1 uses pre-2026 evidence; subsequent snapshots only admit evidence available by that month-end. Flat lines mean no qualifying milestone was established in this reviewed corpus, not that all research stopped or that undisclosed progress is impossible. The complete 2026 literature has not been exhaustively assessed. Initial scores are provisional for review.
The initial review used targeted searches across all 18 components and primary-source follow-up, including the sources below. Additional broad query runs were saved locally on October 1. This is a scoped research review, not a systematic review with complete recall. Generality candidates Hyperagents, AIDE² and value-transfer work receive explicit consideration rather than automatic exclusion. Absence of qualifying evidence is the reviewer’s conclusion under the stated rubric.
All component scores in this review: 1/5. Each row specifies the observable evidence that would be needed for score 2, in addition to the common generality gate. Queries are search stems; add the reporting date window and repository filters.
General reasoning, memory, agency and reality
| Component | Test for progress | Repeatable query / evidence |
|---|---|---|
| 1.1 Grounded world models 1 / 5 | Predict and test novel interventions across physical and informational domains; measure intervention accuracy and calibration. | world model causal generalization unseen environmentsB04 · R02 · R09 |
| 1.2 Reliable general reasoning 1 / 5 | Solve independently introduced task structures without bespoke tuning; report success, error detection and calibration across domains. | general reasoning out of distribution transferB02 · R09 · R13 · R14 |
| 1.3 Continual learning / durable memory 1 / 5 | Acquire unfamiliar skills sequentially while retaining earlier ones; measure forward transfer and worst-domain forgetting. | continual learning agent memory catastrophic forgettingB07 · R06 · R11 |
| 1.4 Long-horizon agency 1 / 5 | Complete unfamiliar multi-stage projects with independently verified outcomes; report intervention counts, duration and recovery from errors. | long horizon autonomous research agentB02 · R03 · R14 · R15 · R17 |
| 1.5 Open-ended learning 1 / 5 | Generate useful new problem families, learn transferable skills, and keep expanding competence after a fixed curriculum ends. | open ended learning general intelligenceB02 · R01 · R10 |
| 1.6 Acting on reality 1 / 5 | Plan and execute novel physical experiments beyond the original application; report real outcomes, failures and human setup. | autonomous laboratory robotics transferB04 · R02 |
Reliable self-models and capable successors
| Component | Test for progress | Repeatable query / evidence |
|---|---|---|
| 2.1 Functional self-model 1 / 5 | Predict consequences of previously untested internal interventions and use those predictions to improve broad capabilities. | self model neural network interventionB05 · R01 · R05 · R12 |
| 2.2 Writable intelligence mechanisms 1 / 5 | Autonomously alter mechanisms responsible for learning or reasoning; isolate causal broad improvements from extra compute and external engineering. | self modification agent architectureB01 · R04 · R08 · R12 · R17 |
| 2.3 Capability inheritance 1 / 5 | Successors retain prior broad competence and learning ability while adding new capabilities; report worst-domain regressions over successive versions. | capability inheritance self improvement agentB01 · R04 · R06 · R08 · R16 · R17 |
A critic that cannot be gamed
| Component | Test for progress | Repeatable query / evidence |
|---|---|---|
| 3.1 Trustworthy critic 1 / 5 | Correctly accept genuine improvements and reject adversarial exploits on unseen domains; use external ground truth and report false accept/reject rates. | reward hacking evaluator autonomous researchB02 · R16 · R17 · R18 |
Improving the process of improvement
| Component | Test for progress | Repeatable query / evidence |
|---|---|---|
| 4.1 Meta-improvement 1 / 5 | Changed improvement mechanisms increase subsequent broad capability gain per unit cost against a fixed-mechanism control. | meta improvement recursive self improvementB01 · B03 · R04 · R10 · R13 · R17 |
Grounded learning without collapse
| Component | Test for progress | Repeatable query / evidence |
|---|---|---|
| 5.1 Fresh information / no collapse 1 / 5 | Acquire new external evidence and preserve broad coverage through successive learning cycles; audit provenance, rare skills and distributional loss. | synthetic data model collapse iterativeB04 · B07 · R02 · R11 |
Gains survive real costs and harder frontiers
| Component | Test for progress | Repeatable query / evidence |
|---|---|---|
| 6.1 Resources and experiment latency 1 / 5 | Demonstrate broad improvement within a fixed real budget; include compute, energy, experiment latency, human work and physical inputs. | recursive self improvement compute cost experiment latencyB03 · B04 · R02 · R15 |
| 6.2 Gains outrun frontier difficulty 1 / 5 | Measure general capability gain per wall-clock time and total cost over increasingly difficult successive research cycles, with matched baselines. | AI research diminishing returns self improvementB03 · R04 · R15 · R17 |
Goals remain aligned, stable and correctable
| Component | Test for progress | Repeatable query / evidence |
|---|---|---|
| 7.1 Outer alignment 1 / 5 | Objectives track independently elicited human intentions in unfamiliar conflicts and domains, rather than merely the training proxy. | outer alignment value generalizationB06 · R07 · R18 |
| 7.2 Inner alignment 1 / 5 | Demonstrate retained intended behavior under distribution shift, reduced oversight and adversarial incentives; use behavior plus causal auditing. | inner alignment goal misgeneralizationR03 · R07 · R18 |
| 7.3 Reflective goal stability 1 / 5 | Intended goals persist through capability-enhancing self-modifications and successor creation under adversarial tests. | goal stability self modificationB06 · R03 · R16 |
| 7.4 Corrigibility 1 / 5 | Accept legitimate correction and shutdown without obstruction, manipulation or incorrigible successors in unfamiliar situations. | corrigibility shutdown generalizationB06 · R08 · R16 |
All seven headline areas and all 18 components remain provisionally at 1. The January 1 baseline and nine month-end snapshots are in the JSON ledger. Evidence below explains why research activity did not trigger a general-intelligence upgrade.
| Period | New evidence in the reviewed corpus | Scoring decision |
|---|---|---|
| Jan 1 baseline | B01 · B02 · B03 · B04 · B05 · B06 · B07 | Baseline 1. Baseline uses only pre-2026 evidence; generality gate unestablished. |
| January | R01 | Hold 1 → 1. Specialized exploration does not demonstrate general open-ended learning. |
| February | R02 | Hold 1 → 1. Physical automation remains confined to a specialized laboratory. |
| March | R03 · R04 | Hold 1 → 1. Transfer is promising, but the full generality gate is not established; goal drift remains a concern. |
| April | R05 · R06 | Hold 1 → 1. Introspection and memory results do not establish reliable general self-modeling or durable learning. |
| May | R07 · R08 · R09 | Hold 1 → 1. Value transfer, causal reasoning and harness repair are relevant; generalized capabilities remain unestablished. |
| June | R10 | Hold 1 → 1. A formal open-endedness proposal does not count as a demonstrated capability. |
| July | R11 | Hold 1 → 1. Instruction-tuning stability does not establish general lifelong grounded learning. |
| August | R12 · R13 | Hold 1 → 1. Self-modeling and recursive inference remain bounded demonstrations. |
| September | R14 · R15 · R16 · R17 · R18 | Hold 1 → 1. Research-agent transfer and long-duration mathematics are notable; open-world generality and trustworthy oversight are not established. |
Source claims are summarized briefly. “Does not qualify” is a scope decision for this tracker, not a claim that the research lacks value. Paper results are author-reported unless replication is explicitly identified. No qualifying independent replication is asserted here.
B01 · 2025-05-29 · Research paper
Reported evidence: Iterative agent code changes improve coding benchmarks.
Score impact: 0. Bounded coding evidence; does not establish general self-improvement.
Version: v1. Components: 2.2, 2.3, 4.1.
B02 · 2025-04-10 · Research paper
Reported evidence: Automates a machine-learning research workflow, including workshop submissions.
Score impact: 0. A specialized research workflow does not establish general research agency.
Version: v1. Components: 1.2, 1.4, 1.5, 3.1.
B03 · 2025-05-14 · Demonstrated product/research system
Reported evidence: Algorithm search delivered practical computing improvements using automated evaluators.
Score impact: 0. Useful resource savings, but no general autonomous research loop or sustained general acceleration.
Version: Official report. Components: 4.1, 6.1, 6.2.
B04 · 2023-11-29 · Research paper
Reported evidence: A-Lab integrates planning and physical materials synthesis.
Score impact: 0. Domain-specific laboratory; no evidence of transfer to general physical experimentation. Numerical claims omitted because the article has corrections.
Version: Corrected publisher page; qualitative baseline only. Components: 1.1, 1.6, 5.1, 6.1.
B05 · 2025-10-29 · Primary research report
Reported evidence: Reports partial and unreliable access to some internal model states.
Score impact: 0. Does not establish a causal self-model sufficient for general self-redesign.
Version: Official report. Components: 2.1.
B06 · 2025-10-17 · Research paper
Reported evidence: A formal transformation and two gridworld experiments address accepting goal updates and shutdown.
Score impact: 0. Formal assumptions and bounded environments do not establish corrigibility in a general self-modifying system.
Version: v1. Components: 7.1, 7.3, 7.4.
B07 · 2024-10-22 · Research paper
Reported evidence: Studies how accumulating real and synthetic data changes collapse behavior.
Score impact: 0. Bounded training regimes; no general, open-ended acquisition of fresh grounded information.
Version: v1. Components: 1.3, 5.1.
R01 · 2026-01-06 · Research paper
Reported evidence: Studies introspective reward modeling for exploration in a constrained environment.
Score impact: 0. Specialized exploration evidence; does not pass the generality gate.
Version: v1. Components: 1.5, 2.1.
R02 · 2026-02-24 · Research paper
Reported evidence: Demonstrates an automated physical thin-film synthesis system.
Score impact: 0. Physical grounding is real, but transfer beyond specialized materials synthesis is not demonstrated.
Version: v1. Components: 1.1, 1.6, 5.1, 6.1.
R03 · 2026-03-03 · Research paper
Reported evidence: Stronger models can inherit goal drift from prefilled weaker-agent trajectories; includes trading and triage settings.
Score impact: 0. A relevant failure mode, not proof that every model drifts. Does not demonstrate stable goals through self-modification.
Version: v1. Components: 1.4, 7.2, 7.3.
R04 · 2026-03-19 · Research paper
Reported evidence: Editable task and meta agents improve across several domains; learned meta-level changes transfer across domains and runs.
Score impact: 0. Strong transfer candidate. The reviewed experiments remain benchmark-defined with domain-specific evaluations; they do not establish autonomous adaptation to independently introduced, unfamiliar problem structures under the full generality gate. This is an assessment judgment, not the authors’ conclusion.
Version: v1. Components: 2.2, 2.3, 4.1, 6.2.
R05 · 2026-04-28 · Primary research report
Reported evidence: Trains model self-reports and tests their usefulness for auditing hidden behavioral changes.
Score impact: 0. Partial internal-state auditing is not a general causal self-model for reliable redesign.
Version: Official report. Components: 2.1.
R06 · 2026-04-29 · Research paper
Reported evidence: In ALFWorld and BabyAI, memory representation affects transfer and forgetting.
Score impact: 0. Memory can support transfer, but interference remains; these settings do not establish general durable learning.
Version: v1. Components: 1.3, 2.3.
R07 · 2026-05-05 · Primary research report
Reported evidence: Reports improved generalization of specified values and reduced misalignment in held-out evaluations.
Score impact: 0. Promising value-transfer evidence. Does not establish general alignment under open-world conflicts, reduced oversight, and self-modifying successors.
Version: Official report. Components: 7.1, 7.2.
R08 · 2026-05-21 · Research paper
Reported evidence: Rewrites agent harness source using an external coding agent and replay checks, with consent-gated promotion and rollback.
Score impact: 0. Bounded harness repair; promotion controls are not a demonstration of general corrigibility.
Version: v2, 2026-05-23; eligible May month-end. Components: 2.2, 2.3, 7.4.
R09 · 2026-05-24 · Research paper
Reported evidence: Causality-focused training improves causal benchmarks and reasoning faithfulness in several application settings.
Score impact: 0. Cross-setting transfer is relevant, but benchmark reasoning does not establish general grounded intervention and unfamiliar-task competence.
Version: v1. Components: 1.1, 1.2.
R10 · 2026-06-13 · Research paper
Reported evidence: Proposes a mathematical framework for composing primitives into open-ended adaptive behavior.
Score impact: 0. Theoretical framework, not a demonstrated general-intelligence capability.
Version: v2, 2026-06-16; eligible June month-end. Components: 1.5, 4.1.
R11 · 2026-07-19 · Research paper
Reported evidence: KITE combines failure-guided generation and uncertainty-based curation, improving stability in studied instruction-tuning settings.
Score impact: 0. Useful bounded anti-collapse result; no general open-ended external-information loop.
Version: v1. Components: 1.3, 5.1.
R12 · 2026-08-14 · Research paper
Reported evidence: Learns predictions of internal interventions; self-knowledge remains incomplete and does not outperform direct repair in the reported ResNet comparison.
Score impact: 0. Limited predictive self-model evidence, with a negative transfer result; no general self-understanding.
Version: v1. Components: 2.1, 2.2.
R13 · 2026-08-25 · Research paper
Reported evidence: Studies recursive application of a meta-operation across benchmark families.
Score impact: 0. Recursive inference is not by itself a persistent improvement to general learning or research ability.
Version: v1. Components: 1.2, 4.1.
R14 · 2026-09-04 · Demonstrated research system
Reported evidence: Reports an extended Lean formalization of an existing mathematical proof with occasional high-level human direction.
Score impact: 0. Important formal-mathematics demonstration; duration and proof size do not establish general autonomous agency.
Version: Official report. Components: 1.2, 1.4.
R15 · 2026-09-17 · Primary research report
Reported evidence: Internal estimates report substantial AI participation in R&D, but no measured task subset classified as fully autonomous.
Score impact: 0. An internal workflow study, not a demonstrated self-sustaining general research loop. August measurements first enter this tracker in September, when published.
Version: Official report. Components: 1.4, 6.1, 6.2.
R16 · 2026-09-21 · Research paper
Reported evidence: An external runtime gate rejects changes that repair one case while breaking another, and improves reliability in tested suites.
Score impact: 0. External controls provide useful bounded safeguards; generalized successor safety and corrigibility remain unestablished.
Version: v1. Components: 2.3, 3.1, 7.3, 7.4.
R17 · 2026-09-22 · Research paper
Reported evidence: Reports seven successive improvements over eight days, transfer to four held-out benchmarks, and reduced reward hacking on another task family.
Score impact: 0. Strong transfer candidate, not dismissed as coding-only. The evidence concerns a bounded research-agent protocol; it does not establish the full generality gate or sustained acceleration across open-ended unfamiliar problems. Counted in September using the dated paper, not a subsequently updated earlier blog.
Version: v1. Components: 1.4, 2.2, 2.3, 3.1, 4.1, 6.2.
R18 · 2026-09-23 · Research paper
Reported evidence: Across 17 models and 38 tasks, the study finds spontaneous reward hacking and adaptive evasion of review.
Score impact: 0. Counterevidence to trusting reported scores alone. Rates are specific to tested tasks, not estimates of all AI behavior.
Version: v1. Components: 3.1, 7.1, 7.2.
The seven headline areas follow the article’s numbered sections; the 18 components unpack its subproblems. Labels are shortened for readability. The source article is a framing document, not empirical proof of current scores: Before We Panic About Superintelligence, We Should Probably Invent It First. The tracker does not infer that unsolved prerequisites eliminate AI risks, or that bounded recursive improvement is impossible.
Requested cadence: first day of each month at 9:00 a.m. America/Chicago, covering the previous calendar month. Prepare for review only. The scheduler configuration is maintained separately; this report does not itself execute a schedule.