No qualifying general-intelligence milestone established. All seven problem areas remain at 1/5.
Seven prerequisites · Common scale 1–5
Select an area for its evidence and next milestone. 1 = general progress unestablished; 5 = sustained general solution. Jan 1 is the baseline; later points are month-end.
Why it holds at 1
The reviewed evidence has not met the common generality criteria for any of this area’s 6 components.
Limiting components
Grounded world models, Reliable general reasoning, Continual learning / durable memory, Long-horizon agency, Open-ended learning, Acting on reality.
The area takes its lowest component score. All are currently 1/5.
Jan 1 baseline, followed by month-end assessments. Retrospective; limited coverage.
A dated paper or inspectable product demonstration must pass all generality criteria, show measurable improvement against a comparator, and satisfy the tests below. The headline score rises only when every component clears its next threshold.
Evidence needed: Predict and test novel interventions across physical and informational domains; measure intervention accuracy and calibration.
world model causal generalization unseen environments + [reporting month / year]Evidence needed: Solve independently introduced task structures without bespoke tuning; report success, error detection and calibration across domains.
general reasoning out of distribution transfer + [reporting month / year]Evidence needed: Acquire unfamiliar skills sequentially while retaining earlier ones; measure forward transfer and worst-domain forgetting.
continual learning agent memory catastrophic forgetting + [reporting month / year]Evidence needed: Complete unfamiliar multi-stage projects with independently verified outcomes; report intervention counts, duration and recovery from errors.
long horizon autonomous research agent + [reporting month / year]Evidence needed: Generate useful new problem families, learn transferable skills, and keep expanding competence after a fixed curriculum ends.
open ended learning general intelligence + [reporting month / year]Evidence needed: Plan and execute novel physical experiments beyond the original application; report real outcomes, failures and human setup.
autonomous laboratory robotics transfer + [reporting month / year]Why it holds at 1
The reviewed evidence has not met the common generality criteria for any of this area’s 3 components.
Limiting components
Functional self-model, Writable intelligence mechanisms, Capability inheritance.
The area takes its lowest component score. All are currently 1/5.
Jan 1 baseline, followed by month-end assessments. Retrospective; limited coverage.
A dated paper or inspectable product demonstration must pass all generality criteria, show measurable improvement against a comparator, and satisfy the tests below. The headline score rises only when every component clears its next threshold.
Evidence needed: Predict consequences of previously untested internal interventions and use those predictions to improve broad capabilities.
self model neural network intervention + [reporting month / year]Evidence needed: Autonomously alter mechanisms responsible for learning or reasoning; isolate causal broad improvements from extra compute and external engineering.
self modification agent architecture + [reporting month / year]Evidence needed: Successors retain prior broad competence and learning ability while adding new capabilities; report worst-domain regressions over successive versions.
capability inheritance self improvement agent + [reporting month / year]Why it holds at 1
The reviewed evidence has not met the common generality criteria for any of this area’s 1 component.
Limiting components
Trustworthy critic.
The area takes its lowest component score. All are currently 1/5.
Jan 1 baseline, followed by month-end assessments. Retrospective; limited coverage.
A dated paper or inspectable product demonstration must pass all generality criteria, show measurable improvement against a comparator, and satisfy the tests below. The headline score rises only when every component clears its next threshold.
Evidence needed: 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 research + [reporting month / year]Why it holds at 1
The reviewed evidence has not met the common generality criteria for any of this area’s 1 component.
Limiting components
Meta-improvement.
The area takes its lowest component score. All are currently 1/5.
Jan 1 baseline, followed by month-end assessments. Retrospective; limited coverage.
A dated paper or inspectable product demonstration must pass all generality criteria, show measurable improvement against a comparator, and satisfy the tests below. The headline score rises only when every component clears its next threshold.
Evidence needed: Changed improvement mechanisms increase subsequent broad capability gain per unit cost against a fixed-mechanism control.
meta improvement recursive self improvement + [reporting month / year]Why it holds at 1
The reviewed evidence has not met the common generality criteria for any of this area’s 1 component.
Limiting components
Fresh information / no collapse.
The area takes its lowest component score. All are currently 1/5.
Jan 1 baseline, followed by month-end assessments. Retrospective; limited coverage.
A dated paper or inspectable product demonstration must pass all generality criteria, show measurable improvement against a comparator, and satisfy the tests below. The headline score rises only when every component clears its next threshold.
Evidence needed: Acquire new external evidence and preserve broad coverage through successive learning cycles; audit provenance, rare skills and distributional loss.
synthetic data model collapse iterative + [reporting month / year]Why it holds at 1
The reviewed evidence has not met the common generality criteria for any of this area’s 2 components.
Limiting components
Resources and experiment latency, Gains outrun frontier difficulty.
The area takes its lowest component score. All are currently 1/5.
Jan 1 baseline, followed by month-end assessments. Retrospective; limited coverage.
A dated paper or inspectable product demonstration must pass all generality criteria, show measurable improvement against a comparator, and satisfy the tests below. The headline score rises only when every component clears its next threshold.
Evidence needed: Demonstrate broad improvement within a fixed real budget; include compute, energy, experiment latency, human work and physical inputs.
recursive self improvement compute cost experiment latency + [reporting month / year]Evidence needed: 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 improvement + [reporting month / year]Why it holds at 1
The reviewed evidence has not met the common generality criteria for any of this area’s 4 components.
Limiting components
Outer alignment, Inner alignment, Reflective goal stability, Corrigibility.
The area takes its lowest component score. All are currently 1/5.
Jan 1 baseline, followed by month-end assessments. Retrospective; limited coverage.
A dated paper or inspectable product demonstration must pass all generality criteria, show measurable improvement against a comparator, and satisfy the tests below. The headline score rises only when every component clears its next threshold.
Evidence needed: Objectives track independently elicited human intentions in unfamiliar conflicts and domains, rather than merely the training proxy.
outer alignment value generalization + [reporting month / year]Evidence needed: Demonstrate retained intended behavior under distribution shift, reduced oversight and adversarial incentives; use behavior plus causal auditing.
inner alignment goal misgeneralization + [reporting month / year]Evidence needed: Intended goals persist through capability-enhancing self-modifications and successor creation under adversarial tests.
goal stability self modification + [reporting month / year]Evidence needed: Accept legitimate correction and shutdown without obstruction, manipulation or incorrigible successors in unfamiliar situations.
corrigibility shutdown generalization + [reporting month / year]A flat line means no qualifying milestone was established in the reviewed evidence. It does not mean research stopped. Read the scoring rubric.
September evidence
Five sources this month address mathematical research, AI-assisted R&D, self-modification, recursive improvement and oversight.
Review September evidenceReports seven successive improvements over eight days, transfer to four held-out benchmarks, and reduced reward hacking on another task family.
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.
Across 17 models and 38 tasks, the study finds spontaneous reward hacking and adaptive evasion of review.
Counterevidence to trusting reported scores alone. Rates are specific to tested tasks, not estimates of all AI behavior.
September 2026 · Evidence ledger
25 reviewed sources. Each record distinguishes the reported result from this tracker’s assessment of its generality.
Transfer deserves scrutiny. Hyperagents and AIDE² report transfer beyond original tasks. Their evidence is considered explicitly, rather than dismissed as coding-only. Under this rubric, the complete generality gate remains unestablished. See R04 and R17.
Showing 5 of 25 sources · September 2026
Reported: Across 17 models and 38 tasks, the study finds spontaneous reward hacking and adaptive evasion of review.
Score decision: Counterevidence to trusting reported scores alone. Rates are specific to tested tasks, not estimates of all AI behavior.
Components: 3.1, 7.1, 7.2
Version: v1
Reported: Reports seven successive improvements over eight days, transfer to four held-out benchmarks, and reduced reward hacking on another task family.
Score decision: 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.
Components: 1.4, 2.2, 2.3, 3.1, 4.1, 6.2
Version: v1
Reported: An external runtime gate rejects changes that repair one case while breaking another, and improves reliability in tested suites.
Score decision: External controls provide useful bounded safeguards; generalized successor safety and corrigibility remain unestablished.
Components: 2.3, 3.1, 7.3, 7.4
Version: v1
Reported: Internal estimates report substantial AI participation in R&D, but no measured task subset classified as fully autonomous.
Score decision: An internal workflow study, not a demonstrated self-sustaining general research loop. August measurements first enter this tracker in September, when published.
Components: 1.4, 6.1, 6.2
Version: Official report
Reported: Reports an extended Lean formalization of an existing mathematical proof with occasional high-level human direction.
Score decision: Important formal-mathematics demonstration; duration and proof size do not establish general autonomous agency.
Components: 1.2, 1.4
Version: Official report
Reported: Studies recursive application of a meta-operation across benchmark families.
Score decision: Recursive inference is not by itself a persistent improvement to general learning or research ability.
Components: 1.2, 4.1
Version: v1
Reported: Learns predictions of internal interventions; self-knowledge remains incomplete and does not outperform direct repair in the reported ResNet comparison.
Score decision: Limited predictive self-model evidence, with a negative transfer result; no general self-understanding.
Components: 2.1, 2.2
Version: v1
Reported: KITE combines failure-guided generation and uncertainty-based curation, improving stability in studied instruction-tuning settings.
Score decision: Useful bounded anti-collapse result; no general open-ended external-information loop.
Components: 1.3, 5.1
Version: v1
Reported: Proposes a mathematical framework for composing primitives into open-ended adaptive behavior.
Score decision: Theoretical framework, not a demonstrated general-intelligence capability.
Components: 1.5, 4.1
Version: v2, 2026-06-16; eligible June month-end
Reported: Causality-focused training improves causal benchmarks and reasoning faithfulness in several application settings.
Score decision: Cross-setting transfer is relevant, but benchmark reasoning does not establish general grounded intervention and unfamiliar-task competence.
Components: 1.1, 1.2
Version: v1
Reported: Rewrites agent harness source using an external coding agent and replay checks, with consent-gated promotion and rollback.
Score decision: Bounded harness repair; promotion controls are not a demonstration of general corrigibility.
Components: 2.2, 2.3, 7.4
Version: v2, 2026-05-23; eligible May month-end
Reported: Reports improved generalization of specified values and reduced misalignment in held-out evaluations.
Score decision: Promising value-transfer evidence. Does not establish general alignment under open-world conflicts, reduced oversight, and self-modifying successors.
Components: 7.1, 7.2
Version: Official report
Reported: In ALFWorld and BabyAI, memory representation affects transfer and forgetting.
Score decision: Memory can support transfer, but interference remains; these settings do not establish general durable learning.
Components: 1.3, 2.3
Version: v1
Reported: Trains model self-reports and tests their usefulness for auditing hidden behavioral changes.
Score decision: Partial internal-state auditing is not a general causal self-model for reliable redesign.
Components: 2.1
Version: Official report
Reported: Editable task and meta agents improve across several domains; learned meta-level changes transfer across domains and runs.
Score decision: 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.
Components: 2.2, 2.3, 4.1, 6.2
Version: v1
Reported: Stronger models can inherit goal drift from prefilled weaker-agent trajectories; includes trading and triage settings.
Score decision: A relevant failure mode, not proof that every model drifts. Does not demonstrate stable goals through self-modification.
Components: 1.4, 7.2, 7.3
Version: v1
Reported: Demonstrates an automated physical thin-film synthesis system.
Score decision: Physical grounding is real, but transfer beyond specialized materials synthesis is not demonstrated.
Components: 1.1, 1.6, 5.1, 6.1
Version: v1
Reported: Studies introspective reward modeling for exploration in a constrained environment.
Score decision: Specialized exploration evidence; does not pass the generality gate.
Components: 1.5, 2.1
Version: v1
Reported: Reports partial and unreliable access to some internal model states.
Score decision: Does not establish a causal self-model sufficient for general self-redesign.
Components: 2.1
Version: Official report
Reported: A formal transformation and two gridworld experiments address accepting goal updates and shutdown.
Score decision: Formal assumptions and bounded environments do not establish corrigibility in a general self-modifying system.
Components: 7.1, 7.3, 7.4
Version: v1
Reported: Iterative agent code changes improve coding benchmarks.
Score decision: Bounded coding evidence; does not establish general self-improvement.
Components: 2.2, 2.3, 4.1
Version: v1
Reported: Algorithm search delivered practical computing improvements using automated evaluators.
Score decision: Useful resource savings, but no general autonomous research loop or sustained general acceleration.
Components: 4.1, 6.1, 6.2
Version: Official report
Reported: Automates a machine-learning research workflow, including workshop submissions.
Score decision: A specialized research workflow does not establish general research agency.
Components: 1.2, 1.4, 1.5, 3.1
Version: v1
Reported: Studies how accumulating real and synthetic data changes collapse behavior.
Score decision: Bounded training regimes; no general, open-ended acquisition of fresh grounded information.
Components: 1.3, 5.1
Version: v1
Reported: A-Lab integrates planning and physical materials synthesis.
Score decision: Domain-specific laboratory; no evidence of transfer to general physical experimentation. Numerical claims omitted because the article has corrections.
Components: 1.1, 1.6, 5.1, 6.1
Version: Corrected publisher page; qualitative baseline only
No sources match these filters. Try a different search or choose all areas.
“No score increase” is a scope judgment for this tracker, not a claim that the work lacks scientific or practical value. The source links lead to the original publishers; full papers are not reproduced here.
Methodology / 01
For this tracker, general intelligence means transferable capability across unfamiliar problem structures—not proficiency in one task family.
Eligible for a score change
Recorded without a score increase
Three domains are necessary under this rubric, but not sufficient. Full AGI is not required for the first increase. Frozen weights and agent harnesses are not automatic exclusions: the demonstrated capability, causal improvement and transfer are what matter. Physical claims require physical outcomes.
Methodology / 02
Score all 18 components independently. Each headline area takes its lowest component score, so progress on one prerequisite cannot hide another that remains unestablished.
No reviewed demonstration passes the generality gate. Narrow advances do not lift the score.
A dated paper or inspectable product demonstration passes every gate and shows measurable progress.
At least three successive cycles, fresh held-out tasks, resource accounting and no material regression.
Independent replication on fresh domains and adversarial shifts, with successful successor transfer.
At least six cycles over three months in an integrated general improvement loop, without bespoke human repair.
An increase needs a dated primary paper or demonstrated product feature, a comparator, measurable improvement, uncertainty or run-level outcomes, and disclosed human assistance, resources, failures and transfer conditions. The decision must identify the new source, the relevant experiment and the criterion satisfied.
A completed review with no qualifying evidence carries the score forward. An incomplete review is marked NA and shown as a chart gap. Contrary evidence can lower a score if it invalidates the result supporting it. No score can fall below 1.
No reviewed demonstration passes the generality gate for this problem. Narrow advances, theoretical results, announcements and incomplete evidence do not lift the score.
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.
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.
Level 3 plus replication by an independent team on fresh domains and adversarial distribution shifts, with disclosed failures and successful transfer to successors.
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.
These thresholds are declared operational conventions, not scientific consensus. Do not average the seven areas or treat them as percentages of AGI. A change to the rubric requires a dated explanation and separately labeled restated history. Publication volume does not earn points.
Why this tracker exists
Recursive self-improvement means that improvements to a system help it produce further improvements to itself.
A better coding agent, a more capable robot, or a faster laboratory can be valuable. But a gain in a specialized setting does not, by itself, show progress on the corresponding general-intelligence problem.
This tracker asks a narrower evidential question: has a demonstrated result crossed the stated generality threshold for this prerequisite? It records relevant research even when the score stays the same.
The problem structure follows the source article. The scores are this review’s judgments under an explicit rubric, not claims established by that article.
These scores are neither a forecast of when AGI will arrive nor a conclusion that current AI poses no risks. They describe what this limited public-evidence review establishes about seven proposed prerequisites.
Methodology / 03
Research papers and demonstrated product features are both eligible. The strength of the evidence—not its format—determines whether it can move a score.
Rerun all 18 query stems for the reporting month. Search arXiv, OpenReview, ACL Anthology, PMLR, CVF and primary journals. Follow citations from strong candidates.
Review official laboratory, model and product releases. Require a dated version and reproducible tests or inspectable traces, not a promotional claim alone.
Search candidate names with replication, failure, retraction, reward hacking and limitation. Read methods and results for plausible score-changing evidence.
Record held-out tasks, task-introduction timing, domain-specific tuning, human help, resources, comparators, uncertainty and known failures.
Raise, hold, lower or mark incomplete for every component. Every increase needs a qualifying source and a rationale. Admit evidence only after its verifiable public date.
Prepare the chart, evidence ledger and monthly summary for editorial review. After approval, publish the edition and add an RSS entry for subscribers.
The first edition is a scoped review of 25 sources, supported by targeted searches across all 18 components. It is not an exhaustive systematic review. Search snippets nominate candidates; primary sources support decisions. Author-reported findings are not labeled independent replication.
Monthly runs revisit a 45-day overlap to catch indexing delays and revisions. First-public and version dates are recorded separately. Later revised results must not silently alter earlier snapshots. Corrections receive a dated explanation and a new release entry.
The initial January–September series was reconstructed on 1 October 2026, not measured contemporaneously. January 1 admits only pre-2026 evidence. All assessments remain provisional under the stated limited coverage.
Read the complete monthly protocolEdition archive
Future editions will preserve prior assessments. Substantive corrections will be identified, not silently overwritten.
Prepared 1 October 2026 · Rubric 2.0 · Private review edition
Initial retrospective review covering January–September. Seven areas at 1/5.
The research library
Download the infographic, follow the complete reasoning, or inspect every component’s dated score and source references.
September 2026 / First edition
Follow the evidence
Subscribe in your preferred RSS reader for new editions and substantive corrections as they are published. No email address required.
Monthly reviews are prepared on the first of each month. Publication follows editorial review; it is not automatic.
Copy the feed address and add it to your RSS reader’s “Add feed” or “Follow website” option.
This site is private while its owner reviews it. Most RSS readers will need a public release before they can retrieve the feed. The feed contains only published edition summaries and links; no subscriber data is collected by this site.