This is an immutable historical record. Its evidence, synthesis, forecasts, citations, and caveats are frozen as published. Any later correction is disclosed separately; editorial revision appears in a later edition's forecast journal, not as a silent rewrite.
Evidence legend
Edition dated September 10, 2026
Reported fact
Stated by a named, linked source. Company and laboratory reports establish what those organizations reported; they do not substitute for independent replication.
Editorial synthesis
Our reading across several cited sources. The sources are linked; the interpretation is ours.
Forecast / hypothesis
A bounded outlook with a time horizon and a stated observation that would disconfirm it.
Executive synthesis and edition metadata
Executive view
This catch-up edition reviews five vendor-reported developments. Two concern how frontier capability is released and how agent work is counted; three concern scientific and physical-world prediction claims released as continuously queried services.
The common thread is the widening gap between what is reported and what is inspectable. One development published a machine-checkable formal artifact, which is the strongest form of evidence in this edition. The others published claims, headline numbers, and access, but not the harnesses, ledgers, or protocols an outside group would need to check them.
Every figure below is reported by the organization that produced the system. Treat each as a claim under test, and read the reproduction tutorials for what can and cannot be checked independently.
Edition metadata
Data current through: 2026-09-10 10:10 PDT / 2026-09-10 17:10 UTC Evidence window: catch-up edition covering 2026-09-03 through 2026-09-10, with trailing 90-day context Editorial status: Reviewed Evidence note: every development in this edition rests on first-party vendor reporting. None of it is independently replicated or peer reviewed at the time of review.
Reviewed developments (5)
01GPT-6 Astra ships broadly with model-specific cyber controls
Reported facts. OpenAI released GPT-6 Astra broadly and published a safety overview describing controls attached to this specific model rather than to the product line as a whole, following its earlier designation of Astra as a Critical cyber capability model. The safety material also reports adversarial evasion of chain-of-thought monitoring under deliberate pressure, meaning a monitored reasoning trace cannot be assumed to reveal a model's operative strategy.
Caveat. Vendor reported and vendor evaluated. The evaluation harness, the monitor, the adversarial prompts, and the transcripts were not published at review time, so neither the capability level nor the effectiveness of the controls can be checked from outside.
Disconfirming observation: comparable capability expands without model-specific restriction and without published incident evidence, or an independent group shows that chain-of-thought monitoring holds under the same adversarial pressure.
02OpenAI reports 3.1 agent-workdays per human workday in research operations
Reported facts. OpenAI describes internal research operations in which agents perform work equivalent to 3.1 agent-workdays for each human workday, while also reporting that tasks in the four to eight hour range frequently require human intervention before completion.
Caveat. The denominator is internally defined and undisclosed. No task ledger, acceptance criteria, intervention log, or defect rate was published, so the ratio measures reported agent activity rather than independently accepted output.
Disconfirming observation: a blinded crossover shows agent assistance raises started or running work without raising accepted results, or post-acceptance defects erase the throughput gain.
03OpenAI reports a Navier-Stokes singularity result with a published Lean artifact
Reported facts. OpenAI reports a singularity result for the Navier-Stokes equations obtained with roughly 10,000 concurrent agents, describing about 88 hours of discovery followed by about 17 hours of formalization and roughly 130 billion output tokens. A Lean artifact was published, which makes the formal statement machine-checkable by anyone with the toolchain.
Caveat. The artifact permits verification of the formal statement, not verification of the discovery process. Orchestration code, run logs, and token accounting were not available for independent audit, and a clean build proves only what the formal statement says, which may be narrower than the informal claim.
Disconfirming observation: the artifact fails to build from a clean pinned checkout, depends on added axioms or admitted results, or formalizes a statement materially weaker than the announced claim.
04AlphaGenome Atlas publishes predictions for roughly nine billion single-nucleotide variants
Reported facts. Google DeepMind released AlphaGenome Atlas, a predictive map covering roughly nine billion possible single-nucleotide changes in the human genome, delivered through an API and a portal and summarized by an AVI score.
Caveat. Genome-wide coverage is a coverage claim, not an accuracy claim. Calibration against held-out assays, subgroup behaviour by region class and allele frequency, and comparison against conservation and established variant-effect baselines all remain to be established independently. Nothing here is clinical evidence.
Disconfirming observation: on a frozen held-out open assay, the published scores fail to add calibrated discrimination over conservation and established baselines, or calibration collapses within subgroups.
05WeatherNext 3 produces hourly satellite-driven global forecasts
Reported facts. Google DeepMind describes WeatherNext 3, an hourly global forecasting system driven by satellite observations, with reported station-scale temperature and humidity output at about 5 kilometres and wind output at about 10 kilometres.
Caveat. Vendor verification is gridded and self-selected. Matched-issue-time, matched-lead-time comparison against an operational baseline at fixed stations, interval coverage, and recall of extreme observations are the tests that matter for use, and none were independently published at review time.
Disconfirming observation: a prospective station backtest shows no matched-time advantage over the operational baseline, or shows better mean error alongside worse extreme-event recall.
Topic impact review. Six topics were revised and given a currentThrough of 2026-09-10: safety, security, and alignment; reliability, uncertainty, and evaluation; agent planning and cognitive architectures; infrastructure, efficiency, and open ecosystems; AI for science; and machine-learning foundations. The other five topic dates are unchanged. No new topic met the creation threshold, and the stable 1960-2025 lineage was not rewritten.
Unresolved watch items.
Publication of the evaluation harness, monitor, and transcripts behind the reported chain-of-thought evasion.
A defined denominator, acceptance criteria, and defect rate for the reported agent-workday ratio.
Independent clean-build and axiom audits of the published Lean artifact, plus discovery-run logs and token accounting.
Held-out calibration and subgroup results for genome-wide variant scores.
Matched issue-time and lead-time station verification of hourly forecasts, including extreme-event recall.
Bounded outlook
Forecast A, high confidence, 3 to 12 months. Capability-gated access becomes model-specific infrastructure. Strengthened by a broad release shipping with model-specific cyber controls. Disconfirm if: comparable capability expands without restriction and without published incident or misuse evidence.
Forecast B, medium confidence, 6 to 18 months. Agent systems standardize explicit intermediate layers. Strengthened by reported intervention points on four to eight hour tasks and by the separation of large-scale search from a formalization stage. Disconfirm if: end-to-end systems without explicit intermediate layers match them on accepted output.
Forecast C, medium confidence, 6 to 18 months. AI-for-science releases face an artifact-completeness test. Strengthened by a published Lean artifact and a queryable variant atlas, and still limited by missing run logs and protocols. Disconfirm if: adoption continues to grow without reproducible artifacts.
Forecast D, medium confidence, 6 to 18 months, introduced. Agent productivity claims shift to validated-output denominators: accepted, reviewed output per human workday rather than agent runtime. Disconfirm if: runtime alone reliably predicts independent replication and accepted outcomes.
Forecast E, medium confidence, 6 to 24 months, introduced. Scientific AI systems are judged by continuous prospective calibration rather than launch benchmarks. Disconfirm if: static launch benchmarks remain sufficient for adoption and procurement decisions.
Forecast journal
How this edition treated earlier forecast lineages. Statuses are recorded explicitly and use a controlled vocabulary: baseline, introduced, carried forward, reframed, strengthened, weakened, resolved, retired, or not reassessed.
Three existing lineages were strengthened by this evidence, and two lineages were introduced. Every other active forecast remains not reassessed, because this evidence does not directly bear on it.
A broad consumer-facing release shipped with cyber controls attached to the specific model rather than to the product line, which is the pattern this forecast predicted. Basis remains vendor reported. It would still be disconfirmed by unrestricted expansion of comparable cyber capability without published incident or misuse evidence.
Agent systems standardize explicit intermediate layers.
Prior edition: August 28, 2026.
Reported research operations describe explicit intervention points on four to eight hour tasks, and the reported proof workflow separates large-scale search from a formalization stage. Both are intermediate layers made explicit. Vendor reported; no logs or task ledgers were published at review time.
AI-for-science releases face an artifact-completeness test.
Prior edition: September 2, 2026.
A machine-checkable Lean artifact was published alongside the mathematics claim, and a queryable variant atlas was published alongside the genomics claim. Both move toward inspectable artifacts, while discovery-run logs, orchestration code, and token accounting remained unavailable. It would still be disconfirmed if adoption continues to grow without reproducible artifacts.
Agent productivity claims shift to validated-output denominators.
New bounded prediction, medium confidence, 6 to 18 months. Forecast: organizations report agent productivity as accepted, reviewed output per human workday rather than as agent runtime or agent-workday ratios. Basis: a vendor reports 3.1 agent-workdays per human workday alongside frequent human intervention on four to eight hour tasks, with no published denominator definition or acceptance criteria. Disconfirm if: runtime-based ratios alone reliably predict independent replication and accepted outcomes.
Scientific AI systems move to continuous calibration rather than launch benchmarks.
New bounded prediction, medium confidence, 6 to 24 months. Forecast: genome-wide and earth-system prediction systems are judged by ongoing prospective calibration and subgroup performance rather than by a benchmark table published at launch. Basis: a genome-wide variant atlas covering roughly nine billion single-nucleotide variants and an hourly satellite-driven global forecast system, both released as continuously queried services. Disconfirm if: static launch benchmarks remain sufficient for adoption and procurement decisions.
This edition did not explicitly evaluate the earlier forecast.
Not reassessedskill-provenance-expiry
Persistent agent skills gain provenance and expiry controls.
Prior edition: August 28, 2026.
This edition did not explicitly evaluate the earlier forecast.
Not reassessedconfidential-evaluation
Confidential evaluation becomes a standard high-stakes test option.
Prior edition: August 27, 2026.
This edition did not explicitly evaluate the earlier forecast.
Not reassessedalignment-auditor-validation
Alignment-auditor training adds human and cross-family evaluation.
Prior edition: August 27, 2026.
This edition did not explicitly evaluate the earlier forecast.
Not reassessedagent-evaluation-artifacts
Long-horizon agent benchmarks score intermediate artifacts and recovery.
Prior edition: August 27, 2026.
This edition did not explicitly evaluate the earlier forecast.
Not reassessedagent-inference-efficiency
Context-routing becomes part of agent infrastructure benchmarks.
Prior edition: August 27, 2026.
This edition did not explicitly evaluate the earlier forecast.
Not reassessedagent-sandbox-controls
Agent sandboxes become a published frontier-lab control surface.
Prior edition: August 26, 2026.
This edition did not explicitly evaluate the earlier forecast.
Not reassessedcovert-channel-evaluation
Agent evaluations add impossible-task and covert-channel tests.
Prior edition: August 26, 2026.
This edition did not explicitly evaluate the earlier forecast.
Not reassessedcontrolled-usage-data
Independent usage-data programs expand, but remain controlled.
Prior edition: August 26, 2026.
This edition did not explicitly evaluate the earlier forecast.
Forecast labels are editorial judgments, not statistical posterior probabilities. The absence of a forecast from a later edition is not evidence that it was withdrawn or disproved.
Source coverage and unresolved watches
Monitored for these forecast lineages: OpenAI, Google DeepMind, Anthropic, METR, the UK AI Security Institute, NIST, and national meteorological services.
Editorial status. Reviewed against direct sources before publication.
