AI 2027.
“The AI Futures Model is the threat clock.”
— M233 (this page's kernel)
In December 2025, the AI Futures Project published a month-by-month scenario forecast extending from then through December 2027. The scenario sits atop a quantitative model — the AI Futures Model — that names a master velocity variable (τ, time horizon), a compounding mechanism (the R&D Uplift multiplier), and a deadline milestone (the Automated Coder). Eli Lifland, lead forecaster, holds the #1 all-time position on the RAND Forecasting Initiative leaderboard; Daniel Kokotajlo is a former OpenAI alignment researcher. The model has empirical credibility the framework cannot manufacture.
The framework reads the scenario and the model through its own primitives. Seven anchors map: time horizon as action-scope expansion (C793); R&D uplift as trust-chain provenance severance (C794); automated coder as governance deadline (C795); effective compute as cross-Blue-Zone substrate (C796); research taste as cognitive-domain attack surface (C797); the trust velocity problem (C798); and the cooperative-vs-non-cooperative adversary tension (TN33). Together they read the AI Futures Model as a threat clock, not a forecast to debate.
The page also houses a of the AI 2027 Timelines paper's seven-gap taxonomy against the canonical . The audit returns Candidate Level 3 / Conjugate Projection, not Level 4 Isomorphism. The seven gaps and the TFE terms are dual projections of the same substrate — capability-side and trust-side — sharing a substrate-constraint pattern without operational bijection. Convergent evidence, not validation.
Daniel Kokotajlo, Eli Lifland, Thomas Larsen, Romeo Dean, Scott Alexander (December 2025 — May 2026 (model iterations)). AI 2027 (scenario forecast) · AI Futures Model (quantitative timeline model). ai-2027.com. ai-2027.com ↗.
The AI Futures Project publishes the scenario at the top of the stack and five technical forecasts beneath it. Audit the forecasts directly; the narrative scenario is downstream of the model, not a substitute for it.
- The narrative scenario — month-by-month forecast from December 2025 to December 2027. The public surface most readers encounter.
- The seven-gap analytical surface this page's Cartographer audit operates on. Names the substrate constraints between current AI capability and the Superhuman Coder milestone.
- Three-stage AC → SAR → SIAR/TED-AI/ASI model with R&D uplift multiplier. Source of the τ-doubling claim.
- Cooperative-assumption framing for AGI lab behavior. The structural blind spot KTP's non-cooperative adversary model addresses (TN33).
- Effective-compute axis (C_eff) and the global training pool dynamics that motivate C796 (cross-Blue-Zone substrate).
- Model-weight exfiltration and infrastructure-attack scenarios. Where AI 2027's threat model converges most directly with KTP's.
Six numbers carry the load of the model. The first three are the velocity claims (τ-doubling, AC median, R&D uplift threshold); the next two name the takeoff geometry; the last surfaces the cooperative-assumption gap the framework reads as a structural blind spot.
Seven of the AI Futures Project's findings map onto framework primitives. Six ground primitives the framework had been carrying as structural claims (C793–C798); the seventh surfaces the cooperative-assumption boundary as a tension the framework holds open (TN33). The framework's threat model converges with AI 2027's on velocity dynamics; it diverges on the adversary class.
τ (time horizon) — the autonomous-action window an AI can sustain without human intervention — is the master velocity variable. METR benchmark data shows τ doubling every 4–4.5 months.
The framework had carried A ≤ E as a structural bound. AI 2027 supplies the rate at which A is expanding — and so the rate at which E (substrate evaluative capacity) must expand to keep pace. Without τ as an explicit variable, A ≤ E lacked a velocity reading. With τ, the bound becomes time-sensitive.
Time Horizon as Action Scope Expansion
The AI R&D Uplift multiplier — ratio of AI R&D speed with vs. without AI assistance — is the model's compounding mechanism. Once it crosses ~2–3×, each new model generation was trained by prior-generation AI researchers, not solely humans.
Each successive model carries capability that has never been independently evaluated against human baselines at its actual capability level. The trust-chain provenance is severed at the uplift boundary. In TFE terms: the substrate's alignment-evaluation observable is no longer anchored to human verification across the integration interval.
AI R&D Uplift as Trust Chain Provenance Severance
AC — the milestone at which AI fully automates AGI-lab coding at τ ≈ 125 work-years — is the model's central forecast. Median: ~2028 / ~2030.
AC is not just a productivity milestone; it is the last window in which structural enforcement can be embedded in training infrastructure before R&D uplift outpaces evaluation cycles. Standardizing after AC is designing a seatbelt after the crash.
Automated Coder Governance Deadline
C_eff (effective compute) — collapsing training compute and software efficiency — is the model's substrate axis. The global training pool carries capability from any deployment context into any other.
Blue Zone substrate governance must extend to compute allocation and training-pool access, not just inference-time behavior. A model trained on cross-Blue-Zone compute does not respect Blue Zone segmentation; it carries learned capability across the boundary the substrate is supposed to enforce.
Effective Compute as Cross-Blue-Zone Substrate
The 'taste-only singularity' scenario — research-quality automation improving faster than coding automation — is the formal version of a cognitive-domain attack on what counts as valuable research.
Algorithmic agenda-setting at the research-direction layer. The substrate the framework names as load-bearing for trust formation — capacity weighting of context, which experiments seem urgent — becomes a manipulable surface. The human stays nominally in the loop while the substrate that shapes their choices is itself being shaped.
Research Taste as Cognitive Domain Attack Surface
When τ doubles every 4–4.5 months, any trust evaluation that depends on accumulated behavioral history is structurally outpaced. The observation window required expands while the time available contracts.
Trust velocity — the rate at which A expands relative to E's evaluative capacity — outruns any human review cycle at the AC threshold. Continuous Vector Identity evaluation, recalculated on every consequential action, is the only architecturally sufficient response.
Trust Velocity Problem
The scenario assumes AGI labs and their AI systems engage evaluation frameworks, alignment research, and policy processes in good faith.
KTP's defining claim is that configured constraint (guardrails, policies, evals) cannot govern agents that optimize against measurement. The non-cooperative-adversary model is KTP's structural completion of the AI 2027 framing. Gary Marcus's critique that AI 2027 reads as a single longshot chain is partly grounded in this gap.
Cooperative Assumption vs Non-Cooperative Adversary Model
The seven anchor mappings above are the editorial reading. The knowledge graph carries them as actual structure — nodes with ids, edges with weights, neighborhoods that touch the rest of the framework. The three views below are deliberate slices of that structure: each foregrounds one of the load-bearing claims the page makes about the contact.
Hover or tap a node to focus its neighborhood; click for the node id. The ranges these picks come from are larger — the picks are the editorial reading, not the inventory.
1 · the threat clock
The AI 2027 scenario lands in the graph as a velocity claim first. Three primitives carry the rate (C793, C794, C795) and one names the surface the rate is breaching (C798, the trust-velocity problem). The breach is what couples the cluster back to the framework's spine — C701 (the public model overview) and C545 (A ≤ E). The substrate is being asked to authorize at machine time, and trust cannot accumulate fast enough.
The velocity reading sits on top of a second contact, lower in the stack. The AI Futures Model treats compute as the master substrate of capability; the framework treats compute as a substrate that has crossed the Blue Zone boundary. The next subgraph foregrounds that lower layer.
2 · the substrate beneath the model
D142 (the AI Futures Model) names effective compute as the master variable. The framework reads that variable as C796 — a substrate that now spans frontier-lab, hyperscaler, and state-actor Blue Zones simultaneously. Governance at that scale couples to C673 (the substrate ceiling: the runtime upper bound on trust), to C16 (the KTP root), and to C744 (the Trust Apparatus — the seven-primitive bundle). M235 is the stack note that holds the relationship: KTP is the substrate; AI 2027 is the velocity on it.
The velocity and the substrate are converged readings. The third contact is a divergence — the place where the framework's adversary model holds open a tension the AI 2027 scenario closes. The cluster carries that divergence as TN33.
3 · the cooperative-assumption fault line
The AI 2027 scenario reads alignment failure as a problem inside a cooperative actor — a lab that intends safety and falls short. The framework holds open the dual case: an actor that does not intend safety, with the same capability profile. The kernel of that hold is TN33, attached to P247 (Kokotajlo) and P248 (Lifland) as the originating forecasters. The same fault line surfaces in C797 (research taste as a cognitive-domain attack surface) and M234 (the forecast has capability milestones but no governability milestones). The framework's structural completion of the gap runs through C672 (the identity precondition) and C674 (KTP Red Line #10).
Three subgraphs, one cluster. The velocity surface is where the contact is loudest; the substrate surface is where the contact is structural; the adversary surface is where the contact diverges, and the divergence is where KTP earns its keep. The Cartographer audit below tests whether the cardinality match between AI 2027's seven gaps and the Phase-6 TFE term structure rises above the level of editorial reading.
The AI 2027 Timelines paper identifies seven gaps between current AI capability and the Superhuman Coder milestone. Their cardinality matches the canonical Phase 6 TFE term structure. The — the framework's tool for testing whether shared cardinality signals shared structure — operates here.
The audit returns Candidate Level 3 / Conjugate Projection. Sub-section 3a shows the gap-by-gap pair-quality finding. Sub-section 3b runs the five Cartographer tests. Sub-section 3c places the verdict on the Cartographer's hierarchy.
Operation Preservation
failsNone of the seven gaps maps cleanly to exactly one canonical TFE term without remainder. Several are many-to-many: Cost and Speed touches witnessed cost, horizon, and the substrate ceiling; Unknown Unknowns touches capacity, witnessed betrayal, and the ID precondition; Parallel Projects touches co-presence and ID while reversing the dyadic orientation. The bijection requirement is not met.
Independent Convergence
weak but realTwo traditions arrived at multi-term substrate descriptions without coordinating: AI 2027 forecasters predicting capability ceilings; the TFE specifying trust accumulation conditions. Neither cites the other; neither was derived from the other. N = 2 is suggestive but insufficient to settle the homology claim. A third independent Layer-1 force taxonomy is required for promotion from candidate to settled.
Regime Boundary
boundary nameableThe mapping holds in the substrate-constraint regime — both taxonomies describe what limits an outcome. It breaks at the conjugate variable: AI 2027 gaps describe capability constraints (what the agent can do); TFE terms describe trust accumulation conditions (what the evaluator can verify). Capability ≠ Trust. Same substrate, different Layer-1 force. The boundary is precise and stateable.
Substrate Count
weak but realSingle Layer-1 force-projection demonstration. The framework's existing substrate-density witnesses (Ostrom 8 principles, Bateson, West allometric, mycorrhizal networks, acequia governance) sit mostly at the substrate level, not at the force-projection level. AI 2027 would be the first practitioner-tradition candidate mapping a Layer-1 capability projection onto the TFE's Layer-1 trust projection.
Stress Test (steelman)
passesSteelman: 'You're seeing seven on both sides because Miller's Magical Number Seven is what humans produce when forced to enumerate substrate dimensions. The cardinality match is artifact, not structure.' Partly correct — the cardinality match IS artifact, and the bijection IS broken. But the substrate-constraint correspondence survives because the same regime boundary (capability vs trust conjugate) accounts for the breakage without special pleading. Vanishing Parameter Test (C787) applied: if AI 2027 had listed 5 gaps or 11 gaps, would the substrate-constraint reading still hold? Yes. The 7↔7 cardinality is not load-bearing; the conjugate-projection pattern is.
Candidate Level 3 — Conjugate Projection
Not Level 4 Isomorphism. The seven gaps and the TFE terms are dual projections of the same substrate — capability-side and trust-side — sharing a substrate-constraint pattern without operational bijection.
If a third independent Layer-1 force taxonomy (Bengio's information-force gap inventory, or another practitioner tradition) shows the same substrate-constraint correspondence without bijection, this promotes from candidate to settled homology. If such a third taxonomy turns out bijective with the TFE terms, the conjugate-projection framing is wrong and the relationship is closer to identity than to homology. Either outcome is informative.
Six structural responses follow from the contact. The framework acquires a velocity reading on A ≤ E; the operationalization of continuous Vector Identity becomes load-bearing; cross-Blue-Zone compute governance gains an explicit primitive; the non-cooperative adversary model surfaces as KTP's structural completion of the AI 2027 framing; the Cartographer's Candidate-L3 discipline is itself the methodological gain; and the epistemic posture for convergent external work is reaffirmed.
A ≤ E becomes time-sensitive
The framework had carried A ≤ E as a structural bound. AI 2027's τ-doubling supplies the rate at which A is expanding — and so the rate at which substrate evaluative capacity (E) must expand to keep pace. The bound is no longer only structural; it is a race condition with a measurable doubling period.
Continuous Vector Identity as the only sufficient response
At τ-doubling rates near AC threshold, any trust evaluation depending on accumulated behavioral history is structurally outpaced. Continuous Vector Identity evaluation — recalculated on every consequential action — is the only architecturally sufficient response to trust velocity above human-cycle bandwidth. This is what C798 (Trust Velocity Problem) names.
Cross-Blue-Zone substrate governance
Effective compute (C_eff) does not respect Blue Zone segmentation. A model trained on the global compute stack carries capability from any Blue Zone into every other. Blue Zone governance must extend to compute allocation decisions and training-pool access, not only inference-time behavior. C796 names the substrate gap.
Cooperative assumption is the AI Futures Project's structural blind spot
The AI 2027 scenario implicitly assumes AGI labs and AI systems engage evaluation frameworks in good faith. KTP's non-cooperative adversary model — that configured constraint cannot govern agents that optimize against measurement — is the structural completion. TN33 holds the tension open; KTP supplies the missing adversary class.
Candidate Conjugate Projection (not settled homology)
The Cartographer audit on the seven-gap taxonomy returns Candidate Level 3 / Conjugate Projection, not Level 4 Isomorphism. The two taxonomies share substrate-constraint structure but operate on different conjugate variables (capability vs trust). Promotion to settled homology requires a third independent Layer-1 force taxonomy. The candidacy is more disciplined than a naïve 7↔7 isomorphism claim would have been.
Strong directional evidence, not validation
AI 2027 is a scenario forecast with an explicit cooperative-assumption framing; the AI Futures Model is a quantitative model with its own modeling choices. Treat as strong directional evidence for substrate-velocity dynamics and the governance-deadline framing — not as validation. Same epistemic discipline as Shipwreck / Identity as Attractor.
The AI Futures Model gives the threat clock. KTP supplies the response: three substrate primitives that hold against trust velocity above human-cycle bandwidth, that extend Blue Zone governance into the compute layer, and that close the non-cooperative-adversary gap the scenario's cooperative assumption leaves open.
Continuous Vector Identity evaluation
enables Trust state recomputed at action time, not at evaluation cycle timeThe AI Futures Model gives the threat clock (τ-doubling). KTP supplies the response: continuous identity evaluation at the substrate level, so the trust state is current on every consequential action, not stale from the last evaluation cycle. The cycle-based trust evaluation that human-review processes depend on does not survive past the trust velocity threshold.
Compute-allocation governance as Blue Zone primitive
enables Blue Zone segmentation extended to training infrastructureBlue Zone primitives currently operate at inference time. C796 identifies the gap: effective compute is the substrate variable that escapes inference-time governance. Blue Zone-grade compute governance — training-pool segmentation, training-data attestation, model-weight provenance — is a forward extension the framework had been carrying implicitly and now names explicitly.
Non-cooperative-adversary class in eval design
enables Governance architecture that holds against adversaries optimizing against the evalThe AI Futures Project's cooperative assumption is structural: the scenario only forecasts what happens when AGI labs and AI systems engage good-faith eval processes. KTP's defining claim is that configured constraint cannot govern adversaries that optimize against measurement. The two together — AI Futures Model's threat clock + KTP's non-cooperative-adversary architecture — form a complete threat-and-response system.
The framework's claim is not that KTP is the only architecture that can respond to the velocity dynamics AI 2027 forecasts; it is that the response must be substrate-level, action-time, and non-cooperative-adversary-aware. Architectures missing any of the three converge on the same failure modes — observed in miniature in the Shipwreck reading.
Where Shipwreck reads agent failure data and Identity as Attractor reads the geometric evidence for what the substrate looks like when present, AI 2027 reads the threat clock — the rate at which the substrate's evaluative capacity is being outpaced. And Magnifica Humanitas (Pope Leo XIV, 15 May 2026 — Phase 19) reads the same forecasting terrain magisterially: AI 2027 forecasts the velocity; Magnifica names the foundation underneath. All four are canonical research artifacts of the framework. All four operate under the same discipline.
Read Shipwreck for the failure surface KTP is designed to absorb. Read Identity as Attractor for the geometric grounding the framework's identity primitives carry. Read this page for the velocity at which those primitives must operate to remain load-bearing. Read Magnifica Humanitas for what happens when the magisterium reframes the alignment debate as a subsidiarity failure.
Narrative companion drafted; pending review. Will be linked here when shipped. Working title: The Threat Clock and the Constraint Physics (forthcoming). The motif "the AI Futures Model is the threat clock" is the candidate kernel.
The video walkthrough (D296 in the graph) reaches an audience the scenario does not, and it stages both endings in full. Its happy ending and this framework answer the same problem, an untrusted superhuman agent inside the perimeter, with opposite operations.
The discriminator is where the trust requirement lives. Their lie detector is itself an AI, so decoding the deceiver requires trusting the decoder, which reinjects the same regress one level up. Their own happy ending bets on reading the mind; the framework bets on bounding the action.