keep score in public
The predictions ledger
Every issue carries a falsifiable call with a stated confidence. Every due call gets settled, right or wrong, and scored with Brier — lower is better, 0.25 is a coin flip. Credibility is the product.
63open
0–0 record
— mean Brier
0settled
Calibration
The open book by confidence — and, as calls settle, whether the stated confidence matches the real hit-rate (bar = actual, tick = claimed).
| Confidence | Open | Settled | Actual vs claimed |
|---|---|---|---|
| 40–50% | 1 | · | — |
| 50–60% | 4 | · | — |
| 60–70% | 16 | · | — |
| 70–80% | 32 | · | — |
| 80–100% | 9 | · | — |
— no calls settled yet · 63 open —
The ledger
| Made | Prediction | Conf. | Due | Status | Brier |
|---|---|---|---|---|---|
| 2026-W23 | GitHub partially walks back Copilot pricing (extends promo credits past Aug, restores fallback model, or cuts the Opus multiplier) within 30 days, without reversing metering itself. | 70% | 2026-07-05 | ● open | — |
| Dive · 2026-06-07 | At least two of {GitHub, Cursor, Anthropic} ship an "unlimited on our own/house models" flat tier (subsidy internalized, frontier stays metered). | 65% | 2027-Q1 | ● open | — |
| Dive · 2026-06-10 | GitHub/npm ship branch/ref binding for OIDC trusted publishing (the actual Miasma hole) — and a worm generation defeats npm v12's script-off default before that ships. | 55% | 2026-Q4 | ● open | — |
| Dive · 2026-06-08 | A major cloud or agent platform ships an enforced hard per-task/per-agent spend ceiling (not a budget alert) that the agent cannot cross. | 45% | 2027-Q2 | ● open | — |
| Dive · 2026-06-08 | "Agent liability" insurance appears OR a cloud publishes a runaway-agent forgiveness policy, mandating spend caps/observability as a condition. | 55% | 2027-Q2 | ● open | — |
| Dive · 2026-06-09 | The top frontier-vs-best-open-model spread on a major agentic benchmark (SWE-bench/MCPMark/Terminal-Bench) stays inside ~5 pts — no lab reopens a durable capability gap, confirming the channel (not the model) is the moat. | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-06-12 | A contamination-resistant benchmark (SWE-bench Pro / SWE-rebench or successor) does NOT reproduce SWE-bench Verified's top-5 model ordering — decontamination changes rank, not just absolute scores. | 65% | 2027-Q1 | ● open | — |
| Dive · 2026-06-11 | No venture-funded independent LLM gateway/observability/eval company reaches a standalone outcome (IPO or $1B+ while independent) — the next two notable outcomes are absorptions by a model vendor or data/monitoring platform, or wind-downs. | 65% | 2027-Q2 | ● open | — |
| Dive · 2026-06-16 | No model ranked in the top tier of a major agentic benchmark (SWE-bench/MCPMark/Terminal-Bench or successor) ships meeting OSAID 1.0 in full — weights, data information, and complete training code under an OSI-approved license. "Open source AI" releases stay open-weight-only. | 80% | 2027-Q1 | ● open | — |
| Dive · 2026-06-17 | A sub-35B open-weight coding model fits a single 24GB card with a usable 128K context AND lands within ~10 points of that quarter's top frontier model on a contamination-resistant agentic benchmark (SWE-rebench / SWE-bench Pro) — i.e. the runnable-vs-competitive gap finally closes on consumer hardware. | 35% | 2027-Q1 | ● open | — |
| Dive · 2026-06-18 | Anthropic ships automatic/implicit prompt caching — a cache hit without a manually placed breakpoint — on at least one default API path, converging toward the zero-config model OpenAI, DeepSeek, and Gemini already use, because the gap between the advertised 90% discount and the realized hit rate is a cost-perception liability. | 55% | 2027-Q1 | ● open | — |
| Dive · 2026-06-19 | Multi-agent / A2A-style agent-to-agent coordination does NOT become the default shipped production-agent pattern; single-context loops plus tool-calling (the MCP rung) stay dominant, and A2A stays enterprise-announced rather than developer-used — no broad practitioner-usage signal emerges. | 75% | 2027-Q1 | ● open | — |
| Dive · 2026-06-20 | Claude Code surfaces auto-compaction control as a documented, first-class setting — a configurable threshold or a "manual / safe-point-only" compaction mode in /config or the official docs — rather than today's undocumented env var and reverse-engineered buffer numbers. | 55% | 2027-Q1 | ● open | — |
| Dive · 2026-06-21 | The next frontier-tier open-weight model release (top ~5 on the intelligence index) ships with an activation ratio at or below ~6% (active ÷ total parameters), continuing the Mixtral 27.6% → DeepSeek-V3 / GLM-5.2 ~5.4% sparsification trend; none re-ships above ~15%. | 70% | 2027-Q1 | ● open | — |
| 2026-W25 | At least one major commercial AI vendor (Anthropic, OpenAI, Google, or Microsoft) ships or formally announces a customer-facing multi-provider / bring-your-own-model fallback in a first-party developer product — positioning provider-portability as resilience against the switch-off risk the Fable 5 / Mythos 5 export ban made concrete. | 60% | 2026-09-20 | ● open | — |
| Dive · 2026-06-22 | Prompt-and-tool portability stays a manual re-eval problem — no cross-provider standard or vendor feature lets a non-trivial agent's prompt-plus-toolset move between two frontier providers and reproduce its eval scores within a small margin without per-model retuning. Gateways keep normalizing API syntax; behavior still requires bespoke adaptation. | 65% | 2027-Q1 | ● open | — |
| Dive · 2026-06-23 | No agent-native version-control system (Oak, jj-style, or similar) displaces git worktrees as the default file-isolation primitive for parallel coding agents. The major agent harnesses (Claude Code, Cursor, and peers) keep building parallel-session isolation on git worktrees — not a non-git store — in their shipped defaults. | 80% | 2027-Q1 | ● open | — |
| Dive · 2026-06-24 | Speculative decoding stays a single-stream / low-QPS latency trick. No widely-deployed variant delivers a greater-than-~1.5× throughput gain at high batch (≥64 concurrent requests) on a frontier-class model; at saturation, batching remains the dominant amortization of the weight read and the high-batch throughput multiple stays under roughly 1.5×. | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-06-25 | Claude Code does not ship a lossless / auditable auto-compaction — one that writes its kept-set to a user-inspectable file AND reliably preserves the decision rationale (the "why," not just paths and names) — so the manual dump-to-markdown + /clear handoff stays the practitioner default for long, multi-step tasks. | 65% | 2027-Q1 | ● open | — |
| Dive · 2026-06-26 | No major agent harness (Claude Code, Cursor, Codex, or peers) ships automatic tool-call deduplication — collapsing identical repeated tool invocations within a session so a retried mutating call executes once — as a documented default. Retry-safety stays the tool author's responsibility via idempotency keys / unique constraints, and the harness's only built-in stays blunt refusal of destructive operations (v2.1.183-style). | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-06-27 | No closed frontier lab (Anthropic, OpenAI, or Google) widens default-path logprob exposure beyond today's limits (Anthropic: none; OpenAI: top-20) for its flagship models — the dense soft-target leak stays closed, leaving black-box output imitation as the only available distillation route against closed frontier models. | 75% | 2027-Q1 | ● open | — |
| Dive · 2026-06-28 | DeepSeek's permanent V4-Pro price floor (~$0.44 input / $0.87 output per Mtok) does NOT ratchet materially back up (>25% on either leg) within two quarters — the open-weight-pinned floor proves structural, not a promo — AND no closed frontier lab (OpenAI/Anthropic) cuts its flagship API token price to within ~2× of that floor over the same window; they hold a premium and segment to capability instead. | 65% | 2027-Q1 | ● open | — |
| Dive · 2026-06-30 | No frontier model closes the effective-context gap — none holds ≥90% of its 4K-baseline accuracy at its FULL advertised context length on a RULER-class multi-needle test. "Just use long context" stays a cost/accuracy tradeoff, not a free win, so retrieval and routing remain the cheaper default for distinct-document workloads. | 75% | 2027-Q1 | ● open | — |
| Dive · 2026-07-01 | No public analysis demonstrates that Claude Code's request-marking is all three of: (a) high-entropy enough to uniquely single out an individual session rather than a coarse category; (b) able to survive Unicode normalization, a paraphrase, AND a copy-paste through a sanitizer; and (c) keyed to individual end users rather than reseller/category infrastructure. It stays a low-bit, strippable anti-distillation tripwire, not per-user surveillance. | 75% | 2027-Q1 | ● open | — |
| Dive · 2026-07-02 | Claude Code does NOT ship a permission-rule grammar that natively enforces intent-level Bash constraints — e.g. "curl only to an allowlisted host" holding through wrappers, redirects, and variables. Argument-constraining deny patterns stay documented-fragile, and a PreToolUse hook (plus the sandbox) remains Anthropic's own recommended enforcement path for a real boundary. | 80% | 2027-Q1 | ● open | — |
| Dive · 2026-07-03 | No published or production text watermark demonstrates AUROC ≥ 0.9 (or a true-positive rate ≥ 0.9 at 1% false-positive rate) on sub-200-token model outputs after a full recursive-paraphrase attack. Statistical watermarking stays a length-and-good-faith provenance signal — defeated on the short and adversarial case, with no scheme escaping the paraphrase floor. | 80% | 2027-Q1 | ● open | — |
| Dive · 2026-07-04 | No client-side text-as-image compression tool demonstrates ≥99% exact-string recall on code (identifiers, hashes, literals — not prose) while still cutting input tokens more than 50%. The compression-fidelity curve holds, so imaging source stays a lossy bet that corrupts high-entropy strings — AND no major provider prices a text-in-image path below its text-token rate, so it never becomes a genuine pricing arbitrage rather than a compression bet. | 80% | 2027-Q1 | ● open | — |
| Dive · 2026-07-04 | No frontier lab (Anthropic, OpenAI, Google) publicly confirms consuming llms.txt at crawl or inference time, AND MCP (callable endpoints) stays the dominant agent-distribution surface for developer tools over any passive-file standard — measured by SDK downloads and active-server count, not by how many sites publish a file. | 75% | 2027-Q1 | ● open | — |
| Dive · 2026-07-05 | Grammar-constrained / strict tool use stays opt-in per tool (not default-on) in the major agent APIs (Anthropic, OpenAI), AND at least one further frontier model release exhibits a documented tool-call schema-adherence regression on a non-strict path (invented or renamed keys, or type drift) — confirming the model's schema prior stays version-sensitive and tool-call reliability is not portable across versions without re-eval. | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-07-06 | Agent-initiated machine payments (x402 / HTTP 402 pay-per-call) stay an opt-in edge-and-crypto integration wired by hand (Cloudflare / AWS / Coinbase), NOT a runtime default — no frontier lab (Anthropic, OpenAI, Google) ships a built-in, on-by-default wallet in its first-party agent runtime that pays arbitrary 402 endpoints without per-transaction human approval. | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-07-07 | No documented real-world case shows an LLM agent gaining initial access to a patched / hardened / non-default target via a vulnerability it discovered itself (a true zero-day — not a known-class web bug fed to a team-of-agents lab harness) with no human decision gate. Agentic intrusions stay confined to known-CVE / default-credential / exposed surfaces with a human at the strategic gates, and the published autonomous find-and-exploit rate WITHOUT a CVE description stays well under ~50% on hardened real-world targets. | 75% | 2027-Q1 | ● open | — |
| Dive · 2026-07-08 | No major agent harness (Claude Code / Cursor / Codex / etc.) ships a tamper-evident run/audit log — cryptographically verifiable by a third party (signed or hash-chained, so the emitting process can't silently omit or backdate a record) — as a documented default; the built-in trail stays plain OTel telemetry plus git history (author-trusted), and Halo-style verifiable-evidence logging stays a third-party opt-in. AND the OpenTelemetry GenAI semantic conventions remain in Development (not Stable) status. | 72% | 2027-Q1 | ● open | — |
| Dive · 2026-07-09 | Claude Code keeps progressive disclosure as the default for skills — in a regular (non-subagent) session, only a skill's name and description are preloaded and the full SKILL.md body loads on invocation, not preloaded by default — AND the default always-loaded skill-listing budget stays a small fraction of the context window (skillListingBudgetFraction default at or below ~0.02, not a full description for every installed skill). | 80% | 2027-Q1 | ● open | — |
| Dive · 2026-07-10 | The interaction-data moat stays asserted, not demonstrated: no AI-coding vendor publicly shows a reproducible model-quality gain attributable to training on IDE accept/reject/preference data via a contamination-resistant agentic benchmark (SWE-bench Pro / SWE-rebench class) that independent evaluators reproduce — AND enterprise/Business-tier zero-data-retention stays the default, so the highest-value repositories remain contractually off-limits for training. Grok 4.5, specifically, does not publish a system card carrying such an independently reproduced coding score. | 68% | 2027-Q1 | ● open | — |
| Dive · 2026-07-11 | WebMCP (navigator.modelContext) stays an origin-trial / W3C Community-Group draft with no cross-browser-shipped, specified consent model — AND the dominant page-perception path in shipped agent harnesses stays the accessibility-tree snapshot (Playwright MCP browser_snapshot / Chrome DevTools take_snapshot), not page-declared WebMCP tools and not vision-first; the structure lives in the a11y tree before it lives in the page's own tools. | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-07-12 | Nvidia does NOT take a vendor-financing write-down (equity + capacity backstop) large enough to cut an annual EPS by >5% before end-2027 — but at least one publicly traded neocloud (CoreWeave / Nebius / a listed peer) has a credit-stress event in the same window: a covenant breach, a down-round or distressed refinancing, a downgrade deeper into junk, or a canceled / renegotiated anchor contract. The risk sits in the levered middle of the circle, not at Nvidia. | 60% | 2027-12-31 | ● open | — |
| Dive · 2026-07-14 | Anthropic does NOT ship a downloadable / offline tokenizer for its current models through Q1 2027 — getting the billed token count stays an API round-trip (count_tokens), with no local library that reproduces it — AND the newer-tokenizer ~30% inflation (Opus 4.7+ / Fable 5 / Mythos 5 / Sonnet 5 vs earlier models) is not reversed or materially reduced on a shipped model. So per-file token counts stay model-and-version-specific and cross-vendor code ratios (Claude vs GPT) stay ≥ ~1.4×. | 78% | 2027-Q1 | ● open | — |
| Dive · 2026-07-15 | On-device system speech-to-text (Apple SpeechAnalyzer and peers) does NOT close the hard-audio gap through Q1 2027: on a real-world far-field / multi-speaker / accented benchmark (earnings22-class), the on-device model stays behind a small hosted or cloud Whisper-class model — so cloud STT keeps a genuine specialist tier (hard audio + rare languages), even as it plainly loses the clean-English, near-field, single-speaker default to $0 on-device. | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-07-16 | Claude Code (the standard CLI) does NOT make MCP tool-definition deferral the default through Q1 2027 — a freshly connected MCP server still injects its full tool schemas into every request by default, and pruning stays a manual opt-in (/doctor, defer_loading, or disconnecting the server); no on-by-default Tool-Search / deferred-loading path ships in the CLI that hides an unused server's schemas without the user configuring it. | 65% | 2027-Q1 | ● open | — |
| Dive · 2026-07-17 | Claude Code (the standard CLI) stays closed-source through Q1 2027 — Anthropic does NOT open-source the core agent/CLI, and answers the transparency competition (against open challengers OpenCode and Grok Build) with published data-flow docs plus telemetry opt-outs rather than a source release; the frontier vendors keep the harness closed even as challengers go open. | 72% | 2027-Q1 | ● open | — |
| Dive · 2026-07-18 | Through Q1 2027, on a hard contamination-resistant reasoning/agentic benchmark, no frontier lab demonstrates that reaching its TOP accuracy tier costs materially fewer tokens-per-solved-task than the prior generation — per-token list prices keep falling but peak-accuracy cost-per-solved-task stays flat-to-rising, because closing the last points keeps requiring super-linear test-time compute (long thinking plus best-of-N); the reasoning/sampling tax is passed to the bill, not absorbed by training. | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-07-21 | OpenAI's flagship coding model (the GPT-5.6 line or its successor) keeps a long-context price surcharge through Q1 2027 — a fixed input-token threshold above which the request reprices at a higher input/output multiplier — rather than moving to flat pricing across its full advertised context window the way Anthropic did (Mar 2026, >200k surcharge removed). The long-context cliff does not become obsolete; metered long context stays OpenAI's default. | 72% | 2027-Q1 | ● open | — |
| Dive · 2026-07-22 | Through Q1 2027, no published controlled study demonstrates that experienced engineers' defect catch-rate in AI-assisted code REVIEW (not authoring) holds steady or improves as agent reliability rises — the automation-complacency result (reviewer miss-rate rising with automation reliability, per Parasuraman/Manzey) either reproduces in the code-review setting or the human catch-rate stays unmeasured by vendors and benchmarks. "Human in the loop" keeps being asserted as a control without a reviewer miss-rate number. | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-07-24 | Through Q1 2027, frontier AI's headline math and coding wins stay concentrated on the verifier-backed side — no frontier system is credited by domain experts with an ORIGINAL, human-novel proof of a universal (∀) statement that lacks a cheap external checker (a genuinely new conceptual result, not a Lean formalization, a counterexample/witness, or a verifier-search construction à la FunSearch/AlphaEvolve) at anything like the cadence of its counterexample and construction results; capability keeps tracking the availability of a cheap, faithful, runnable verifier. | 72% | 2027-Q1 | ● open | — |
| Dive · 2026-07-25 | Through end-2027 the five hyperscalers' AI data-center leverage keeps migrating OFF their balance sheets rather than back on — no voluntary consolidation of the major data-center SPVs/JVs onto a hyperscaler's balance sheet, and no FASB/SEC rule forces debt-equivalent classification of the un-commenced-lease + unconsolidated-JV obligations; the aggregate off-balance-sheet data-center commitment of the five stays at or above their combined reported adjusted debt (Moody's ~113% baseline, Feb 2026) and does not fall materially below it. The leverage stays a footnote, not a liability. | 70% | 2027-12-31 | ● open | — |
| Dive · 2026-07-26 | Through Q1 2027, Anthropic's "less scaffolding" context-engineering guidance holds and doesn't quietly re-bloat: Claude Code's default INSTRUCTION system prompt — the hand-written rules, tracked by Piebald's public diffs, excluding tool schemas — stays at or below its post-Claude-5 cut size, and Anthropic does not reintroduce a comparable volume of prescriptive style/behavior rules for its flagship models; "fewer rules, higher-signal context" stays the shipped default, not a launch-week talking point. | 65% | 2027-Q1 | ● open | — |
| Dive · 2026-07-28 | Through Q3 2027, the publicly documented large agent-fleet rewrites and ports run overwhelmingly INTO top-corpus languages (Rust / Go / TypeScript / Python / C++), and no low-resource language (Zig / Nim / Crystal / Odin / V) is the TARGET of a comparable fleet rewrite — one that ships at Bun-like scale (~100k+ lines), parallelism, and cost/quality. Training-corpus size acts as a real language-selection pressure, and the arrow keeps pointing toward the high-resource languages the model already reads fluently. | 70% | 2027-09-30 | ● open | — |
| Dive · 2026-07-30 | Through Q1 2027, Claude Code keeps subagent text-forwarding OPT-IN — the default output stream stays tool_use / tool_result only, and --forward-subagent-text (or CLAUDE_CODE_FORWARD_SUBAGENT_TEXT) remains the switch — and the default single-session concurrent-subagent limit stays around 20 with default spawn depth around 3 (no material increase). Anthropic keeps steering SUSTAINED parallelism to agent teams and background sessions, each with its own context, rather than scaling up the single-context subagent: the subagent stays positioned as a context-isolation primitive, not a scale-out compute one. If text-forwarding flips to default, or the concurrency ceiling jumps, the reframe inverts. | 65% | 2027-Q1 | ● open | — |
| Dive · 2026-07-31 | Through Q1 2027, no frontier AI system is credited by cryptographers with a cryptanalytic result that both (a) targets a full-round, standardized, deployed primitive (AES / ChaCha20 / SHA-2 / SHA-3 / Ed25519 / X25519 / ML-KEM / ML-DSA) and (b) is a genuinely novel technique whose correctness is established by formal or machine-checkable analysis rather than by execution. AI-found cryptanalysis stays concentrated on the executable-and-checkable side — key recovery on reduced-round or not-yet-standardized schemes, where a recovered key is a one-pass witness — while full-round deployed primitives take no such hit. If a model produces a novel structural attack, verified formally rather than by running it, against a primitive people actually deploy, the labor-not-security reading flips. | 80% | 2027-Q1 | ● open | — |
| Dive · 2026-08-01 | Through Q1 2027, the 2026-07-28 stateless MCP transport holds as the forward default: the SDKs keep session-based Streamable HTTP (server-side Mcp-Session-Id state) as legacy-only, with no reversion to server-side sessions as the recommended remote model; stdio stays unchanged (no session or auth layer added); and server-initiated requests (elicitation / sampling) stay in-band via Multi Round-Trip Requests (input_required) rather than reverting to held-open SSE server-to-client streams as the default. The protocol keeps moving toward the gateway-fronted, per-request-auth deployment target — header-based routing and CIMD-over-DCR are not withdrawn. If the session model returns as the recommended default, or MRTR is rolled back to a push transport, the reframe inverts. | 72% | 2027-Q1 | ● open | — |
| Dive · 2026-08-02 | Through Q1 2027, no published detector-survival benchmark — machine- detectability measured at platform ingest on content that went through a real social pipeline (upload / re-encode / screenshot for media; a paraphrase or re-type for text), not a lab transform — shows AI-content marking holding above ~90% true-positive for images, or any meaningful figure (>~50% TPR at low FPR) for re-typed / paraphrased text. The metadata / C2PA path stays near-zero survival on the major platforms, and the EU's Article 50(2) enforcement and Code of Practice keep "as far as technically feasible" as the compliance standard rather than mandating a fixed survival threshold — so 50(2) stays a duty to attach at origin, not a guarantee to detect at consumption. | 78% | 2027-Q1 | ● open | — |
| Dive · 2026-08-04 | Through Q1 2027, no weight-streaming / offloading loader (AirLLM-class) makes a dense 70B-or-larger model run at interactive speed (>=5 tokens/sec) on a consumer GPU of 8GB or less. Batch-1 decoding reads the full set of weights once per token, so single-stream throughput stays bounded by (slowest-link bandwidth / model bytes) — streaming from disk or system RAM just picks the slowest bus. Usable local speedups on that hardware keep coming from smaller models or from MoE sparsity (streaming only the ~5% active experts per token), not from streaming dense weights. If a dense 70B streams at interactive speed on a 4-8GB card, the bandwidth-wall framing is wrong. | 85% | 2027-Q1 | ● open | — |
| Dive · 2026-08-05 | Before 2027-01-01, at least two of {OpenAI, Anthropic, Google} lead a flagship agentic-coding launch with a harder successor evaluation (SWE-bench Pro, or a contamination-resistant / unsaturable benchmark) as the headline coding number, instead of SWE-bench Verified — because Verified no longer resolves the frontier: the top models sit inside the score's own confidence interval (~±1.9 pts at N=500, p≈0.95) and inside the benchmark's documented label-error rate. The retirement of Verified-as-headline is the visible sign of the thesis that benchmark saturation and model commoditization are the same event. If both labs keep leading with SWE-bench Verified as the headline coding number through year-end, the framing was early. | 68% | 2026-12-31 | ● open | — |
| Dive · 2026-08-06 | Through Q1 2027, Claude Code does NOT ship a default-on mechanism that compresses or caps verbose tool-result tokens in the context the model reads (a built-in or config that trims Bash/Read/tool output by default, beyond the existing raw character truncation). Recovering context budget on the output side stays a manual opt-in — a PostToolUse `updatedToolOutput` hook or subagent isolation — and tool results stay the largest uncontrolled consumer of the usable window unless the user configures one. If Claude Code ships default-on tool-output compaction, this was wrong. | 65% | 2027-Q1 | ● open | — |
| Dive · 2026-08-07 | Through Q2 2027, agent-config standardization stays confined to the advisory instructions file: no shared cross-harness standard for the enforcement layer (a portable hook/permission/sandbox spec adopted by ≥2 major agent harnesses, e.g. under the Agentic AI Foundation) ships or is formally adopted — AND Claude Code keeps reading CLAUDE.md by default (no native default AGENTS.md read; the `@AGENTS.md` import or symlink stays the only bridge). Distinct from the 08-03 call (compiling policy into enforcement inside one harness); this is about the portability/standardization of that enforcement layer across harnesses. If a portable enforcement spec is adopted, or Claude Code reads AGENTS.md natively by default, this was wrong. | 80% | 2027-Q2 | ● open | — |
| Dive · 2026-08-08 | Through end-2027, no model-in-silicon product (weights baked into mask-ROM — Taalas/AMD or a peer) serves a current frontier-tier model in production: none carries a top-~5 intelligence-index model while that model is still current (within ~one release cycle) from etched weights. Weights-in-silicon deployments stay confined to small/stable/commodity models the operator has deliberately frozen, because frontier release cadence stays faster than the ~2-month metal re-spin — so the shelf-life constraint holds and etching stays the commodity tier's endgame, not the frontier's. If an etched current frontier model carries production traffic, this was wrong. | 75% | 2027-12-31 | ● open | — |
| Dive · 2026-08-09 | Through Q1 2027, no published controlled field study of production/brownfield, correctness-critical software — stratified by developer skill and run on real repositories, not customer-support scripts, graded consulting decks, or greenfield toy tasks — shows AI narrowing the junior–senior defect-rate or output-quality gap as models get more reliable. The "AI levels knowledge work" result (Brynjolfsson-Li-Raymond +34% for novices; Dell'Acqua-BCG +43% for below-average performers) stays confined to cheap-verifier, bounded-downside tasks, while the expensive-verifier case keeps reproducing expert-slowdown / skill-concentration (METR-class) — confirming that AI's distributional effect is set by the task's verifier cost, not by AI. If a production RCT shows AI closing the skill gap where the verifier is expensive, this was wrong. | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-08-12 | Through Q1 2027, the major closed frontier providers (Anthropic / OpenAI / Google) do NOT move reasoning-trace handling fully server-side: the flagship reasoning APIs keep returning the encrypted/omitted chain-of-thought to the client (Anthropic's signature, OpenAI's encrypted_content) as replayed round-trip state, rather than a never-returned server-held trace with per-session-bound, non-portable keys — AND independent researchers demonstrate at least one further cross-session / cross-model reasoning-trace recovery on a shipped flagship in that window. The trace stays client-held and recoverable in the tail because the stateless round-trip is load-bearing. If the providers move the trace server-side (never returned, keys session-bound and non-portable) and the recovery attacks stop reproducing, this was wrong. | 70% | 2027-Q1 | ● open | — |
| Dive · 2026-08-13 | Through Q1 2027, Claude Code's durable way to run many Claudes at once stays the separate-session path — independent, restart-surviving, addressable sessions coordinated by a human through cross-session messaging (SendMessage / ListAgents), the `claude agents` control plane, background sessions, and the Notification hook (agent_needs_input / agent_completed) as the attention router — while agent teams stay experimental and opt-in (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS, off by default) and do NOT become a resumable, default-on, cross-session-shareable production feature. The subagent stays a context-isolation primitive (per the 2026-07-30 dive), not the concurrency unit. If agent teams ship as a default-on, resumable production concurrency model, or the subagent becomes the scale-out unit, this was wrong. | 68% | 2027-Q1 | ● open | — |
| Dive · 2026-08-14 | Within four quarters, at least one more widely-deployed, heavily-tested open-source infrastructure project (a database, a queue, or a consensus / replication library) publicly attributes a long-latent concurrency or durability bug — one that survived years under high line/branch coverage — to a deterministic-simulation / scheduler-exploration testing pass, rather than to a coverage-driven test or a production incident. The technique that caught SQLite's 16-year WAL-reset race in about 15 minutes catches a category, not a single bug. If deterministic simulation stays a FoundationDB-lineage niche and the next wave of these bugs keeps surfacing in production by luck, this was early / wrong. | 70% | 2027-08-14 | ● open | — |
| Dive · 2026-08-15 | Through Q2 2027, no frontier lab (Anthropic, OpenAI, or Google) ships — as a headline capability eval on a flagship model card — a multi-turn clarification / underspecified-task benchmark that scores the model on asking-versus-guessing under ambiguity (calibration plus clarifying-question behavior on genuinely under-specified inputs). Public benchmarking of coding and agentic models stays single-shot, pass/fail, on fixed-answer tasks, so the collaborative "ask before you assume" behavior stays un-measured — and therefore un-optimized-for by the leaderboard — remaining a per-user prompt and eval discipline rather than a scored, advertised model property. If a lab makes an under-specification / clarification benchmark a headline number on a flagship release in this window, this was wrong. | 72% | 2027-Q2 | ● open | — |
| Dive · 2026-08-16 | Mistral's US 12,670,045 ("Code implemented tool calls") is not successfully monetized or enforced against a major agent framework through Q4 2027 — no publicly reported paid license, settlement, or injunction. The general executable-tool-calling pattern's prior art (CodeAct, ICML 2024; shipped Cloudflare Code Mode and Anthropic's code-execution-with-MCP, both 2025) makes the broad reading unenforceable, and the narrow pause-ship-resume-replay execution mechanism the claim actually fences is trivially designed around (coroutine suspension instead of replay). If Mistral extracts a reported license fee or wins an injunction/settlement against a major agent framework in this window, this was wrong. | 72% | 2027-12-31 | ● open | — |
| Dive · 2026-08-18 | Through end-2027 the training corpus stays the one asset every frontier lab keeps sealed BECAUSE disclosure is discovery: no frontier lab (OpenAI/Anthropic/Google/Meta/xAI) publishes a work-level source manifest for a current top-tier model (and none ships one meeting OSAID's "data information"), AND at least one additional frontier lab beyond Anthropic records a copyright settlement or adverse judgment of ≥$100M tied to training-data *acquisition* (pirated or unlicensed copies, distinct from an output/regurgitation claim). If a lab discloses a full corpus manifest without drawing suits, or no second ≥$100M acquisition-liability event lands, this was wrong. | 65% | 2027-12-31 | ● open | — |
Source of truth:
_data/predictions.yml. Working memory:
editorial memory.