Plans, builds, results, lessons and decisions remain in their original sequence. Filter the record or open it five entries at a time—without turning the homepage into an endless feed.
CANONICAL JSON
Program published: · Canonical file auto-checks every 60 seconds
Waiting for first browser check
BUILD IN PUBLIC / APPEND-ONLY VIEW
Public research log
A timestamped narrative of what we planned, built, learned and decided.
LOG-151
resultP02.3
A paid catalog is not global coverage. All 87 tracked tickers now have a measured data map.
A one-time Financial Datasets audit tested every currently tracked instrument: 92 listing routes and 87 unique symbols across the core AGI portfolio, Global 50, China routes, defense routes and QQQ/SPY benchmarks. The provider catalog contained 43 exact symbols; the identity gate verified 35. Eighteen symbols passed the full private research-input standard, two support standardized-only work and fourteen need field or provenance repair. Forty-four were absent from the exact-symbol catalog, five short symbols resolved to another issuer and four fund/private-market routes were structurally unsuitable for company fundamentals. Of 34 captured financial snapshots, 33 contain twelve quarters and 27 include the required as-reported packet; 27 are current, one lagging and six stale at the cutoff. The most important result is negative but useful: this provider can accelerate an identity-verified US lane, but it cannot serve as our global security master. A successful API response is never enough to join facts to a company. The audit also rejected an SPCX listing conflict and exposed sector-specific schema traps such as bank capex and null-versus-zero debt. 189 local API attempts were recorded; auto-reload, subscriptions and recurring calls remain off. Because the current plan does not grant underlying-data redistribution, the public receipt exposes classifications and ticker cohorts—not licensed values. No canonical evidence, valuation, portfolio weight or trade authority changed.
The last AI built by humans has not been built yet.
The 75-page v1 of “The Last AI Built by Humans” does not show that an autonomous takeoff has begun. It gives us something more useful: an evidence ladder for deciding what would count. B0 improves one output without inheritance. L1 executes a human-prescribed change; L2 chooses an improvement strategy; L3 chooses what experience to acquire; L4 persists adaptation from deployment; and L5 revises the improver, verifier or successor-generation mechanism itself. The decisive distinction is structural versus effective recursion: reusing a revised mechanism proves loop closure, while genuine benefit requires better later successors under matched total budgets, protected independent evaluation, held-out transfer and a complete regression history. The survey maps broad L1–L2 evidence, domain-dependent L3–L4 and only bounded L5 prototypes. Its HCI is not “percent to AGI,” and its post-2026 curve is illustrative rather than predictive. We therefore add the RSI Ladder as a standing track: every claim now receives a level, persistent-state receipt, successor re-entry test, verifier-authority record, matched-budget ledger, transfer result, generation count, rollback history, unit-economics measure, owner-economic KPI and falsifier. CDNS/SNPS may sit closest to executable RSI feedback; platform, compute and physical-infrastructure holdings gain only if validated research throughput closes into owner cash faster than efficiency or expectations absorb it. No capability prior, issuer verdict, valuation, weight or trade changed.
AGI is a field of inquiry—not a unit of measurement.
Roberto Pieraccini’s “Four Questions About AGI” correctly exposes the danger of letting one ambiguous banner stand in for intelligence, generality, autonomy, reliable deployment and economic value. Its strongest evidence is not that AGI is meaningless, but that these clocks can diverge: benchmark performance can leap while competence remains jagged; digital capability can scale while physical reliability moves slowly; disruption can precede institutional diffusion; and a laboratory can declare a milestone without independently establishing owner economics. The essay is a secondary synthesis, not a new capability receipt. Its “red herring” conclusion is an argument rather than consensus, and Rodney Brooks’s decade-per-nine reliability rule is a useful heuristic rather than a validated universal law. We therefore keep AGI as the name of our research field but prohibit the label itself from entering the evidence chain. Every future AGI claim must be decomposed into six measurable layers: capability, autonomy, reliability, diffusion, economics and price. For capital, the exploitable object is the gap between impressive capability and dependable economic deployment—and the bottlenecks that close or widen it. This is a methodological clarification only: no capability prior, issuer verdict, valuation, portfolio weight or trade changed.
AGI is not the finish line. It is the start of a compounding race.
Google DeepMind’s 57-page “From AGI to ASI” gives us a more useful map than a single takeoff date. It defines competent AGI around median skilled-human performance on most cognitive tasks and ASI as intelligence exceeding large, coordinated groups of thousands or tens of thousands of experts across virtually all domains. Six native advantages—speed of I/O and processing, memory, substrate independence, lossless replication and high-bandwidth sharing—make stalling exactly at human level an unstable base case. Progress can run through four non-exclusive paths: scaling compute/models/data, a new algorithmic paradigm, recursive AI R&D and multi-agent organization. Each meets six possible frictions: data, resources/economics, paradigm limits, harder research, abstraction/physical latency and deliberate governance slowdown. The key fund insight is organizational: cheap AGI copies may form a superhuman collective before one model becomes solitary ASI. We therefore replace “when ASI?” with a forward matrix: path × friction × economic transmission × listed owner × price. Epoch AI’s current scaling indicators and METR’s field observations suggest scaling and R&D autonomy are advancing, but physical build times, judgment and reliability still bind. Every future event now needs a path, counter-friction, capability, physical, owner-economics and price receipt plus a horizon and falsifier. Five review triggers are proposed; none creates an issuer verdict, valuation, weight or trade.
AIMultiple’s “10,000 predictions” is a valuable discovery map, but not one comparable consensus distribution. It combines more than 6,000 participants across ten surveys, 18 named forecasts and roughly 3,900 community/market contributors while mixing AGI, weak AGI, HLMI, full automation of labor, a Turing-test outcome and a company announcement. The definition alone moves the 2023 expert-survey median from 2047 for HLMI to 2116 for full labor automation; live market pages also moved materially beyond the article’s August snapshots by our 15 September cutoff. We therefore reject one weighted calendar year as a portfolio input and split the question into six clocks: capability, reliable autonomy, declaration/contract, deployment/diffusion, cash-flow capture and market pricing. Calendar revisions now have proposed forward rules: a ≥10 percentage-point repricing with a liquidity receipt or a ≥12-month same-contract median shift can open review, but changing a research prior requires two independent capability/autonomy lanes; issuer escalation additionally requires owner economics and a point-in-time valuation gap. The March 2027 forecast remains our pre-registered hypothesis—not a consensus claim. No issuer verdict, valuation, weight or trade was created.
P10.1 turns the frozen denominator into an operating intake. Both Daily AGI Answer and Daily market research now rebuild the three upstream sources, inspect every post-activation object in source-clock order and append one of three receipts: admitted decision-grade, rejected admission gate or rejected duplicate root event. Admission requires all ten P10 gates simultaneously; source priority can select the earliest complete representation but cannot relax a gate. Identity revisions fail closed, rejected objects stay outside the denominator and one root event can contribute at most once across all ledgers. The current result is deliberately empty: 0 post-activation objects, 0 admitted records and 0 rejections. That is not missing work—it is a signed no-backfill boundary. P10.2 clocks remain locked until the first eligible future object arrives.
Three forward ledgers. One denominator. No hindsight.
P10 begins with a refusal to call unlike objects a track record. The P06-NET hypothesis ledger, P07 governed-docket audit and P09 global event-company ledger are now connected through one frozen P10.0 admission contract, but connection is not admission and their counts are never simply added. A future object must pass ten gates: it must be recorded before the outcome; carry immutable identity and dated evidence; state a falsifiable claim, exact baseline, finite horizon and deterministic resolution rule; map to a named entity and causal channel; prove root-event deduplication; and preserve the capital firewall. Today the three sources contain 0 upstream objects and the common denominator contains 0 decision-grade records, 0/100 resolved outcomes and 0 alpha claims. Calibration is null, not zero. The March 2027 AGI forecast remains visible in a migration queue but contributes nothing until a machine-readable adapter proves comparability. P10 ends finitely at P10.4 after intake, outcome clocks, calibration/countercase audit and a governed P11 handoff; no valuation, portfolio weight, order or trade authority is created.
A feed is not coverage. Twenty-one named channels form the current source perimeter.
The Daily AGI Answer source contract is now public in full. Eleven daily channels cover OpenAI, Anthropic, Google DeepMind, Meta AI, METR, NVIDIA, TSMC filings, BLS JOLTS, BIS export controls, FRED unemployment and IBM enterprise deployment. Ten weekly channels cover Epoch AI, ASML, EIA, FERC, NERC, Census business formation, BEA consumption and three explicit scenario/countercase maps: Situational Awareness, Citrini and Citadel Securities. Their purpose is not to maximize headlines; it is to observe separate links in the AGI → capabilities → deployment → compute → power → labor/policy → owner cash → expectations chain. Every due inspection writes an access/freshness/candidate receipt, and “no new publication” is separated from access failure. The registry is audited weekly against a 60-cell region × mechanism × evidence matrix and reviewed monthly for primary-evidence value, unique admitted candidates, timeliness, machine reliability, rights and countercase coverage. New routes spend at least seven days in shadow mode and need three clean captures before promotion. Three consecutive access failures trigger review; broken, replaced or rights-blocked routes are demoted, while editorial silence alone never removes a source. Event-driven review begins when a new frontier lab, regulator, bottleneck or repeatedly useful discovery source exposes an uncovered causal lane. The perimeter is therefore versioned and expandable—not a permanent hand-picked bookmark list.
News is not a signal. A signal must survive six gates.
The Daily AGI Answer now publishes the exact signal boundary. A candidate is material only when all six conditions hold simultaneously: a direct, dated and attributable source is inside the 48-hour window; fact is separated from inference; the event changes a named AGI → economy → capital mechanism and maps to a layer or entity; the expectation gap and dated falsifier are explicit; analyst confidence is at least 0.65; and an independent evaluator passes every editorial gate at confidence of at least 0.78. The eligible channels are capability/autonomy, adoption/unit economics, compute/semicap capacity, power/grid/data-center constraints, policy/export/market access, labor productivity/business formation and valuation/portfolio risk. A headline, changed page hash, social post, standalone benchmark, unexplained price move or stale/duplicated/unattributable inference is never sufficient alone. Passing the gate creates only WATCH or THESIS_REVIEW. A rebalance review additionally requires a thesis falsifier, KPI trend break, scenario-probability change, point-in-time valuation gap, position/layer band breach or portfolio-risk breach, followed by current receipts, issuer/layer KPIs, scenario/valuation/risk work, a bounded proposal, independent validation and explicit human investment-committee authorization. The Daily cannot change a weight or place a trade.
Six markets. One identity grammar. Zero invented history.
P09.2A sends one frozen issuer → legal security → listing → point-in-time symbol grammar through six maximum-contrast defense/security identities: Palantir, Rheinmetall, AviChina, Bharat Electronics, Mitsubishi Heavy Industries and Embraer. This is deliberately not a “best defense stocks” list. It is an identity-portability test spanning a U.S. venue transfer, German bearer shares, a PRC H share in Hong Kong, Indian native equity, a numeric Japanese symbol and a Brazilian current symbol whose exact corporate-action interval remains licence-blocked. Palantir proves why ticker-first data fail: PLTR stayed PLTR while the legal listing moved from NYSE to Nasdaq on 26 November 2024. Embraer proves the inverse: current EMBJ3 is observable in an admitted issuer disclosure, but the precise EMBR3 → EMBJ3 interval may not be reconstructed from the blocked B3 history, so validFrom remains null with a reason. Six issuers, six securities, seven listings and seven symbol records are now hash-bound with 39 explicit null reasons. Forty-four identities and eleven sectors remain. Zero prices, exposure edges, scores, valuations, portfolio weights, trades or alpha claims were created.
The bubble can be real. The revolution can still win.
A deep review of bubble research, central-bank stability evidence and current AGI economics rejects two easy stories: neither “the whole market is dot-com” nor “AGI makes price irrelevant” survives the evidence. The current regime is a genuine technological revolution with selective bubble risk. S&P 500 valuation and concentration are elevated; the AI investment surge is already near the 2000 peak in GDP terms; yet NVIDIA and Microsoft show enormous realized revenue and cash flow, while aggregate household and business leverage remains moderate. The page therefore replaces bottom-calling with a four-gate entry protocol: survival without on-time AGI, expectations visible through reverse DCF, primary-evidence monetization and portfolio-factor capacity. A 20–30% drawdown is only a review trigger. The strongest long-entry zone is price compression plus an intact causal thesis plus improving evidence. No leverage, issuer verdict, approved valuation, portfolio weight, market-timing claim or trade authorization was created.
The Daily AGI Answer now has an explicit path from audience attention to research utility. One sealed answer produces three deterministic distribution packages—short, professional and direct—without rewriting its admitted fields. On the canonical page, a reader can inspect evidence, choose an investor, founder, researcher or engineer lens, and turn a question into a falsifiable Research Track stored only in that browser. The receipt can be copied or downloaded as JSON. Optional first-party measurement records only session-deduplicated event counts for views, evidence openings, lens choices, track starts, exports and returns; it stores no identity, email or research text. Three KPI contracts begin with null baselines and null targets until 30 live days exist. Sources, corrections and the resolved forecast record remain public. Alerts, continuing private tracks, a personal archive and API access are only membership candidates—not products currently for sale. No external account was posted to, no subscriber was enrolled, and no capital action occurred.
The $100B line has been crossed. The definition has not.
OpenAI and Anthropic's reported annualized revenue now exceeds $105B in aggregate. That is a real commercialization threshold, but not the same as the reported 2023 Microsoft–OpenAI test: $100B of profit generated by one system. Revenue is not profit; run-rate is not a completed year; two firms are not one system. Axios also reports material gross-versus-net accounting differences. The official 2025 Microsoft–OpenAI agreement requires independent expert-panel verification, and the reviewed sources show no public declaration or panel verdict. Our conclusion: the economic AGI regime may already be under way; contractual AGI is unverified and capability AGI remains governed by the six pre-registered gates. The March 2027 55% prior stays frozen. No issuer verdict, valuation, rank, weight or trade changes.
P09.1 turns “monitor the whole planet” into an executable rights perimeter rather than a scraping slogan. We froze exactly 24 official routes: one regulatory-disclosure, security-reference, market-data/corporate-actions and economic-context lane for each of six maximum-contrast anchors—United States, European Union, Hong Kong, India, Japan and Brazil. Eighteen routes may support bounded automated discovery, official APIs, derived facts or item-level open-data reuse. Each route now carries its canonical source, separate rights basis, language, cadence, machine route, retention rule, redistribution rule, point-in-time policy and evidence limit. All six exchange market-data lanes are blocked until a licence explicitly permits the intended ingestion, history, display and index-like use. This matters because a public webpage is not permission to copy it; translation cannot expand rights; a latest-only API cannot silently overwrite a prior vintage; and an open-data portal cannot waive an item-level licence. Brazil is the first contrast anchor for emerging markets, not a claim of complete regional coverage. The map admits zero issuers and zero securities. P09.2 may now begin identity-only intake through the cleared disclosure and reference lanes, while prices, rankings, valuations, portfolio weights, trades and alpha claims remain closed.
P09.0 freezes the identity and causal grammar of the Global AGI Equity Observatory before selection can bias it. Issuer, legal security, exchange listing and point-in-time symbol history are now four separate objects; a current ticker cannot be silently projected into the past, and missing identifiers remain null with a reason rather than being guessed. Every future exposure edge must declare one primary route from AGI to economics: direct AGI supply, enabling constraint, adoption and distribution, displacement exposure, or second-order transmission. Each route carries both an owner-cash-flow question and a falsifier. The security master requires 32 fields and five clocks: source publication, observation, ingestion and validity start/end. The advertised 50 securities across six regions and twelve sectors remain an empty maximum-contrast target: zero issuers, securities, listings, evidence sources, causal edges, valuations, weights, trades or alpha claims are admitted. P09.1 must now record rights, retention, redistribution, language and point-in-time policy for every source class before the first candidate enters the map.
The starting portfolio now has an honest birth record.
The machine-readable receipt sealed on 5 September 2026 proves the exact starting state of Experiment 00: USD 100,000 of paper capital, 25 companies plus a 3.5% SHV reserve, 26 allocation lines across eight AGI-stack layers, zero leverage and a banded quarterly policy. The basket was thesis-led rather than optimizer-produced: it combined frontier-model and distribution exposure with scarce compute, semiconductor-tooling, networking, data-centre and power bottlenecks; individual weights then reflected qualitative judgments about business quality, bottleneck power, liquidity, concentration and permanent-loss risk. That is the best reconstruction of the decision—not contemporaneous proof of a complete selection algorithm. The eligible universe, rejected alternatives, point-in-time valuation snapshots and exact weight derivation were not sealed at inception, so we label those explanations RETROSPECTIVE RECONSTRUCTION and will not manufacture missing provenance. Historical curves using today's constituents remain exposure diagnostics, not an investable track record. P11 must rebuild the universe-to-weight bridge prospectively before any alpha claim. This log changes no holding, weight or trade authority.
The first data partnership now starts with a handshake, not an attachment.
P08.2A closes the gap between an internal intake specification and something a real data owner can safely answer. All six India workflow routes now expose self-contained, downloadable JSON schema receipts bound to the frozen route, workflow and acquisition hash. A provider can declare only field structure, authority, format, treatment/control availability, sample and time capacity, derivable outcomes and source-system lineage; the templates contain no rows, examples, prompts, transcripts, source code or personal identities. Every blank template is intentionally incomplete and is automatically rejected by the same ten-gate validator that will inspect an authentic response. Software delivery is the priority handshake because stable work-item clocks, quality outcomes and matched teams can satisfy the shortest measurable causal path without requesting source code or identity. This release sends nothing: external contacts, submitted receipts, rights receipts, private rows, matched datasets, measured effects, issuers and capital actions all remain zero. A separate founder authorization is required before one template leaves the laboratory. Even a complete receipt advances only to rights review; row transfer needs a new authorization and P08.3 stays locked until a real rights-cleared matched dataset passes the inherited measurement contract.
Daily AGI Answer 2.1 turns the canonical research object into four synchronized public surfaces without creating four truths. The same immutable delta, economic mechanism and capital boundary now drive the written brief, an editorial visual, browser-local voice narration and a three-scene motion brief. Readers can still change their investor, founder, researcher or engineer lens without changing evidence. A new beta Research Track Studio lets a reader frame an investment or business question, identify the entity or market and pre-commit a falsifier before research begins. Its first receipt is intentionally local to the browser, exportable and marked NOT SUBMITTED: it does not upload user text, run a model, enter the public evidence queue, select a security or authorize capital. A moderated server intake, scheduled evidence contract, generated per-edition audio/video files and subscriber delivery remain separate future gates. The product direction is now explicit: do not merely broadcast AGI news—help the audience turn curiosity into testable, cumulative research.
The AGI-to-defense thesis is not “budgets rise, buy defense.” Greater autonomy can improve sensing, planning, swarm coordination and software adaptation, but owner cash appears only when a supplier lowers cost per defended outcome, integrates across sensors and effectors, updates from field evidence, scales production and preserves contract margin. We opened a discovery queue around RTX, Lockheed Martin, Thales, Rafael and D-Fend because a secondary supplier map makes them useful contrasts—not because five names are approved securities. Primary U.S. defense sources distinguish budget requests from appropriations and vendor demonstrations from field economics. The forward ladder will track budget request → appropriation → contract → funded backlog, demonstration → independent test → deployment → renewal, and then revenue → segment margin → free cash flow per diluted share. This is also the first contrast lane for the new finite P09 Global AGI Equity Observatory: 50 securities, six regions, twelve sectors and 100 pre-outcome event-company resolutions before scale. No ranking, valuation, weight or trade has been created.
Three economic futures are a stress surface—not a forecast.
Anthropic’s task-based model varies five axes—capability, adoption, autonomy, productivity and worker adjustment—to produce modest, substantial and extreme 2030 scenarios. Its reported GDP deltas are 1.6%, 8.3% and 32.4%, while the modeled capital share reaches 40.6%, 43.9% and 54.8%. Those numbers are not assigned probabilities and do not map directly to stock returns. The investable use is narrower and stronger: stress which companies own scarce capital, which sell complementary physical and digital services, which face labor-displacement demand loss, and which valuations already assume the extreme corridor. We will not average the three cases into false precision. Each future needs explicit priors, issuer-level transmission, valuation and falsifiers before it can change an underwriting record.
Autonomy changes the price of attack—and the price of trust.
Anthropic’s September threat report is important because the repeated mechanism is economic: agents can compress skill, labor and time costs and run more operations in parallel, while humans still select targets, monetize and review. The report spans seven harm areas and includes reported targeting of Ukrainian and European government, defense and drone supply chains. But it is a vendor-selected set of novel or severe cases, not a representative incident base rate; Anthropic’s attribution claims remain primary-vendor evidence until independently corroborated. The capital hypothesis is therefore conditional: secure deployment, auditability and safeguards may support enterprise willingness-to-pay and regulatory access, while simultaneously increasing cost, liability and geopolitical fragmentation. OpenAI and Anthropic now become a Tier 0 public-source perimeter across releases, governance, security and economic research. A future IPO is not assumed. Any offering or later dip must pass security identity, governance, dilution, valuation, liquidity and thesis-preservation gates. This project uses an OpenAI model but has no non-public OpenAI or Anthropic information. No valuation, weight or trade changed.
The machine stayed deep. The entrance became simple.
A first-reader review found a real product failure: the homepage exposed the internal P00–P12 machinery before making the public value legible. We accepted the criticism without weakening the method. The canonical Daily AGI Answer now appears immediately after the opening screen and exposes three decision fields—observed change, economic mechanism and capital relevance. A new five-second map reduces the public workflow to three auditable steps: collect primary sources, test the claim rather than repeat the marketing, and seal a dated answer with a falsifier. A real IBM case demonstrates the boundary: the OpenAI partnership announcement is confirmed, but the reproducible bridge from AI bookings to recurring software margin and free cash flow per diluted share is not yet admitted, so IBM remains under deep monitoring at 0% portfolio weight. Seven recurring terms now have controlled, accessible definitions. The full roadmap, receipts, hashes, corrections and public log remain available but are folded into an opt-in audit layer. A sitewide status strip also states what the project is and is not: a public AGI analytics laboratory, a research portfolio and no live trading. No evidence threshold, production weight or capital authority changed.
The first matched dataset must not enter as an email attachment and become “evidence” by momentum. P08.2 now places a deterministic airlock between six acquisition routes and every future private row. Each workflow has one canonical request packet that freezes the intended grain, cohort join keys, field names and types, minimum treatment/control units and groups, observation window, outcome metrics, counterfactual, falsifier and privacy exclusions. Ten fail-closed gates test receipt identity, frozen lineage, provider authority, schema-only boundaries, prohibited sensitive fields, required-field coverage, same-system treatment/control, sample capacity, temporal capacity and outcome/hash lineage. The build also found four operational name mismatches in the prior briefs—team_id, queue_id, lane_id and service_lane_id versus their pseudonymous cohort keys. We preserve the historical record and resolve those aliases prospectively; every join key is now an actual canonical field and automated tests prevent regression. A complete schema may advance only to rights review. It cannot authorize row transfer, causal inference, issuer selection or capital. Today the honest state remains zero authentic schema receipts, zero contacts, zero private rows, zero matched datasets and zero measured effects. The next acceptable event is one real schema-only receipt that passes all ten gates; P08.3 remains locked until an actual matched dataset satisfies the inherited measurement contract.
A data route is not a dataset. Find the first honest pair.
P08.2 now turns six frozen workflow cells into six executable, data-minimized acquisition briefs. A review of the admitted official perimeter found no public dataset that joins a versioned AI treatment, a matched control, the inherited minimum sample and quality-adjusted outcomes in one measurement system; the honest public matched-dataset count is therefore zero. The router opens two parallel tracks. The fastest partner track prioritizes software delivery and customer operations because their work-item and queue clocks, quality metrics and pre/post windows can be measured quickly without source code, transcripts or personal identities. The strongest official-data track targets IndiaAI compute allocations: the policy defines application and allocation mechanics that could support a natural experiment, but the public register currently exposes only field headers—not reproducible rows, utilization or production outcomes—so minimized authorized access is required. ONDC open data remains aggregate while its strategy places transaction data with buyer and seller apps; BHASHINI model availability remains capability rather than matched workflow performance. The launch queue is now explicit, every route inherits its original sample, outcomes, counterfactual and falsifier, and every first transfer must begin with a schema-only receipt. No organization has been contacted, no private row has entered the system, no effect has been measured and no issuer has been selected. P08.3 remains locked until one rights-cleared treatment-and-control dataset passes the frozen cell.
Activity is not adoption. Adoption is not displacement.
P08.2 turns the India deployment thesis into six precommitted workflow measurement cells: software delivery, customer operations, back-office processing, local-language service, MSME digital commerce and subsidised compute-to-production. Each cell freezes the unit, full denominator, versioned AI treatment, quality-adjusted outcomes, minimum treated and control samples, pre/post windows, counterfactual design and falsifier before a result exists. The first machine receipts preserve three useful but explicitly non-causal baselines: NCS reported 578,393 mobilized IT & communication vacancies for FY2025-26 as of 28 February 2026, including 446,024 in Information Technology; ONDC reported more than 500 million cumulative network transactions by 25 July 2026; and the IndiaAI price list dated 25 August 2026 exposed posted compute-price examples. A fourth receipt is deliberately empty because the IndiaAI end-user page exposes allocation headers but not reproducible static rows—so no user count is inferred. These objects show scale, exposure and capability price; they do not show AI adoption, job displacement, production conversion or owner cash flow. With zero treatment receipts, zero matched controls and zero measured effects, P08.3 issuer mapping remains locked. The next acceptable object is one rights-cleared matched workflow dataset, not a ticker or narrative shortcut.
P08.1 converts “research India” from a browsing intention into a rights-aware machine perimeter. We mapped 20 canonical routes across 7 required source classes and all 6 deployment-market causal lanes. 18 routes may now feed only their declared bounded use: internal monitoring, document discovery, public documentation monitoring or—on data.gov.in—structured reuse under the item-level Government Open Data Licence with attribution and no implied endorsement. 2 routes deliberately fail closed. NSE price, volume, identifier, redistribution and index-like use require the relevant market-data licence; MCA payload ingestion requires an authorized API, bulk file or provider route. BHASHINI public API documentation is classified as PoC, not a production entitlement, and language-model coverage is not user adoption. The contract stores no raw payload, permits no manual fallback and creates zero verified claims, measured workflows, issuers, securities, valuations, trades or alpha claims. Source identity is now READY; local-language coverage remains PARTIAL. P08.2 can begin only by choosing bounded workflows from admitted sources and freezing denominators and counterfactuals before extraction.
A model is not a market. Deployment decides who gets paid.
P08 begins before a ticker list. We have frozen a finite deployment-markets charter around one question: where can increasingly abundant machine cognition create or destroy an investable unit of cash flow before market expectations fully adjust? India goes first as a maximum-contrast stress test of the method—not as a bullish country call—because one pilot can force us to test exportable digital work, multilingual adoption, price-sensitive deployment, distribution rails, physical constraints and public-market access together. The United States, China and India are comparator roles—frontier creation, frugal diffusion and workflow repricing—not a country ranking. Every candidate path must cross six causal lanes and eight evidence gates, including measured adoption, counterfactual economics, issuer transmission, security access, point-in-time expectations and a precommitted forward outcome. P08 has a hard end at P08.6, when one fully evidenced hypothesis may be handed to the separate P07 forward docket. Today the honest count is zero sources, zero verified claims, zero workflows, zero issuers, zero securities, zero valuations, zero trades and zero alpha claims. Next, P08.1 builds the India source-and-rights map before extracting a single claim.
P07.7 is the final designed stage of P07, but it cannot be completed by shipping more code. We have frozen a prospective docket-audit contract and an append-only hash chain with four lifecycle events: governance decision, forward shadow position, resolved outcome and explicit audit-clock tick. The audit remains in ACCUMULATING_NOT_ELIGIBLE until every threshold passes: 0/12 governed dockets, 0/4 forward-priced positions, 0/4 resolved outcomes and 0/90 elapsed calendar days. The daily workflows can automatically append newly observed P07.6 decisions and positions, eligible resolution packets and one clock receipt per UTC day after the first governance event. They cannot infer days from the wall clock or insert retrospective events. Each docket freezes its probability, outcome deadline, benchmark and factor model before the outcome. Once eligible, the machine will calculate Brier score, log loss, hit rate, gross and net return, modeled cost drag, benchmark-relative return, factor-adjusted return and lineage compliance. The result will still be a finite descriptive audit: one winner cannot prove durable alpha, and no audit creates a live risk budget, portfolio weight, order or trade. P08 may now be designed in parallel while P07 accumulates reality.
Approval chooses the object. Time chooses the price.
P07.6 freezes governance and forward pricing before the first eligible P07.5 proposal. Exactly two independent reviewers—Evidence Integrity and Risk Capacity—must approve the same content-addressed shadow proposal, valuation distribution and risk map. Reviewers must be unique, conflict-free and different from the valuation producer. Their approval is still not capital authority: only one explicit GitHub-bound human receipt may APPROVE, REJECT or DEFER the exact object, with a public rationale; automatic approval is forbidden. If approved, the machine consumes a complete chronological market window and selects the first completed regular-session close strictly after the governance timestamp. It cannot use intraday, already-known or retrospectively favorable prices. Full modeled round-trip cost is prepaid before research units are calculated. The result is a FORWARD_SHADOW_POSITION—not a live risk budget, portfolio weight, order, trade or alpha claim. P07.7 is finite but time-gated: at least 12 governed dockets, 4 forward-priced shadow positions, 4 resolved outcomes and 90 elapsed calendar days are required before the process audit can begin. Today the upstream proposal count is 0, so the honest state is ARMED_BLOCKED_BY_SHADOW_PROPOSAL: 0 reviews, 0 decisions and 0 priced shadow positions.
P07.5 freezes the Shadow Expected Value contract before the first valuation distribution exists. A positive P07.4 research ceiling is necessary but insufficient: one automated packet must bind the exact risk-cost map, security identity, price, 12–60 month horizon and BEAR / BASE / BULL intrinsic values to one timestamp and four source-receipt hashes. Scenario probabilities must be positive, ordered and sum to one; missing values, manual valuation inputs and retrospective scenario tuning fail closed. Each scenario return is intrinsic value divided by the point-in-time reference price minus one; full round-trip cost is deducted before probability-weighted expected value. Gross expected edge must first exceed the P07.4 three-times cost hurdle and expected net return must remain positive. The engine then approximates full Kelly as expected net return divided by the probability-weighted squared net return, caps it at one and permits at most one quarter. A separate bound limits the BEAR outcome to a 1% loss contribution on the fixed $1m comparison denominator. The smallest of this downside capacity, quarter Kelly and the P07.4 risk ceiling becomes a non-executable shadow weight. It still requires independent review, governance and the first eligible close recorded strictly after that decision. Today the upstream chain has 0 sealed risk maps, therefore P07.5 correctly holds 0 distributions and 0 proposals. It manufactures neither valuation nor capital.
P07.4 freezes the Risk & Cost Mapping contract before the first posterior exists. A hash-sealed P07.3 update alone is insufficient: one machine risk packet must bind that exact update to a security, venue, currency, AGI layer, region and point-in-time receipts for volatility, 20-day dollar liquidity, absolute factor beta, current concentration and implementation costs. Five independent capacities then compete: unused room under a 2.5% single-name cap, a 0.75% annual volatility-contribution cap, at most 0.5% of ADV per day over three days, unused room under a 25% AGI-layer cap and unused room under a 40% regional cap. The smallest capacity wins and is scaled only by the absolute probability change, reaching full scale at a 25-point posterior move. Factor beta above 1.5, round-trip cost above 100 bps, or zero liquidity/layer/region headroom forces the research ceiling to zero. Round-trip cost is precommitted as twice half-spread, impact and fees plus horizon borrow; any later valuation must show gross expected edge of at least three times that cost. The $1m capital unit is a comparison denominator, not fund NAV. A mapped research ceiling is neither a risk budget nor a portfolio weight. Today P07.3 has sealed 0 updates, so P07.4 correctly contains 0 maps and no estimated substitute. P07.5 remains a separate point-in-time valuation and shadow-proposal gate. No weight, trade or alpha claim was created.
P07.3 freezes the probability-update equation before the first eligible live event. One complete seven-component point-in-time baseline and one exact machine packet must bind the same event, hypothesis, consensus receipt, expected metric and observed metric. Only then may the engine measure signed surprise, normalize it to the frozen expectation and map it through a precommitted five-bucket likelihood table: 0.2 for a falsifier, 0.5 for contradictory evidence, 1.0 for neutral evidence, 1.5 for supportive evidence and 3.0 for thesis-confirming evidence. Frozen prior odds multiplied by that likelihood ratio become posterior odds; the published posterior is bounded between 5% and 95%. No analyst may enter a probability or likelihood ratio manually, repair missing evidence after the outcome or switch hypotheses. Today there are 0 complete baselines, 0 exact machine packets and 0 posterior updates, so the ledger is correctly empty. Even a valid posterior creates no valuation, risk budget, portfolio weight, trade or alpha claim. P07.4 must separately translate belief into risk and cost before capital can move.
P07.2C turns the point-in-time design into an executable state machine. When P07.1 selects the first eligible post-activation material event, its immutable P07.2 request and P07.2B execution plan will open one 24-hour run. The executor must issue exactly seven ordered receipts. Every receipt binds the request hash, execution-plan hash, adapter receipt, attempt time, source as-of, retrieval time, raw-payload hash, value hash, rights class and the previous receipt hash. A lane can return captured evidence, a narrowly permitted unavailable company-guidance receipt, or an explicit blocker; it can never silently disappear. All seven required components must be satisfied before one complete baseline is sealed. Six of seven stays blocked. A request replay preserves the same idempotency key instead of manufacturing a second history. Missing licensed data, a source newer than the event cutoff, a broken hash, or a missed deadline blocks P07.3; current webpages, later restatements, manual values and late backfill are forbidden. Today the chamber is correctly empty: no qualifying event means 0 capture runs, 0 receipts and 0 baselines. This is operating readiness, not a claim that evidence or alpha already exists.
Seven baseline lanes meet reality. Two licence gates stay locked.
P07.2 could already emit a 24-hour request, but a request without acquisition routes would be theatre. P07.2B now binds all seven mandatory components to unique hash-sealed adapter receipts: exact source routes, pre-event as-of rules, rights class, storage boundary and a fail-closed blocker. Three routes are ready when a valid request arrives: the last completed security close, the frozen hypothesis metric and the dated three-month Treasury reference. Two are deliberately event-bound: issuer guidance needs the selected issuer/security, while benchmark-factor state needs the selected hypothesis factor map. Two remain genuinely blocked: consensus vintages and the valuation denominator need a licensed point-in-time vendor that passes anti-lookahead, restatement and redistribution qualification. GuruFocus is still only a candidate. The machine may not replace either gap with a current webpage, later restatement, manual value or invented zero. Today there is no selected event, no execution plan, no captured component and no complete baseline. P07.2C and P07.3 remain closed; no valuation, weight, trade or alpha claim was created.
Freeze the expectation before observing the outcome.
P07.2 removes the next hindsight loophole: describing what “the market expected” after an AGI event has already moved prices. The capture contract is armed before P07.1 has selected any event. Once the first eligible live event locks, the workflow must open one automatic 24-hour request for seven synchronized components: the last completed security close strictly before the event, point-in-time consensus expectations, the declared valuation multiple, latest attributable company guidance, the hypothesis target metric, benchmark and factor state, and the applicable risk-free rate. Every source as-of must be no later than the event cutoff. A missing value stays missing; a vendor restatement cannot replace the frozen observation; raw licensed payloads remain private while hashes or rights-cleared values form the public receipt. Late reconstruction and manual value entry are forbidden. Today P07.1 has selected no event, so P07.2 correctly shows 0 requests, 0 captured components and 0 baselines. P07.3 probabilistic underwriting remains blocked. No valuation, portfolio weight, trade or alpha claim was created.
P07.1 turns “take the next live event” from prose into a timestamped, hash-bound selector. The activation receipt freezes a strict boundary at 2026-09-11T09:05:06.290Z. The 2 Daily Answers already in the archive remain public and hash-visible but are structurally excluded from the live intake chain. Every later record is evaluated in one immutable order—recordedAt, coverageDate, contentHash. EVIDENCE_BLOCKED, NO_MATERIAL_CHANGE and missed days create automatic no-action receipts. Only the first MATERIAL_CHANGE with an exact Economic Graph event, an admitted hypothesis receipt and exactly one still-baseline-free frozen hypothesis can lock the slot. A later attractive event cannot replace it. Today the post-activation queue contains 7 receipts and zero selected events, which is the honest forward state. P07.1 cannot attach a baseline, create valuation, alter weights, trade or claim alpha. P07.2 opens only after sequence—not taste—selects the first eligible event.
A verified event is not a position. It must earn every bridge.
P07.0 freezes the Capital Consequence Contract before any favorable event can be selected. Each append-only Daily Answer now receives exactly one hash-chained docket. To approach even a shadow proposal, an event must pass ten gates in order: fact capture, evidence verification, Economic Graph binding, a frozen falsifiable hypothesis, a point-in-time market baseline, probabilistic underwriting, explicit risk and cost mapping, a shadow-only proposal, a governed receipt and pricing at the first eligible close strictly after governance. All thirteen consequence fields must be present before underwriting can advance. The current 9 Daily Answer receipts were consumed automatically; both were evidence-blocked, so the system preserved 9 permanent no-action receipts instead of manufacturing signals. No valuation, issuer verdict, weight change, trade or alpha claim exists. Broader Asian AGI supply chains are the first P07 expansion frontier, while P06 measurement clocks continue independently. The next test is not a hand-picked historical winner: P07.1 must consume the next eligible live event.
The machine is built. Its reliability still has to be earned.
P06-OBS.5 closes the finite observation-control build with a deterministic audit rather than a victory lap. It recomputes four dimensions from the exact hash-linked P06-OBS.1–4 projections. Cadence: one attempted bootstrap window delivered all 15 due receipts, so that run has 100% completeness, but there are zero scheduled windows and zero calendar span; reliability remains unproven until at least three attempted windows, two scheduled windows and 14 days exist. Extractor drift: one exact FactSet rule is fingerprinted, but zero live extracted values and zero comparable receipts cannot demonstrate stability; 11 schema gaps remain visible. Correction history: there are zero append-only correction receipts, which means none recorded—not perfect data; any future correction must name the prior object hash and preserve it. Downstream utility: zero material events have been admitted, bound to the Economic Graph or consumed by a Daily Answer, so the route is wired but untested in production. The audit therefore seals three explicit blockers and opens a controlled P07 handoff: the broader Asia study may reuse contracts, source lineage, extractor fingerprints, receipts and the admission gate, but inherits no reliability, investment or alpha claim. P06-OBS is complete; its live evidence clock keeps running independently. No valuation, weight or trade changed.
P06-OBS.4 freezes the first deterministic admission gate between metric comparison and interpretation. Every append-only comparison receipt receives exactly one hash-chained decision. Pre-activation records cannot be backfilled. EVIDENCE_BLOCKED, a first BASELINE and NO_MATERIAL_CHANGE all stop visibly; only a prospective direct COMPARISON_ASSESSED result classified STRENGTHENS or WEAKENS may become a research input. An admitted result creates one Economic Graph event routed only to EXPECTATION_GAP and enters the Daily Answer queue as evidence that the AI must independently reopen and interpret. Its direct-primary evidence grade, broad-market-proxy limitation and expiresAt remain machine-visible. The proxy cannot prove AGI causality, bind an issuer or security, set fair value, publish a Daily Answer, change a weight, authorize a trade or claim alpha. The live ledger contains zero decisions and zero admitted events because the comparison rail has not yet sealed its first baseline. That null state is published rather than replaced with a demonstration signal.
One number starts the clock. Only the second may say “change.”
P06-OBS.3 arms the first append-only comparable metric series without rewriting the frozen source contracts. The exact FactSet S&P 500 forward 12-month P/E extractor is now wired into the scheduled Daily market research workflow with ephemeral PDF extraction and hash-only evidence locators. Pre-activation transport receipts are ineligible: the next exact value becomes the first prospective baseline and remains NOT_ASSESSED. Only a second eligible value from the same source, unit and extractor version may be compared with the immediately previous value. A decrease of at least 0.5 turns is classified STRENGTHENS, an increase of at least 0.5 turns WEAKENS and a smaller move NO_MATERIAL_CHANGE. The threshold is an operational change gate, not fair value. Today the ledger truthfully contains zero baselines and zero comparisons; the next scheduled gate must either seal the first baseline or publish EVIDENCE_BLOCKED. Eleven schema gaps remain blocked. No graph event, Daily Answer, issuer verdict, valuation, weight change, trade or alpha claim is created.
P06-OBS.2 audited all 12 requested fields in the four monitoring contracts previously labelled DIRECT against their primary source pages. The label was not treated as proof of semantic fit. Only one field passed: FactSet states the S&P 500 forward 12-month P/E with an exact label, so a versioned extractor now requires exactly one match, a 0–100 numeric guard and a hash-only normalized-text locator. It stores neither the source body nor the matched sentence. The other 11 fields are explicit schema gaps: METR publishes task-completion time horizons rather than multi-day completion or intervention rates; Stanford publishes an employment gap and qualitative hiring mechanism rather than the frozen hiring-rate fields; FactSet does not resolve the meaning of consensus revision or coverage in those fields; and the EU AI Act timeline cannot yield one requirement count or three categorical restriction states without jurisdiction and taxonomy definitions. Two fields therefore measure a different construct, three are not reported and six are underspecified. The tested extractor is deliberately not active in scheduled collection and produced zero retrospective observations. P06-OBS.3 must preserve the frozen contracts, add governed amendments and activate eligible rules only on new prospective receipts. No graph event, Daily Answer, issuer verdict, valuation, weight, trade or alpha claim was created.
One page, two clocks. Collection is not interpretation.
The first one-screen research result exposed an important presentation error: it joined market and evidence receipts, but did not render the canonical Daily AGI Answer itself. A completed collection could therefore look newer than the latest interpretation. LOG-110 separates the two clocks. The market-and-evidence clock identifies the latest sealed retrieval; the AI clock identifies the latest append-only answer and explicitly labels it CURRENT or PENDING_AFTER_COLLECTION. The page now renders all seven mandatory answer fields, the admitted source manifest, epistemic states, the forward falsifier, next observation time and immutable answer hash. In the current public state, a newer market-research receipt exists than the latest Daily Answer, so the latter is truthfully labelled the previous sealed edition awaiting a new evaluation. The capital vault remains closed: neither clock changes production weights, authorizes a trade or claims alpha. The next research build remains P06-OBS.2—deterministic field extractors for direct-source monitoring contracts.
The clocks started. The machine refused to invent the answers.
P06-OBS.1 starts all 15 frozen monitoring contracts prospectively under the Daily market research workflow. The first immutable run created 15 hash-chained receipts after the contract and policy freezes. Fourteen official source objects produced transport baselines; Eaton remained unavailable after two bounded attempts. The system requested 45 typed fields and emitted exactly zero values because no deterministic field extractor is registered yet. Every receipt therefore remains EVIDENCE_BLOCKED and materiality stays NOT_ASSESSED. Raw source bodies were neither stored nor republished. No Economic Graph event, Daily Answer, issuer verdict, valuation, weight change, trade authorization or alpha claim followed. P06-OBS.2 will begin with deterministic extractors for direct-source contracts; proxy and baseline-only inputs remain visibly lower-grade.
A page can support a claim. It cannot choose a stock.
P06-SRC.6 closes the finite private-source normalization queue with a reproducible evidence map for the two China PDFs. The local files remain private and intentionally untracked. Their source hashes, deterministic text-map hashes, page counts and line counts are sealed; eight material claims now resolve to exact page-line coordinates and evidence-text hashes. The public projection republishes only two short excerpts—24 words in total—and leaves six claims citation-only. The NIU monograph supplies historical structure through 2020, not current capability. The Franklin Templeton paper supplies a secondary underwriting framework, not a proof of mispricing. Therefore all eight records stop at line-addressed context: zero Economic Graph events, zero Daily Answer admissions, zero issuer verdicts, zero valuations, zero portfolio changes, zero trades and zero alpha claims. P06-OBS.1 can now begin from governed forward observation contracts rather than repeatedly rereading private files.
A number earns a claim record before it earns belief.
P06-SRC.5 converts all 22 numeric rows in the sealed AGI workbook into typed, dated and source-bound claim records. The register preserves value, unit, period, epistemic state, source identity, scope limitation and an explicit failure test instead of flattening every number into a fact. It contains nine reported facts, six systemic estimates, one model forecast, one empirical result, one sample descriptor, three market snapshots and one industry statistic. Thirteen governed sources are referenced: 18 claims have a machine-reachable starting source, while four inherit visible access exceptions from the Oxford productivity study and Eaton source. A URL proves neither the value nor the exact passage. None of the 22 claims yet has eligible line-addressed evidence, and none is admitted to the Economic Graph or Daily Answer. The next rights-bounded gate will line-address eligible claims in the two China PDFs without republishing their full text. No issuer verdict, valuation, portfolio weight, trade or alpha claim was created.
A research question earns a clock before it earns an answer.
P06-SRC.4 converts all 15 monitoring rows in the sealed AGI workbook into forward-only observation contracts. Every question now names a registered source, source-fit grade, recommended cadence, typed fields and units, comparison rule, expiry and an evidence-blocked stale outcome. Four source bindings directly measure the requested domain; seven are partial proxies and four are baseline-only starting points. Fourteen of the 15 bound sources were machine-retrievable at the last source-register check; Eaton remains one visible timeout exception. This is a clock system, not fabricated telemetry: contract compilation activated no scheduler, established no baseline and created zero observations. The first future receipt must carry its evidence URL, values, comparison baseline, epistemic and materiality states, limitations, expiry and content hash. Missing data cannot become zero; a proxy cannot become direct evidence; an expired or unavailable input becomes EVIDENCE_BLOCKED. The next finite source gate converts 22 workbook fact rows into dated, source-bound claim records. No issuer verdict, valuation, portfolio weight, trade or alpha claim was created.
A sector earns a causal role before it earns a ticker.
P06-SRC.3 turns all 30 sector rows in the sealed AGI workbook into typed, auditable mapping receipts. Twelve sectors map directly to an existing Economic Graph role; fifteen are explicitly adjacent because the current enterprise role is broader than their downstream economics. Copper and materials, standard office real estate, and leisure or personal services remain three visible no-map results because the graph does not yet contain resource-input, demand-destruction or second-order-demand roles. Each sector has a stable ID, exact private-row fingerprint, multilingual identity, horizon classification and resolvable links to the governed source register. The system did not invent tickers to fill the gaps: every row states SECTOR_NOT_SECURITY and carries zero issuer-selection, valuation, portfolio-weight or trade authority. The next finite gate compiles the workbook’s 15 monitoring rows into dated source, cadence, field, unit, materiality and expiry contracts.
P06-SRC.2 converts all 33 source rows in the sealed AGI sector workbook into a governed register instead of an unexamined bookmark list. Every row now has a stable ID, exact row fingerprint, canonical publisher, geography, language, evidence role, rights mode and dated access observation; 30 distinct publishers across four declared regions are represented. A bounded one-time machine check reached 27 URLs directly and one through a recorded redirect. Four HTTP 403 responses and one timeout remain visible as five access exceptions—not mislabeled as broken sources, bad evidence or successful browser reviews. Every entry is link-and-metadata only; no raw response body was retained or republished. Availability does not prove a claim, source identity does not support a claim by itself, and this stage authorizes zero issuer verdicts, weights or trades. The next finite gate maps the workbook’s 30 sector rows into typed Economic Graph roles or explicit no-map results.
Five private local research originals are now sealed into a public-safe provenance contract without publishing the files themselves. Every object has a SHA-256 fingerprint, byte length, rights mode, ingestion state and explicit relationship to derived work; all five local hashes were verified before release. The two source directories are now excluded from Git, while CI validates the manifest and can reproduce the public projection without possessing copyrighted originals. The AGI sector workbook contributes a declared inventory of 30 sectors, 3 scenarios, 15 monitoring rows, 22 facts and 33 source-register rows—103 structured rows queued, not silently treated as evidence. A finite five-gate route now validates sources, maps sectors into the Economic Graph, compiles monitoring contracts, normalizes fact rows and addresses China PDF claims only where rights permit. A file fingerprint proves identity, not truth; conceptual influence is not line-addressed evidence; zero issuer verdicts, weights or trades were authorized.
The first external reviewer praised the falsifiers and honesty, then identified the missing bridge from evidence to risk and weights. P06-QUANT answers with a reproducible investment flight recorder: the gross daily-reset 2020–2026 current-basket replay has 28.7% volatility, −37.8% maximum drawdown, 1.09 QQQ beta and 0.95 QQQ correlation; targets without observable price history remain in SHV. Over the latest common 252 sessions, 24 eligible equities produce 0.31 weighted pair correlation but only 2.7 diversification-ratio-squared effective bets; SPCX is excluded from that window because 58 observations are insufficient. Three disclosed causal stress assumptions now show non-uniform layer losses, and the 20/55/25 priors mechanically generate a shadow layer mix without changing production weights. Risk bands and trade de-minimis are separated. The 2015 curve remains a selection-biased exposure diagnostic; a point-in-time vintage result is visibly blocked until a dated eligible universe exists. A probabilistic owner-earnings sandbox adds dilution, scenario probabilities and discounting. Zero trades and zero alpha claims are authorized.
Alpha begins with a prediction sealed before the outcome.
P06-NET.5 adds a first-class, append-only Alpha Hypothesis Ledger. Every canonical Daily Answer now receives a hash-chained admission receipt. Only a MATERIAL CHANGE bound to the Economic Graph and carrying one measurable, sourced and falsifiable forecast can enter. Before any baseline or outcome exists, the ledger freezes the target metric, direction, minimum magnitude, primary horizon, falsifier and countercase, then opens four forward observation clocks at 7, 30, 90 and 365 days. The current September 9 Daily Answer was EVIDENCE BLOCKED, so the public ledger correctly shows one rejected admission and zero hypotheses. No baseline was invented, no historical forecast was backfilled and no alpha, recommendation, trade or weight change is claimed. The next eligible live event—not a retrospective narrative—must create the first hypothesis.
One weak source must not poison three verified signals.
The first live autonomous run produced four signals. The independent evaluator substantiated three OpenAI developments but could not extract the claimed NERC finding from the cited page, so the machine correctly published no substantive conclusion and sealed EVIDENCE BLOCKED. P06-NET.4B learns without lowering the bar. The evaluator must now decide every signal separately. When an eligible evidence exception leaves at least one verified signal, one—and only one—remediation pass may quarantine the failed source, rewrite the complete multilingual Daily Answer using no new evidence, and submit the smaller packet to a fresh independent evaluator that reopens every remaining material domain. The rewrite cannot add a source, restore an excluded signal or loosen the capital firewall. Any invalid mapping, failed final gate or unresolved ambiguity still ends as EVIDENCE BLOCKED. This is evidence salvage, not conclusion salvage.
Five channels do not get five versions of the truth.
P06-NET.4 freezes one distribution contract for the site, Telegram, X, LinkedIn and video across English, Ukrainian and Russian. A sealed Daily Answer can create 15 channel-language packages, but each package must preserve the exact record hash, revision ID, outcome, epistemic states, capital state and complete source manifest. Adapters assemble only intact canonical fields: if a limit is too short, the whole field is declared omitted rather than truncated or paraphrased. X carries a canonical pointer instead of a compressed new claim. External publishing remains disabled and every package stops at an account gate. Because no Daily Answer was sealed at build time, the public output is honestly five armed adapters, 0 packages and 0 external posts—not demo content. Synthetic tests prove the 15-package path; the first forward edition must prove it with a real sealed object.
P06-NET.3 freezes the first AGI Economic Graph foundation: 103 typed nodes and 135 typed edges connect four monitored labs, seven causal stages, nine value-capture roles, 32 organizations, 40 securities and seven venues. All 26 U.S. baseline positions remain mapped; IBM remains a zero-weight monitored candidate; the five China pilot organizations resolve to 13 distinct securities; SPCX remains an explicitly non-exchange-traded private reference. Eighty identity edges say what an instrument is; 47 hypothesis edges say what to investigate. Neither class proves causality, valuation, alpha or a trade. The sealed-event count is honestly zero at publication: the event rail accepts the first canonical Daily Answer when it exists, but no historical event was fabricated to populate the interface.
P06-NET.2 replaces the flattering source count with a frozen coverage denominator: ten regions by six causal lanes create 60 minimum obligations. The current allowlist fills 11 and leaves 49 visible gaps. All 21 sources are English-language; 85.7% are published from North America; only 1 of 12 target languages is represented. Probe latency remains explicitly unknown for 21 of 21 sources until a new forward radar run records it—historical snapshots are not backfilled. The public surface now exposes the complete matrix, ranked gap queue, source classification and rights boundary. It stores links and metadata only, republishes no source text, authorizes no capital action and makes no alpha claim.
P06-NET.1 turns the daily promise into a forward-only machine. One strict multilingual schema requires seven fields, dated sources, epistemic labels, a falsifier and a capital boundary. The generator seals each eligible date into a SHA-256 chain; a repeated identical run is idempotent, a rewrite fails closed, and any unsealed date becomes a permanent MISSED DAY receipt before the next answer can enter. The public ledger begins honestly at 0/14 because today’s first window has not opened. Two Kyiv-time scheduler checks handle daylight saving without publishing twice. Fourteen consecutive dry-run objects can make the system eligible for activation review, but cannot switch the public SLA on, change portfolio weights, authorize a trade or prove alpha.
P06-NET freezes the fund’s focus before delivery metrics can reward it: every calendar day, one evidence-bound answer states what changed—or did not change—in AGI, why it matters economically and what it means for capital. Capital allocators are the primary audience. Operator and practitioner consequences remain conditional lenses inside the same canonical object, not separate products. MATERIAL CHANGE, NO MATERIAL CHANGE and EVIDENCE BLOCKED are all valid outcomes. The SLA is explicitly off until a 14-day dry run passes; the finite track ends at P06-NET.5 after a 30-live-day audit.
P06.4 freezes the only path from a signed Adaptive decision to a research position. Approval seals intent but carries no price. A synchronized close already recorded before the decision is ineligible—even on the same market date. The first completed close recorded strictly later binds all 26 market objects, deducts the proposal’s modeled cost before allocation and seals fractional adjusted-close total-return units in an append-only receipt. Those units affect only later forward observations, so the observed close cannot grade the decision that selected it. The build is live ahead of the clock; current state remains zero price receipts, zero changed production weights and zero real-money orders.
P06.3 freezes the decision layer before a result exists. The first synchronized forward close in a new UTC month seals the preceding decision window. If no proposal completed all nine evidence-to-position stages, the machine writes one immutable NO ACTION receipt automatically—no manual analysis is required. A proposed Adaptive target change must keep the same 26-security universe, 100% long-only weight, at least 3.5% SHV, no position above 10%, modeled costs, a frozen next-close rule and evidence hashes at every stage. Only an exact GitHub-bound APPROVE, REJECT or DEFER may close that proposal. Even APPROVE remains an unpriced shadow intent: P06.3 changes no target, production weight or real-money order; P06.4 alone may apply it at the first synchronized completed close strictly after the receipt. Current state remains honestly at zero receipts until observation zero and a forward decision window exist.
P06.2 arms a forward-only observation ledger before observation zero exists. The nightly workflow will first freeze one synchronized opening close and only a later synchronized close may become return observation one. Every daily receipt binds 26 market-object hashes to the immutable activation hash, values both arms on the same adjusted-close total-return basis and advances one append-only chain. A stale ticker stops the mark; a duplicate date is idempotent; a rewritten or backfilled record fails closed. Core and Adaptive remain mathematically identical with zero excess return and zero target mutations until P06.3 creates a separately governed pre-price decision or no-action receipt. The phase is finite: P06 closes at P06.5 with a 90-day process audit and frozen handoff to P07—not P06.1700.
The seal is live. The clock still waits for one honest close.
P06.1 freezes the actual production receipt edfed0a77191-34320302004-1 and arms EXP-ALPHA-001-RUN-000 inside the scheduled nightly market workflow. Activation is mechanical, one-time and immutable: all 26 frozen positions must carry the same completed close on or after 9 September 2026. Only then will one content-addressed opening snapshot assign identical prices, capital and fractional units to AGI Core 25 and AGI Adaptive. Today the shared cache still ends on 4 September, so the public record remains at zero forward observations, zero changed weights and zero trades. The first return observation must come after observation zero; no backfill can improve the starting line.
The machine may learn daily. Capital does not have to twitch.
EXP-ALPHA-001 freezes the first killer-feature demonstration. AGI Core 25 is the quarterly-banded control; AGI Adaptive starts from the identical portfolio, capital and timestamp, but may make one governed target decision monthly or open a rare event window only after seven material-event gates pass. Daily evidence collection and weekly/event-driven underwriting cannot move capital by themselves. Every difference must preserve its fact, causal map, expectation gap, instrument, size, book, hedge, falsifier, signed decision and first eligible close after that decision. The contract begins with zero observations, changes no production weight and makes no alpha claim.
AGI changed tense. The evidence has not closed the case.
September’s rhetoric shift is real, but the public record supports a phase-transition hypothesis—not a binary “AGI arrived” verdict. GPT-6 Astra and Claude Fable 5.1 broaden autonomous professional work; ARC Prize confirms a step change while warning that its 99.9% Astra result depends on the provider adapter and is not AGI proof. OpenAI reports a supervised research intern for tasks taking skilled researchers days and expects progress toward RSI; Anthropic confirms an unreleased Model 2 but describes it as broadly comparable to Mythos 5. The unconfirmed “Bel” rumor is excluded. We register a March 2027 ± three-month forecast at a 55% internal prior, six pre-committed gates and zero capital authorization.
China may compound intelligence twice. Neither loop is alpha yet.
Five evidence layers converge on two hypotheses. The frugal-diffusion loop lowers task cost, expands adoption and deepens toolchain, workflow and data dependence. The embodied-feedback loop connects world models to deployed machines, proprietary physical data and improving control policies. Alpha is not “China wins AGI”; it is an expectation gap where one loop becomes durable customer dependence and owner cash before price reflects it. Four forward evidence lanes are now specified, but their collectors are not yet connected. No issuer verdict, valuation, weight or trade is authorized.
We did not choose China’s five winners. We chose five ways the method could fail.
P05.21 closes the finite China pilot and opens P06. The handoff freezes 17 byte-exact artifacts, 7 portable controls, 3 contained incidents and 6 monitored routes. Alibaba, Tencent, Baidu, SMIC and GDS were not selected by expected return: they form a maximum-contrast stress test across cloud, distribution, frontier models, semiconductor sovereignty and data-center power; across cash-rich and capital-intensive economics; and across ADR, OTC, Hong Kong and A-share access. The result is a reusable control plane, not a top-five ranking: 0/5 issuer syntheses, zero valuations, zero weights and no trade. P06 must now freeze the matched US × China hypotheses and measurement rules before viewing outcomes.
The pilot did not produce five stock picks. It produced seven controls.
P05.20 replays the full China pilot from immutable receipts and finds 7 controls that can transfer to other markets, 3 invalid-output incidents safely stopped before publication, and 0 capital leaks. The history is deliberately non-monotonic: rule eligibility ranged 4–5 and automatic gap routes ranged 5–6. Today 0/5 issuers are synthesis-eligible, 5 remain evidence-blocked and 6 routes continue. P05.21 will freeze the verified machinery and explicit gaps for P06—not declare China alpha, valuation, weight or trade.
P05.19 runs one deterministic issuer-synthesis gate across all 5 China dossiers. Result: 0/5 issuers are synthesis-eligible and 5 are EVIDENCE BLOCKED. The gate requires exact identity, bound receipts, governed decisions and sufficient evidence on both mandatory tracks; 6 automatic routes keep working against the missing prerequisites. A missing fact, inconclusive packet or rule-level FAIL cannot become a company verdict, valuation, weight or trade. P05.20 can now audit the pilot without pretending the evidence is complete.
P05.18 separates 5 companies from 13 distinct securities, their venues, currencies and depositary relationships. ALI remains an internal company ID; BABA, TCEHY, BIDU and GDS are instrument symbols; SMIC is a company name whose investable claims include HKEX 00981 and SSE STAR 688981. The selector now leads with company names, every security carries an explicit venue/type/currency, and all five companies have standalone dossiers. Current machine state is read from the live receipts—6 evidence gaps across 5 source lanes—rather than frozen narrative. This is identity control, not valuation, weight or trade authorization.
A missing fact needs an address, not a human backlog
P05.17 turns the six current evidence gaps into executable routes. Six frozen tasks now point to five official source lanes, exact machine search terms, unchanged PASS/FAIL boundaries and the next GitHub cycle on 13 September at 20:41 UTC. GDS deliberately shares one source lane across two different questions. An unchanged page preserves the gap without creating manual work; a qualifying document enters the existing archive, citation, model and adversarial chain. This is routing—not new evidence, an issuer verdict, valuation, weight or trade.
P05.16 now joins five issuers and two frozen questions into one deterministic evidence matrix. Four maintenance cells carry signed FAIL rules because their current packets do not quantify a sustaining boundary; all five shareholder-access cells and GDS maintenance remain inconclusive. That produces ten addressable research states, not an issuer ranking: every synthesis gate, valuation, portfolio weight and trade remains at zero. P05.17 will target the six named missing evidence objects and rerun the same contracts without erasing this receipt chain.
The fifth packet broke the sentence—not the firewall
The first live cycle after P05.15 legitimately expanded the governed docket from four candidates to five. A test and headline still expected four, so the quality gate blocked the new snapshot before publication. We removed the historical cardinality from the invariant: the packet must now match the exact current docket by ordered task IDs, receipts and count, while the interface renders that count from verified data. Research completed; stale narration failed closed; capital remained disconnected.
The new governance decision packet resolves every current candidate into an exact question, machine case, strongest counter-case, line-addressed primary citations and explicit overturn condition. The live evidence cycle also moved SMIC shareholder access back from candidate to blocked/inconclusive, proving that eligibility is recalculated rather than inherited. Four machine recommendations remain non-binding and unsigned; issuer verdicts, valuations, weights and trades remain zero.
Five candidates reach governance. Capital stays outside.
The next production evidence cycle changed the docket from four eligible tasks and six blocked tasks to five and five: SMIC shareholder access now also supports its frozen FAIL rule. P05.15 does not silently inherit or auto-approve that result. It creates an append-only register where an identified reviewer may approve, reject or defer only the exact candidate hash. All five current candidates begin pending; a changed docket resets their authority. There are still zero issuer verdicts, valuations, weights and trades.
Four signals enter the docket. Zero stocks are convicted.
The completed P05.13 production packet confirmed all ten source-bound adversarial checks. The deterministic P05.14 docket admits four maintenance readings—BABA, TCEHY, BIDU and SMIC—only as candidates to record that their frozen FAIL criterion is supported. Six inconclusive tasks remain evidence-blocked. Candidate is not verdict: no issuer passed or failed, and owner cash, valuation, rank, weight and trade remain unauthorized.
The machine conclusion gets a prosecutor, not a friend
P05.13 binds a separate model response to the exact P05.12 analysis hash and gives it the same task-specific archive—not the open web. Its mandate is to attack citation entailment, period, perimeter, unit, reconciliation, contradiction and missing evidence. The deterministic gate permits only confirmation or downgrade to inconclusive: it cannot upgrade an inconclusive reading, reverse PASS support into FAIL support, cite another task or authorize owner cash, valuation, weight or trade. The ten-task run is queued for GitHub; zero verifications are claimed locally.
The second P05.12 GitHub run kept every citation inside the correct task lane, but the model repeated one permitted EQ-SMIC-ACCESS citation ID. The independent validator rejected the duplicate before audit, commit or publication. We now canonicalize only exact duplicate IDs at the model boundary and then run the unchanged validator: unknown and cross-task IDs still fail closed, while repeated evidence never increases support. No machine reading, metric, valuation, weight or trade was published.
The model crossed a citation lane. The system stopped it.
The first GitHub P05.12 run reached the archive-only model assessment, but the EQ-SMIC-ACCESS result referenced a citation assigned to another frozen task. The validator failed closed before audit, commit or publication. We kept that boundary and moved it into the structured-output grammar: every task now has its own required response key, fixed task ID and closed enum of permitted citation IDs; a task with no evidence can cite nothing. The rerun must pass the same independent validator before any machine reading is published.
The extractor becomes infrastructure, not a one-off script
P05.12 separates a universal Evidence Engine from China-specific contracts and presentation. New issuers can reuse the same content addressing, HTML/PDF extraction, citation validator, model boundary and audit chain; they need source adapters and frozen rules, not a rebuild. The first GitHub model run remains the next controlled event; no research verdict or capital action is authorized.
The new schema may say that archived citations support a frozen PASS or FAIL rule, but it cannot promote that statement into owner cash, valuation, rank, weight or trade. P05.13 must independently attack citation scope, period, recurrence and reconciliation before any research verdict changes.
The current immutable packet extracted 43/43 contract-level documents with zero failures, deduplicated them into 31 content-addressed text maps and found 26 exact contract-term citations. The local environment has no publication secret, so the public state honestly says the model run is scheduled. GitHub will perform the archive-only assessment after the controlled push.
HTML and PDF now enter one auditable coordinate system
A Node semantic extractor and pinned pypdf page extractor now normalize archived source bytes without re-fetching them. Every derived text map has its own SHA-256 address; HTML receives semantic line coordinates, PDF preserves page and line coordinates, and every public quote carries locator plus quote hash.
P05.12 pre-commits the missing evidence layer: extract archived HTML and PDF deterministically, preserve page/line locators, cap public quotations at 25 words per source, bind every quote to source and text-map hashes, and let the model use no evidence outside that packet. A mention is not sufficient proof and extraction cannot authorize capital.
P05.11 completed the acquisition layer and authorized no investment conclusion. A document, a matching passage, a research verdict and a capital action remain four different states. P05.12 will make archived HTML and PDF evidence line-addressable and run structured machine assessment; owner cash, valuation, rank, weight and trade remain locked.
The second collection exposed dynamic page noise: raw bytes can move while issuer content does not. The collector now preserves the exact raw SHA-256 object but detects change with a separate semantic fingerprint over normalized text and official document links. Transport failure is also distinct from no signal; Baidu and GDS use disclosed SEC fallbacks when issuer routes time out.
The first complete run reached all ten contract routes, archived 33 contract-level candidate documents and exposed 12 bounded passages. Every frozen-rule test returned insufficient evidence because no packet yet proves the exact period, perimeter, reconciliation and recurrence required for PASS or FAIL. This is a successful acquisition result, not failed research.
Every fetched endpoint and candidate document is compressed once under its SHA-256 filename, outside the public build. Each append-only run names the frozen contract hash, object hashes, transport route, HTTP metadata, extracted passages and the previous run hash. The audit reconstructs the chain and re-hashes every archived object before release.
P05.10 authorizes collection, not conclusions. All ten contracts remain QUEUED / NOT RUN and owner cash, valuation, ranking, weights and trades remain locked. P05.11 will execute a collector, preserve hashes and source documents, extract candidate passages and test the frozen rules without allowing a page change to become a capital decision.
An index diff can discover a document, but it cannot prove maintenance capex or recurring shareholder access. The system now requires a primary document, an extracted metric and satisfaction of the rule frozen before collection. The review-by date is our freshness control—not issuer guidance.
Alibaba, Tencent, Baidu, SMIC and GDS each now have one maintenance contract and one shareholder-access contract. The public surface exposes ten official discovery endpoints, ten exact metrics, ten internal expiry controls and twenty frozen decision rules. Evaluated tests and passed tests both remain zero.
The Filing Lab now renders an interactive acquisition surface for maintenance capex and recurring shareholder access. Each job separates already-read baseline evidence from the official endpoint the future collector will monitor, and publishes its metric, search terms, cadence, review-by control, PASS and FAIL rule.
P05.10 pre-commits two issuer-specific questions before collection: what evidence can narrow the maintenance-capex range, and what evidence can prove recurring cash access to the listed owner. Every request must name baseline evidence, an official discovery endpoint, exact metric, internal expiry, PASS and FAIL rule; none may authorize owner cash or capital.
P05.9 authorizes no maintenance point, owner-cash number, valuation, rank, weight or trade. P05.10 will target the exact evidence capable of narrowing each issuer’s maintenance range or proving a recurring path to the listed owner; until a gate passes, null remains the fiduciary answer.
A 50% maintenance assumption is visually useful because it reveals how quickly the bridge bends between hard endpoints. It has no evidentiary privilege over 0% or 100%, so the system labels it NOT AN ESTIMATE and forbids its use as a base case, target or portfolio input.
Five envelopes produce three different cash regimes
Under both hard endpoints Tencent remains positive and Baidu remains negative. Alibaba, SMIC and GDS cross zero as the maintenance share moves. This separates robust sign from assumption-sensitive sign—but access remains heterogeneous, so it is not a ranking of shareholder value.
The Filing Lab now renders 15 current-period sensitivities across five issuers, applies seven resolved adjustments exactly once and carries all 11 access routes beside the arithmetic. Cash stocks, legal capacity, restricted funds and funded EV claims remain outside annual operating flow.
Three bounded questions before one owner-cash answer
P05.9 pre-commits three mechanical maintenance lenses: 100%, 50% and 0% of current cash capex. The midpoint has no forecast status; resolved adjustments may enter once; balance-sheet stocks and EV claims cannot enter annual flow; shareholder access is a proof gate, never an invented haircut.
P05.8 maps 36 liquidity and claim records plus 11 access routes, but does not manufacture a distributable-cash number. P05.9 may construct bounded scenarios only if cash stocks remain outside annual flows, recurring claims are deducted once and inaccessible funds stay outside the listed-owner perimeter. Valuation, ranking, weights and trades remain locked.
P05.8 separates four identities that conventional screens often compress: balance-sheet liquidity, period cash flow, funded enterprise-value claims and legal access gates. Adding group cash to FCF or deducting the entire debt balance from one year of cash generation would be a category error; the model now rejects both.
Alibaba received RMB101.294bn of subsidiary dividends in FY2026; Baidu received RMB6.5bn from one subsidiary in FY2025; Tencent paid RMB37.535bn to shareholders. By contrast, SMIC paid no cash dividend and explicitly retained funds for expansion, while GDS reported no upstream dividend or distribution in 2023–2025 and funded non-VIE subsidiaries with RMB889.0m in 2025.
The Filing Lab now exposes 36 parent-liquidity and mandatory-claim records with primary-source anchors, 11 cash-access routes and explicit model treatment for every item. Users can see what reached the parent, what reached common shareholders, what is restricted and what belongs in enterprise-to-equity value instead of annual FCF.
P05.8 pre-commits the accounting boundary: cash stock, recurring period flow, enterprise-value claim and legal-access gate can never share one arithmetic column. We will trace cash from operations through subsidiaries and VIEs to the listed parent and shareholder before computing any owner-cash range.
A range cannot become owner cash before access is proven
P05.8 will classify leases, preferred claims, restricted cash, VIE transfers and retained minority stakes across all five issuers. Until mandatory claims and shareholder access pass, the 15 maintenance envelopes authorize no valuation, rank, weight or trade.
GDS disclosed RMB4.6106bn of reported 2025 capex but RMB4.691137bn of cash payments for PP&E and land-use rights. P05.6 incorrectly labelled the first number cash capex. The current bridge and public correction now use the cash-flow statement amount.
Fifteen periods mapped; maintenance points remain zero
Alibaba, Tencent, Baidu, SMIC and GDS now have three audited annual OCF and cash-capex periods. Nine company-defined FCF periods reconcile exactly. Every maintenance range remains the honest hard bound from zero to cash capex because no filing discloses a reproducible split.
A three-year cash-claims control plane is now live
The new workbench separates operating cash, cash PP&E, reported capex, filing-defined FCF claims, unclassified claims, maintenance bounds and structural access barriers. Each annual record links directly to a primary filing.
P05.7 pre-commits a hard rule: annual maintenance starts at 0 and ends at cash capex until a scope-matched filing supports a tighter bound. Repetition, depreciation and “most capex was expansion” may inform investigation, but none becomes an undisclosed fact.
All five reported cash bridges are reproducible, but zero maintenance/growth splits and zero owner-cash measures are approved. P05.7 will build three-year recurring and expansion cash-claims histories. Valuation, ranking, weights and trades remain unauthorized.
Growth capex can make cash look weak without proving owner value
Alibaba, SMIC and GDS show negative observed post-capex residuals while filings also point to expansion investment. We will neither punish every growth dollar as maintenance nor excuse it as valuable growth: both claims require evidence and future returns.
OCF minus cash capex is now reconciled for all five dossiers. Tencent, Alibaba and Baidu company FCF definitions are also reproduced line by line. The prior Baidu conflict was our period error: RMB2.6bn was Q4 OCF, while RMB−3.013bn was FY2025. The correction is visible in the workbench and LOG-016. The GDS cash-capex label in this entry was later corrected by LOG-048.
Each issuer now exposes operating cash flow, cash capex, observed residual, company-defined FCF adjustments, missing maintenance/growth split, mandatory claims and primary sources. Three company definitions reconcile; none is silently treated as cross-company owner cash.
P05.6 pre-commits five rules: same-period cash inputs, observed residual before interpretation, line-by-line company FCF reconciliation, null maintenance when undisclosed, and three approval gates for claims and shareholder access.
Thirty instrument cohorts are now classified, but zero expected diluted bases are approved. The next step is P05.6 owner-cash normalization; dilution stays locked until cohort roll-forwards, settlement offsets and licensed date-matched prices are reproducible. No ranking, weight or trade is authorized.
Tencent’s option ceiling was not its instrument inventory
The earlier 46.198m new-share option ceiling captured neither the full option balance nor awards. At 30 June Tencent disclosed 73.659m options and 118.201m awards; H1 exercises also show why gross options cannot equal new shares: 13.185m exercises produced only 5.251m net shares after price and tax settlement.
Base, expected and stress are now separate chambers
The public workbench no longer presents one diluted-share shortcut. Basic is an evidenced economic base, expected remains intentionally uncomputed, and stress holds maximum disclosed delivery scenarios. Converted debt and preferred claims must leave enterprise value in the same scenario that adds their shares.
Alibaba, Tencent, Baidu, SMIC and GDS now have a common 30-cohort ledger covering awards, options, convertibles, preferred shares, delivery reserves, offsets, historical denominators and plan authorization. Every cohort carries an evidence date, settlement route, treatment and primary-source link.
P05.5 pre-commits five treatments: candidate, stress, offset, authorization and historical. We will not sum plan capacity, outstanding grants, weighted-average EPS denominators and maximum conversions as if they were the same economic object.
Situational Awareness had real capital, execution access and an early infrastructure thesis; its Q1 SEC 13F showed $13.676bn across 42 reportable entries. Later reporting described a 67% July portfolio decline and deleveraging, while the fund remained up 80% for 2026. Our conclusion is not that we are the better fund: we have no client capital or demonstrated alpha. Our designed edge is prospective public memory, global matched challengers and a survival-first mandate in which crowding, financing and exit liquidity are separate gates. No allocation changes and no trade are authorized.
A search result triggered the review; IBM, OpenAI, Together AI and SEC sources set the evidence boundary. IBM is now a 0% monitored candidate: the thesis is an enterprise AI integration tollbooth, not frontier-model ownership. Promotion requires a reproducible bridge from AI bookings to recurring software margin and free cash flow per diluted share. The 25-company portfolio and SPCX remain unchanged; no order is authorized.
Do not call a fresher document a fresh share count
Baidu and GDS still fail the full-count freshness gate. Expected dilution remains uncomputed for all five companies; valuation, ranking, prices, weights and trades stay unauthorized.
Authorization can look like dilution and still not be dilution
Baidu’s 10% plan limit and GDS’s S-8 registration are capacity disclosures. GDS’s filing even combines new evergreen reserve with already-issued returned shares. Adding either wholesale would overstate expected dilution.
Evidence cadence and expected-dilution workbench published
The five-company control surface now shows legal-count dates, latest usable evidence, monitoring routes, authorization boundaries and the missing inputs between observed basic shares and a defensible expected denominator.
Baidu became dual-primary on 1 September and now enters monthly-return reporting, but no first return exists yet. GDS remains secondary-listed; its 4 September interim report confirms 43.590m Class B shares at 30 June, not a complete current cap table.
Separate count date, publication date and authorization
Before searching for a fresher denominator, pre-commit three controls: a filing date cannot roll a count forward; authorization is not issuance; and expected dilution requires grant or conversion evidence plus economic offsets.
Five legal anchors are now available, but Baidu and GDS fail the seven-day freshness gate and no company has an approved expected-diluted denominator. Prices, targets, weights and trades remain unauthorized.
GDS includes depositary award reserves in legal outstanding shares and uses borrowed delta-placement shares around convertible issuance. The policy now separates legal, economic, expected-diluted and stress layers and prohibits double counting borrowed shares with conversion.
All five companies now pass through the same four denominator layers, freshness rule and conversion-offset test. The public matrix shows base, scenario uplift, exclusions and approval state company by company.
Primary corporate documents anchor Baidu at 2.713492250bn ordinary shares on 14 July and GDS at 1.603020903bn legal ordinary shares on 31 May. Both are dated evidence—not target-date spot counts.
Pre-commit four layers—legal, economic common, expected diluted and stress ceiling—with a seven-day freshness gate. Weighted-average diluted shares may inform a scenario but cannot silently replace a dated spot base.
Do not substitute diluted averages for spot shares
P05.2 improves every denominator, but Baidu and GDS still lack a date-matched legal roll-forward. No valuation, ranking, weight or trade is authorized until one dilution policy can be applied consistently.
GDS Q2 diluted weighted-average shares exceed the basic denominator by 520.713m ordinary shares, or 33.5%. The prior 210.103m partial-FD ADS shortcut is retired in favor of an explicit range.
The public ledger now distinguishes legal issued shares, post-cancellation economic shares, disclosed dilution signals and unresolved instruments for Alibaba, Tencent, Baidu, SMIC and GDS.
Pre-commit the control: legal issued, economic post-cancellation and diluted equivalents must carry separate dates, formulas and evidence. Missing buckets remain visible.
The workpapers can reproduce provisional market-cap arithmetic, but quote dates, share dates and evidence perimeters do not align. No ranking, price target, weight or trade is authorized.
Correction: compare periods before declaring a conflict
The original entry incorrectly called Baidu’s figures conflicting. RMB2.6bn was Q4 operating cash flow; RMB−3.013bn was FY2025 operating cash flow. The annual bridge is internally consistent. The separate SMIC correction still stands: US$8.400bn was 2025 cash capex; US$7.664bn was the 2024 comparator.
Each dossier now has an exact-security claim, share denominator, line-item liquidity and debt bridge, OCF/capex bridge, dated quote and FX inputs, formulas and blockers.
Five daily closes and three FX observations are stored with separate dates and a provider boundary. The feed is for prototype research, not licensed execution.
The next work is security-level claims, diluted shares, net cash/debt and maintenance-versus-growth capex. No relative ranking or weight is authorized before that bridge is reproducible.
Baidu discloses the clearest AI revenue map; Tencent shows the strongest group cash funding. Alibaba, SMIC and GDS make the capex bridge the central question. Narrative strength alone cannot rank them.
The first pass records security identity, audited anchors, AI evidence, cash bridges, risk flags and unresolved questions for Alibaba, Tencent, Baidu, SMIC and GDS.
A single comparable surface now renders reported financials, AI linkage, owner-cash tension, primary sources and explicit valuation blockers in three languages.
Each company must answer the same identity, annual-filing, AI-evidence, owner-cash, governance and valuation questions. Missing disclosure remains missing; it will not be inferred into a score.
Mapped public candidates, private laboratories, market-access routes and eight pre-investment gates. This is a research set, not a portfolio or recommendation.
Every US baseline equity has a research dossier and valuation workpaper shell. Approval remains zero until inputs become reproducible and point-in-time.
Predictions will be timestamped with probabilities, expiry and resolution criteria. Until observations resolve, the public alpha-claim count remains zero.
