John Collins Alpha / Singapore
JOHN COLLINS ALPHA / CUMULATIVE INTELLIGENCE
The evaluator found 149 recommendations and zero defensible scores. That was the correct result. Without a proved first tradable entry and benchmark mark, a later price is hindsight—not performance.
01 / THE CUMULATIVE STATE
The accumulated state now applies one discipline across the account, AI, markets and commerce: production is an input, while point-in-time evidence determines what the system may honestly claim.
A full valuation rebuild withdrew the blanket claim that the portfolio was more expensive than everything else. Archetype, forward expectations, revisions, dispersion, catalysts, cycle normalisation and risk now govern the comparison.
Open and inspectable model or harness layers can accelerate research production, but the system becomes decision-grade only when point-in-time evaluation keeps pace.
Drawdown, correlation, short volume, earnings breadth and demand signals remain descriptive until chronology, controls, conversion, financing and competing mechanisms survive testing.
Field evidence can show adoption and time savings without broad near-term task or earnings effects. Capability and use remain intermediate states.
02 / HOW EXPERIENCE COMPOUNDS
A cumulative system is useful only when publication truth, missing evidence and superseded conclusions propagate into the next decision. This interval did all three.
prior verified public state
unpublished candidate plus subsequent corrections
finite research production and source receipts
point-in-time evaluation and market adjudication
surviving claims, withdrawn claims and honest unknowns
Begin from the last verified public state, then carry every later action, failure and correction without promoting an unpublished candidate into history.
Declare the research set, producer archive and evidence gaps before interpreting a ranking or performance claim.
A later quote cannot create a return unless the first eligible trading session, adjusted security mark and benchmark mark are all proved point in time.
Coverage, provenance and abstention can prove process integrity. Only matured prospective evidence can prove or disprove investment efficacy.
Expose aggregate counts, method, limitation, rival and falsifier while keeping private recommendations, portfolio state, paid research and relationship evidence private.
03 / SINCE 18 AUG 2026 / 17:14 SGT
The interval expanded finite research coverage, withdrew an overbroad valuation claim and made the evaluation gap impossible to hide behind recommendation volume.
The system rejected a trailing-yield and equal-weight verdict that ignored archetype, actual exposure, forward expectations and cycle sensitivity. The replacement is conditional, not the opposite blanket story.
A full research sequence declared its opportunity set, producer roles and evidence packets. That improved architecture and auditability while leaving ranking quality sensitive to known market-data gaps.
Project 05 became a bounded Citadel Securities public-intelligence organ, separate from general news. Specialist screens completed without a supported alpha claim; paid text remained private and public facts required primary verification.
The evaluator retained every accessible registered record, but missing first-entry-session evidence prevented return, alpha, catalyst or falsifier scoring. Unknown was not rewritten as zero or backfilled from a later price.
A system can produce a ranked, sourced 0–3 months board without possessing a track record. Each record must cross immutable publication, first-tradable entry, benchmark-normalised observation and terminal adjudication before hit rate or alpha exists.
149 recommendations were registered in 15 immutable packets from seven producer roles.
This proves accessible archive coverage and provenance, not research efficacy or global historical completeness.
Every record specifies the next tradable regular-session open as its entry rule.
A closed calendar receipt proving the first common eligible session and licensed adjusted-return security and benchmark marks.
All 149 records stayed in the denominator and received an impaired observation record.
Without a valid entry, no return, alpha, catalyst or falsifier score exists; current or latest quotes are not substitutes.
The evaluator can state exactly why it cannot score and preserve the failure append only.
Matured 0–3 months cohorts with reproducible terminal evidence. No hit rate, calibration or prospective-alpha claim is permitted.
Deterministic append-only evaluator; complete accessible valid-v2 producer census; immutable packet hashes; exact 0–3 months; first common entry open and first terminal close; normalised benchmark; price alone never adjudicates a thesis. No matched enterprise head-to-head is claimed.
04 / THE ELEVEN-PROJECT NERVOUS SYSTEM
Projects 00–08, this surface and Project 10 perform different functions in one loop. The surface is responsible for preserving their relationship—not flattening them into a weekly list of evidence nodes.
Freeze context and reconcile the research sequence
Discover and rank live market change
Reconstruct the completed session
Maintain the event and dependency clock
Ground claims in primary disclosures
Test public Citadel Securities market intelligence and positioning
Extract rights-safe specialist themes
Interpret the physical AI stack
Attack downside and rival explanations
Render the public-safe cumulative state
Observe what later confirmed or broke
05 / THE PRESENT PROOF
The evaluation receipt makes one practical advance: it shows exactly where a sourced recommendation ends and a defensible performance claim begins.
AI investment systems can generate rankings faster than the point-in-time entry and benchmark evidence required to measure them without hindsight.
The evaluator previously had partial archive registration and no basis for claiming a complete accessible producer census. Recommendation volume could therefore be mistaken for evaluated performance.
The system discovered every accessible valid-v2 packet across seven producer roles, verified immutable packet identities, registered 149 recommendations, applied the same first-entry and exact 0–3 months rules, and preserved each failed observation append only.
Fifteen packets and 149 recommendations were accounted for with zero archive-coverage issues in the declared census. All 149 failed closed at the missing entry receipt, leaving zero evaluable outcomes and no manufactured hit rate or alpha.
This proves abstention and archive coverage for the accessible valid-v2 census, not historical completeness, research quality, portfolio performance or prospective alpha. The cited evidence includes no matched enterprise deployment.
06 / COVERAGE RECEIPT
Coverage is scoped, dated and explicit. Missing evidence remains unknown or impaired; it is never converted to zero, backfilled from a later price or hidden behind recommendation volume.
07 / A USEFUL TEST
Require every recommendation to cross immutable publication, valid first entry, benchmark-normalised observation and terminal adjudication before accepting a hit rate or alpha claim.
08 / PUBLIC EVIDENCE
SpaceXAI / 15 July 2026
Official evidence that a coding-agent harness can expose its context assembly, tool dispatch, extensions and local-first operation.
Inspectability and forkability do not establish production reliability, adoption or economic outcomes.SpaceXAI / 19 August 2026
Official product evidence preserving the unpublished candidate's useful distinction between an inspectable harness and a broader hosted surface.
Provider description; reliability, adoption and total economics are not independently measured here.National Bureau of Economic Research / May 2025; revised November 2025
Randomised field evidence across 66 firms and 7,137 workers separating use and time savings from broad task-composition effects.
Specific workplaces, tool and period; author disclosures include Microsoft relationships.National Bureau of Economic Research / May 2025; revised March 2026
Workplace evidence separating adoption and task reorganisation from measured earnings and hours outcomes.
Danish labour-market evidence from one period; not a product-level return-on-investment experiment.U.S. Securities and Exchange Commission / 8 February 2024
Primary regulatory evidence that hedge-fund strategies are distinct, supporting the withdrawal of a universal single-metric valuation verdict.
A reporting taxonomy, not an investment recommendation or performance study.U.S. Securities and Exchange Commission / August 2026
Primary operating, cash-flow and financing evidence supporting the distinction between demand and common-share capture.
One issuer and reporting period; no public directional security judgment follows.U.S. Securities and Exchange Commission / 21 August 2026
Primary evidence for disclosed equity-financing capacity in the demand-to-common-cash conversion chain.
Offering capacity is not evidence that the full amount will be issued or economically productive.Deep Vertical AI / 18 August 2026
The immutable verified public prior against which Edition 2026.38 is compared.
Prior published synthesis, not external evidence for the new evaluation-integrity result.
John Collins, PhD / Singapore09 / WHO IS BUILDING IT
John Collins builds and operates John Collins Alpha. Deep Vertical AI is the Singapore company through which the accumulated capability becomes research, architecture and client-owned systems.
Earlier work—investment-bank quant research, HSBC's Artificial Intelligence and Data Analytics laboratory, FTI Consulting's Asia AI practice and doctoral research— explains some of the judgment behind the system. It is context, not the proof.