file PR-2026 · portfolio of record
Philip Roy
co-founder · Dicta
BC, Canada · UTC−7
systems in production since february 2026
I build and operate AI systems inside real businesses: agent pipelines that ship client work around the clock, memory infrastructure measured across months of daily use, and the instruments that say what actually works. Every number on this page carries its source.
34,782
memory entries in production
SOURCE: live store, queried 2026-06-10
3,300+
logged agent sessions, one deployment
SOURCE: live store, June 2026 · includes scheduled automation runs
80→100%
retrieval precision@3
SOURCE: 4 repair rounds; a 40% regression reported en route
112
research cycles, hypothesis to verdict
SOURCE: Deep Applied Research live store · 18 focus areas
file 01 — systems
Not demos. Systems with operators, paying clients, failure logs, and uptime to answer for.
A production pipeline that drafts, reviews, and ships voice-matched content for a real client roster 24/7: dispatcher, drafter, a blind cold-read agent, a five-dimension review gate, and bounded auto-revision with named repair strategies. Every gate flag becomes a learning signal that must be dispositioned before the next batch runs.
3-stage QC chain · append-only signal ledger · 22 major prompt versions
Memory infrastructure for LLM agents that goes past fact retrieval: verbatim knowledge graph, behavioral drift tracking, experiential encodings, and a personality layer measured across hundreds of sessions. Persona trait alignment improved from 2.25 to 4.43 out of 5 as the identity graph accumulated evidence.
34,782 entries live · P@3 80→100% · trait alignment 2.25→4.43
A hypothesis-and-verdict protocol that wraps any research agent in structured cycles with cross-session memory. Its defining feature is self-audit: it measured a 77% adopt-rate bias across 96 of its own reports, shipped corrections, and re-measured. The audit trail is the product.
1.9k LOC · ~250 passing tests · bias caught on the record
file 02 — instruments
Instruments built to survive scrutiny, including the author's own results.
2.25 → 4.43 / 5
Five trait scenarios scored before and after identity-graph accumulation, with the largest gains exactly where the base model's compliance gravity had to be overridden.
SOURCE: From Memory to Partnership, Track 2
33 categories
Surface, structural, and interactional failure modes of AI text and behavior, with detection heuristics per category. Published at 22; production use grew it to 33. Runs daily as a live gate.
SOURCE: From Memory to Partnership, Appendix A; expanded in production
First runnable implementation of the NCT (Natangelo 2025): 25 behavioral cases across five axes of agent identity. The harness verifies that seeded facts actually reached the model's context and runs a bare-model baseline arm, so every point in a score is attributed: retrieval, behavior, or base-model credit. The instrument is the contribution; it measures its own author without flinching.
v1 run, June 2026 · full results and attribution ship with the benchmark
file 03 — papers
Longitudinal evidence from a deployment most labs can't run: one dyad, four months, every session logged.
Systems paper: how evolving persistent context transforms human-AI interaction. Four evaluation tracks over 400+ sessions, a triple-confound caveat stated up front, and a 40% retrieval regression reported instead of buried.
Read the paper (PDF)Position paper: why the AGI threshold is the wrong target. Intelligence as a property of the human-AI dyad, with falsification criteria and the honest limits of a single-dyad case study.
Read the paper (PDF)A case against compaction: rolling eviction plus structured offload as the alternative to ever-larger context windows. Testable, partially validated, openly unfinished.
file 04 — method
Disclosed because it's the point, and increasingly the job.
“This paper is, in a sense, its own evidence: an artifact of the partnership it argues for.”
acknowledgments, From Memory to Partnership (2026)
Human direction, AI implementation. Every system here was designed and operated by me, built in disciplined partnership with the AI it studies. The discipline structure runs to hundreds of documented rules, each one traced to a caught failure.
QC gates in both directions. The AI's output passes through independent review agents before it ships. My claims pass through the same standard: a result gets published with its baseline, its scorer repairs, and its regressions.
Receipts over adjectives. Where a number appears on this page, the artifact behind it exists: result files with provenance fields, append-only signal ledgers, papers with the confounds stated in the abstract.