State of the Framework • Internal Audit

The Knowledge Base, the Instrument, and the Three Lines of Work

ShurIQ’s state of the framework, audited 2026-05-26. Lines 1 and 2 shipped 2026-05-15.

Internal May 2026 Author: Jonny Dubowsky Shur Creative Partners

What We Set Out to Build I

What We’ve Built II

When	Move
2026-03-21	First version of the scoring vocabulary. Knowledge base running locally for the first time.
2026-03-27	Billion-fact thesis formalized in the investor pitch.
2026-03-28	Audit-trail vocabulary added — every score now traces back to its sources.
2026-04-02	First week’s prediction validation loaded into the knowledge base.
2026-05-02	System architecture iteration. Account / project / report hierarchy designed.
2026-05-11	Data model v0.7. Schema migration added accounts, projects, business-model-canvas tables.
2026-05-13–14	Brand dashboard pilot live. ReelShort shipped; AHA and Reffer queued.
2026-05-15	Lines 1 and 2 shipped. Micro-drama ingested across 10 weeks; λ measurable for the first time.

What Is Working as Designed III

The Three Lines of Work IV

Line 2Shipped 2026-05-15

Track λ and surface decay in reporting

Advances variable λ directly. Made the Lyapunov dashboard layer load-bearing.

The problem

We named λ as one of our four thesis variables, but we had no way to measure it. Without multi-week history in the knowledge base, the calculation was not possible. The pitch deck and the editorial briefs all referenced λ as if it existed. It did not.

What we did

With 10 weeks of micro-drama scoring now in the knowledge base, we wrote code that looks at each company’s score in week N and its score in week N+1, computes how much it moved, and stores that as one DecayMetric record. We do this for every pair of consecutive weeks we have. We then aggregate the DecayMetric records into a per-week LambdaSnapshot — the median, mean, and 90th-percentile of how much the whole cohort moved that week.

For micro-drama this produced 180 DecayMetric records and 9 LambdaSnapshot records, derived from the source data and stored in their own labeled section of the knowledge base so we can recompute them anytime without touching the original scores.

The λ trajectory across 9 weeks

Week	Median λ	What was happening
W11-2026	0.90	Reporting cadence still ramping; cohort growing from 16 to 17.
W12-2026	1.85	Peak volatility. New entrants, calibration in progress.
W13-2026	1.25	Mid-volatility as the cohort filled out to 21 companies.
W14-2026	0.60	Cohort settling.
W15-2026	0.30	Continued settling.
W16-2026	0.20	Continued settling.
W17-2026	0.00	Structural floor — no movement in the median.
W18-2026	0.00	Continued at the floor.
W19-2026	0.10	Nudges up on Amazon Prime Video Clips deploy.

Top movers, W10 → W19

Company	Net composite change
JioHotstar	+16.80
COL Group / BeLive	+14.45
GoodShort	+6.40
Disney	+6.20
DramaBox	+6.05
Lifetime / A+E	+5.90
Amazon	+3.55

These are the brands whose structural position moved the most across the tracked period.

Why it matters

λ is no longer an unmeasured promise. We can quote it in the pitch deck. We can use it to set refresh cadences per vertical (high λ verticals need nightly updates; low λ verticals can hold for a week or more). The Lyapunov layer in the customer dashboard — the part that frames a brand’s “structural basin” and “shock response” — can now read from this derived data instead of from numbers Jonny enters by hand.

What is left

The brand dashboard’s Lyapunov layer still reads hand-curated inputs. The DecayMetric pipeline produces the principled equivalents. Wiring is a small follow-up.

We also have not yet generated λ for the AI-agent vertical. That vertical has 1,672 companies but only one week of scores. We need at least two consecutive weeks to compute λ, and at least four for it to be reliable. The work: run the auto-research cycle for W13, W14, W15 and onward. The methodology is locked; the pipeline exists. What is missing is execution.

Output: λ goes from a guess to a measured rate. The Lyapunov layer becomes load-bearing.

Naming Discipline V

Where We Are on the 1B Path VI

Variable	Status	Line that advances it
D — Node Density	105,738 of 1,000,000,000 target across the knowledge base (0.011%). Grew 39% since the investor deck.	Line 1 (shipped)
A — Ontological Alpha	Established for AI-agent and micro-drama verticals.	Line 1 (shipped)
E — Extraction Efficiency	Modeled, not running as a standing job.	Standing auto-research
λ — Decay Rate	Measured for micro-drama. Median \|Δ\| settled 1.85 (W12) → 0.10 (W19). Top movers: JioHotstar +16.80, COL Group +14.45.	Line 2 (shipped)

Appendix

For internal eyes

Methodology, schema state, and open questions. This section is intentionally more technical — use the explainer guide to translate.

Knowledge base summary — AI-agent vertical, live state 2026-05-26

Thirteen OWL classes under shurai.com/ontology/ai-agent-sbpi#:

AgentCompany (1,672 instances)
ScoreRecord (1,672 instances)
DimensionScore (8,360 instances)
Attestation (1,672 instances)
Dimension (5 instances — Model Capability, Market Traction, Platform Ecosystem, Autonomy Depth, Capital & Defensibility)
Tier (5 instances)
DomainCategory (12 instances)
AIAgentVertical, plus supporting classes

ScoreRecord predicates: forCompany, forWeek, inTier, hasDimensionScore, hasAttestation, compositeScore, delta.

DimensionScore predicates: forDimension, dimensionValue.

Attestation predicates: confidence, sourceType, plus PROV-O wiring (prov:wasGeneratedBy, prov:Entity, prov:Activity).

Dimension averages — W12-2026, n=1,672

Dimension	Avg	Min	Max
Model Capability	47.61	40	90
Market Traction	46.60	25	90
Autonomy Depth	44.49	25	100
Capital & Defensibility	43.39	40	75
Platform Ecosystem	39.92	30	95

Platform Ecosystem trails by 7.7 points — most YC AI agent companies are individual products, not ecosystem plays. Model Capability tops, consistent with self-reported framing in YC application data.

Top composite scores — W12-2026

Rank	Company	Composite
1	Persana AI	69
2	Fiber AI	68
3	Rescale	66
3	Warmly	66
5	Fini	65
6	Tenyks	64.75
7	Athelas	64.5
7	careCycle	64.5
7	Mutiny	64.5
10	MindsDB	63.5

Score range: 25 to 69. No company exceeds 70 — the rubric was designed with headroom.

DecayMetric class — Line 2, now shipped

sbpi:DecayMetric a owl:Class ;
    rdfs:subClassOf prov:Entity ;
    rdfs:label "Decay Metric" ;
    rdfs:comment "Cross-week delta between two ScoreRecords for the same company. Computed at ingest." .

sbpi:forFromWeek    rdfs:domain sbpi:DecayMetric ; rdfs:range sbpi:Week .
sbpi:forToWeek      rdfs:domain sbpi:DecayMetric ; rdfs:range sbpi:Week .
sbpi:forCompany     rdfs:domain sbpi:DecayMetric ; rdfs:range sbpi:Company .
sbpi:compositeDelta rdfs:domain sbpi:DecayMetric ; rdfs:range xsd:decimal .
sbpi:decayRate      rdfs:domain sbpi:DecayMetric ; rdfs:range xsd:decimal .  # annualized
sbpi:volatility     rdfs:domain sbpi:DecayMetric ; rdfs:range xsd:decimal .  # rolling 4-week std-dev

Named graph for derived decay data: shurai.com/graphs/derived/decay/{vertical}. Source scoring data lives in shurai.com/graphs/{vertical}/{week}.

Architectural decisions worth re-examining

Oxigraph (2026-03-21 initial selection) versus GraphDB (current running instance) — switch made without written rationale. Working stack stays unless scale forces a change.
Discourse grammar (Claim, Evidence, Source plus supports, opposes, informs, triggers, predicts, annotates, grounds, supersedes, createdBy) is agent-side scaffolding. Joinable to the AgentCompany scoring graph but separately tracked. Stays internal.

Open questions for Line 1

Does the auto-research cycle deposit into the same named graph as client engagements, or into an auto: named graph for hygiene?
The dashboard renderer reads JSON. The nightly export job runs where — local cron, GitHub Action, or a Cloudflare Worker triggered cron?
Is the historical W11–W19 micro-drama data clean enough for the JSON-replace gate, or does the export need a per-week review step before publishing?

Open questions for Line 2

Volatility window — 4 weeks was the MVP choice. 8 or 12 weeks would smooth more.
Annualization formula — currently delta × 52. If decay turns out to be multiplicative rather than additive, this needs revisiting.
Per-dimension λ or composite-only? Currently composite-only. Per-dimension is more diagnostic.

The Knowledge Base, the Instrument, and the Three Lines of Work

ShurIQ is an intelligence instrument grounded in a knowledge base.

Three pieces of the instrument are live.

Chronology — how the framework came together

Two pieces look like gaps at first read and are actually design choices.

Three concrete lines close the gap between thesis and current state. Two shipped on 2026-05-15.

Make the knowledge base the canonical source for SBPI scoring

Track λ and surface decay in reporting

Name the collective-intelligence framing in the canonical narrative

Use SBPI exclusively.

Three of the four variables hinge on the same underlying operation.