§ 01 ── PRODUCT ── AUTO-GENERATIVE OPERATING SUBSTRATE
12 agents. Two classes.
One audit chain.
Auto-generated per customer in 1–3 weeks of Phase 0.
Seven customer-operational agents (Class A · drafters · T2 ceiling) and five architectural agents (Class B · stewards · T4 ceiling) operate under one customer-tier audit chain. The fabric is auto-generated from your interview answers · Brand DNA research · and our Customization Decision Tree. By Year 2 it grows to ~25 agents across 6 loops via accretion. Methodology is the IP. The substrate generates itself. The model is rented.
Per Diana Hu's YC framing of AI-native operating companies: AI as the operating system the company runs on · not a tool layered on top. Prove one high-value loop. Expand only after the system demonstrates reliability · governance · and measurable operating value.
§ 02 ── IN 60 SECONDS
What it does, end to end. Auto-generated in 1–3 weeks of Phase 0 (free).
01 · CONNECT
We ingest from your existing stack — Slack · Linear · GitHub · Notion · Zoom · Datadog · and the rest. Permission-bounded · per-category opt-in (F2 floor). Real-time on high-signal sources. No re-platforming. 20 default source integrations · custom sources added per Phase 0 interview.
02 · QUERY
Ask any question about your company in natural language. Every answer carries evidence modes (verified · inferred · model-derived · stale · contradictory · human-override) and confidence topology. We cite the source · not summarize it. Cohen's weighted κ ≥ 0.85 floor on substantive decisions.
03 · FLAG
Continuous monitoring against intended outcomes. Strategy Lead surfaces drift — "engineering trajectory diverges from stated OKR" — within hours · not retros. Diverge-and-reconcile routes load-bearing disagreement to humans.
04 · CLOSE
Class A agents translate goals into operational specs. Class B agents keep audit chain sole-emit · L8 floors operational · κ within baseline. Humans at the edge approve substantive decisions. The autonomous-action code path for money-movement · customer data · compliance · code-shipping · audit emission does not exist.
§ 03 ── THE CLOSED LOOP
Observe → Define → Plan → Execute → Measure → Reconcile → Recalibrate.
Seven steps. Every step traceable. Every action graded against reality. Loop latency under 5 minutes for standard workflows. The substrate doesn't run open-loop — outputs are continuously validated against observed outcomes · and the methodology overlay updates when calibration drifts.
FIG.01 — Seven-step closed loop · canonical agent attribution per step · RECALIBRATE → OBSERVE arc closes the loop. Per L0 doctrine + L1 framework.
FIG.01 — Seven-step closed loop · canonical agent attribution per step · RECALIBRATE → OBSERVE arc closes the loop. Per L0 doctrine + L1 framework.
RECONCILE detects drift between projected and actual outcomes. RECALIBRATE feeds back into the methodology overlay when measured drift exceeds threshold. Replay-test verifies structural integrity on every model swap. Your company's reasoning improves under load · not in spite of it.
§ 04 ── THE 12 AGENTS · TWO CLASSES
Seven Class A. Five Class B. Canonical. Kyash-validated.
Each agent occupies one tier. Each tier has a strict authority ceiling. The autonomous-action code path for substantive decisions does not exist. This is architectural · not policy. Per-customer Class A count varies 5–9 based on Customization Decision Tree adaptive flags. Class B always 5 (architectural · stage-invariant).
FIG.02 — 12 canonical agents · 2 classes · tier ceilings architecturally enforced. Class A drafts · Class B stewards substrate. Per-customer Class A count varies 5–9. Class B fixed at 5 (architectural · stage-invariant).
FIG.02 — 12 canonical agents · 2 classes · tier ceilings architecturally enforced. Class A drafts · Class B stewards substrate. Per-customer Class A count varies 5–9. Class B fixed at 5 (architectural · stage-invariant).
CLASS A · 7 · CUSTOMER-OPERATIONAL · T2 DRAFTER CEILING
Each Class A agent drafts · synthesizes · routes evidence. Humans decide. T2 ceiling: the autonomous-action code path does not exist.
#
AGENT · TIER · FLOOR
ROLE
01
Decision Brief Drafter
T2 · F3 ceiling
Drafts decision memos for substantive product / risk decisions. End-to-end synthesis (evidence · risk · customer impact · regulatory pathway · reversibility · cost) into a ready-to-judge brief. F3 advisory ceiling.
02
Risk & Compliance Lead
T2 · F3 ceiling
Per-decision regulatory pathway check. Maps every in-scope decision to applicable frameworks (GDPR · HIPAA · SOC 2 · ISO 27001 · sector-specific). Drafts compliance briefs. Never autonomously approves regulatory decisions.
03
Customer Operations Lead
T2 · F2 ceiling
Customer-facing workflows · consent-gated per F2. Drafts operational responses · synthesizes customer context · routes substantive decisions to human decider.
04
Finance Operations Lead
T2 · F1 ceiling
Variance · runway · budget · spend categorization. F1 money-movement ceiling: approval routing required.
05
Knowledge Curator
T2
Decision precedent library · regulatory pathway library · cross-loop knowledge · founder-methodology capture (consent-bounded). Surfaces precedent for novel decisions.
06
Strategy Lead
T2
Strategic synthesis · roadmap calibration · OKR amendment drafts · loop accretion candidate surfacing. Founders commit · Strategy Lead proposes.
07
Engineering Steward
T2 · F6 ceiling (F7 if unregulated)
Code-shipping advisory. PR draft + review. F6/F7 ceiling: the autonomous deploy code path does not exist. Operates the "AI software factory" pattern per Diana Hu's YC framing (next evolution of TDD): humans write specs + tests · Engineering Steward drafts implementation + iterates until tests pass. The 1000x-engineer pattern is not autonomous code-shipping; it is a substrate that surrounds an engineer with spec-test-iterate capability under audit.
CLASS B · 5 · ARCHITECTURAL / META · T4 STEWARD CEILING
Each Class B agent stewards substrate discipline. Recursively self-protected where applicable. Tier-4 architectural ceiling. Cannot be silenced by any Class A agent or human user below executive authority.
#
AGENT · TIER · FLOOR
ROLE
08
Audit Curator
T4 · F4 sole-emit
Sole-emission authority per tier. Hash-chains every entry. NEVER-deletable on severity:critical at filesystem level. F4 is L0 doctrine · always applies.
09
L8-Enforcer
T4
Continuous surveillance of all applicable L8 floor categories. Quarterly self-tests. Severity:critical on near-miss. Cannot be silenced.
10
Registry Maintainer
T4
Agent registry · capability matrix · integration source health · cross-source schema normalization. Weekly sweep.
11
Eval Suite Runner
T4
Operates canonical eval cases. Cohen's κ ≥ 0.85 verification on scoring + judgment agents. Halt authority on κ drift detection. Re-baseline annual (Methodology Council ratified).
12
Version Controller
T4
Methodology overlay version progression (v0.5 → v0.9 → v1.0 → v1.x). Pre-snapshots before any modification. Replay-tests for MAJOR cycles. Stakeholder concurrence routing.
YEAR-2 TRAJECTORY · 6-LOOP ACCRETION
Phase 1 starts with 12 agents in 1 loop. Phase 3 quarterly cadence accretes loops 2–6. Each loop adds 2–3 Class A agents. Class B stays at 5. By Year 2: ~25 agents across 6 loops · all operating in the same Y2 9-folder substrate · one audit chain · κ ≥ 0.85 maintained.
L8 HARD CAPS · ARCHITECTURAL · NOT POLICY
L8 caps on Class A agents (F1 money-movement · F2 customer data · F3 compliance/risk · F6 code-shipping) are architectural · not policy. The code path that would escalate them to autonomous decisioning does not exist in the codebase. Human judgment is preserved at the substrate level for the functions where consequences are highest. F1–F6 are canonical for any Company OS engagement. F7–F9 optional (F7 unregulated code-shipping ceiling · F8 multi-tenant isolation · F9 employee privacy) per customer pattern.
HUMAN ARCHETYPES SERVED BY THE FABRIC
The 12-agent fabric serves three human archetypes per Jack Dorsey's framing at Block: IC (builder-operator · the person who directly makes and runs things · served by Engineering Steward + Knowledge Curator + Decision Brief Drafter) · DRI (directly responsible individual · one person · one outcome · no hiding · served by Class A leads + Strategy Lead) · AI Founder (still builds · still coaches · leads by example · served by Methodology Council + Knowledge Curator's founder-methodology capture). The substrate routes information. The archetypes judge at the edge.
CANONICAL EVAL CASES · AVAILABLE UNDER NDA
Each agent has canonical eval cases that verify κ ≥ 0.85 + L8 ceiling enforcement + diverge-and-reconcile triggers + audit emission patterns. Per-agent eval case detail available in technical deep-dive under NDA — see § 10 below.
§ 05 ── TIER FRAMEWORK · HUMANS AT THE EDGE
Four tiers. Each agent occupies one.
Per Jack Dorsey at Block: humans at the edge guiding the company · intelligence layer routing information through the middle. The tier framework is how that division is architecturally enforced — each agent has a declared ceiling · and the lifting code does not exist beyond it.
FIG.03 — Authority ladder · ceiling architecturally enforced per agent. Canonical fabric is T2-heavy by design: drafters in the middle · humans at the edge · architectural stewards above. The absence of T3 OPERATOR in the canonical fabric is deliberate — substantive decision-making remains with humans at the edge.
FIG.03 — Authority ladder · ceiling architecturally enforced per agent. Canonical fabric is T2-heavy by design: drafters in the middle · humans at the edge · architectural stewards above. The absence of T3 OPERATOR in the canonical fabric is deliberate — substantive decision-making remains with humans at the edge.
TIER
NAME
AUTHORITY
EXAMPLES
T1
OBSERVER
Read-only · reports findings · cannot draft.
(none currently in canonical fabric)
T2
DRAFTER
Drafts for human review · cannot commit · advisory only.
All 7 Class A · Customer-Operational
T3
OPERATOR
Drafts AND executes within bounded scope.
(none in canonical fabric · architectural choice)
T4
STEWARD
Substrate discipline authority.
All 5 Class B · Architectural
§ 06 ── AUDIT CHAIN · F4 · L0 DOCTRINE
Append-only. Hash-chained. Sole-emit via Audit Curator. NEVER-deletable on severity:critical.
Every system action — every query · every spec · every flag · every comms draft · every L8 refusal — writes to an append-only audit chain via the Audit Curator (sole-emission authority per tier). Each entry hash-links to the prior entry. Tampering breaks continuity end-to-end and is detectable.
F4 (Audit Sole-Emit) is L0 doctrine · always applies · architecturally enforced. The direct-emission code path does not exist. No other agent — Class A or Class B — can write directly to the audit chain. Only Audit Curator. By construction.
# Example schematic: audit chain entry (emitted by Audit Curator · sole-emit per tier)
2026-W19-T0042 | agent=[CLASS_A_AGENT] | overlay_v=[N.N.N]
evidence_modes: { verified · inferred · model_derived · stale · contradictory · human_override }
confidence_band: [computed]
severity: [info | warn | critical (NEVER-deletable)]
diverge_and_reconcile: [triggered | not_triggered]
hash_prev: [SHA]
hash_self: [SHA]
emission_curator: audit_curator_org_tier_v1Every field is queryable. Every entry is replayable. Every claim carries lineage. Full schema + emission patterns available under NDA — see § 10.
κ ≥ 0.85 · COHEN'S WEIGHTED · FALSIFIABILITY FLOOR
Inter-rater agreement on substantive decisions is held to Cohen's weighted κ ≥ 0.85. Eval Suite Runner re-computes quarterly · halts on κ drift exceeding tolerance · re-baselines annually (Methodology Council ratified). This is the falsifiability floor that distinguishes substrate-grade from "AI marketing." If κ drops · Eval Suite Runner halts emissions until the underlying drift is investigated.
§ 07 ── DIVERGE-AND-RECONCILE
Load-bearing decisions are solved twice.
For substantive-and-above decision class · the substrate solves the problem twice via independent routes. Different reasoning paths · different evidence sets · different model invocations. Then compares.
If both routes agree (within threshold) · the decision proceeds with both traces logged. If they disagree · the substrate stops and surfaces — never silently picks one. Hidden disagreement is the most expensive failure mode in AI deployment.
FIG.04 — Diverge-and-reconcile · independent routes · agreement proceeds with both traces · disagreement surfaces. Methodology Council adjudicates persistent divergence.
FIG.04 — Diverge-and-reconcile · independent routes · agreement proceeds with both traces · disagreement surfaces. Methodology Council adjudicates persistent divergence.
DECISION CANDIDATE
├── ROUTE A
│ method: primary
│ evidence: source set 1
│ output: claim_A
│
├── ROUTE B
│ method: counter-evidence search
│ evidence: source set 2
│ output: claim_B
│
└── RECONCILE
├── AGREE (within δ) → proceed · both traces logged
└── DISAGREE → SURFACE TO HUMANS · never silently picked§ 08 ── EVIDENCE MODES
Trust is distributed · not scalar.
Every output carries an evidence-mode label that signals trust character. Six modes · plus reconcile-flow for contradiction.
FIG.05 — Six evidence modes · trust character per mode · [contradictory] triggers reconcile-flow.
FIG.05 — Six evidence modes · trust character per mode · [contradictory] triggers reconcile-flow.
EVIDENCE MODE
TRUST CHARACTER
EXAMPLE
[verified]
Deterministic
Specific Slack / Linear / GitHub citation inline
[inferred]
Partially-bounded
Cross-source pattern match across N events
[model-derived]
Probabilistic
LLM synthesis without specific grounding
[stale]
Time-degraded
Source older than freshness threshold
[contradictory]
Requires-resolution
Counter-evidence detected · reconcile-flow per § 07
[human-override]
Authoritative-with-audit
Explicit override + rationale logged
§ 09 ── BOUNDARIES · WHAT WE DON'T BUILD
Six categories. Architectural choice · not roadmap omission.
FIG.06 — Six anti-categories · architectural choice · not roadmap omission.
FIG.06 — Six anti-categories · architectural choice · not roadmap omission.
CATEGORY
EXAMPLE
WHY WE DON'T BUILD IT
SEARCH
Glean · Otherside
We close loops · they answer questions.
NOTE-TAKING
Notion AI · Coda Brain
We unify artifacts of every tool · not improve one tool.
CHANNEL-BOUND AI
Slack AI
We live above your stack · not inside one tool.
INTEGRATION PLATFORM
Zapier · n8n
We're the substrate that consumes integrations · not the platform that builds them.
AGENT FRAMEWORK
LangChain · LlamaIndex
We run agents under governance · not the framework that builds them.
DASHBOARD
(any BI tool)
We close loops · not report.
§ 10 ── UNDER THE HOOD · NDA-TIER
Substrate-grade depth available to qualified prospects.
What's public on this page: the architecture · the 12 agents · the tier framework · the audit chain · the evidence modes · the boundaries. What's available in a 30-min technical deep-dive under NDA:
Customization Decision Tree
how interview answers route to per-customer agent fabric + L8 floor selection + stage-adapted KPI calibration.
13-step generation protocol
the exact sequence that produces a complete L2 substrate in 1–3 weeks of Phase 0.
Canonical eval cases per agent
the κ verification tests + L8 ceiling pressure tests + diverge-and-reconcile triggers.
Methodology Council Workshop
the 7-seat composition + κ baseline + monthly + quarterly + annual cadence playbook.
D7 §12 inspection protocol
the 5-step customer-invokable inspection drill (Inspect · Replay · Verify · Cross-check · Escalate).
Pre-flight validation toolkit
operator-side substrate discipline tooling.
Compliance Mapping Library
regulatory framework → L8 floor + audit chain tethering · per-jurisdiction (GDPR · HIPAA · SOC 2 · ISO 27001 · PSA · APPI · JFSA · FCA · sector-specific).
Replay-test infrastructure
structural fingerprint methodology for model swap survival.
Phase 0 → Phase 1 → Phase 2 → Phase 3 cycle mechanics
pricing · falsifiable KPI gate at Week 7–9 · monthly cancellation post-gate-PASS.
Sophisticated buyers · YC partners · M&A diligence teams · procurement teams · all welcome. The substrate is auditor-ready any time per D7 §12 protocol.
§ END ── DEEPER
If the architecture interests you · the doctrine is at /substrate. If the engagement model interests you · the four phases are at /engagement. If the verification layer interests you · the security + compliance posture is at /security. If you want to see this generate a substrate for your company · book Phase 0 (free · 1–3 weeks · no obligation) at /contact.