docs: ecosystem deck, grounded in verified local state (cloud-session takeover)

docs/deck/01-github-profile-analysis.md

@@ -0,0 +1,119 @@
+# GitHub profile analysis -- HarperZ9 / ZentropyLabs
+
+> Census taken 2026-07-30 from the live account listing. Counts are exact at
+> that timestamp. Companion files: 02-vertical-value.md, 03-target-audiences.md,
+> 04-property-assessment.md. Slide deck: flywheel-telos-deck.html.
+
+## Headline numbers
+
+| Measure | Count | Note |
+|---|---|---|
+| Total repositories | 159 | 158 under HarperZ9 + ZentropyLabs-ai/.github |
+| Original (non-fork) | 104 | the portfolio proper |
+| Forks | 55 | upstream OSS + curated awesome-lists + eval/agent tooling |
+| Private | 37 | security line, dev monorepos, site experiments |
+| Public | 122 | |
+
+## The five-tier taxonomy (per the site's own catalog)
+
+### Tier 1 -- the fourteen public engines
+flywheel (route + verify; desktop client v0.2.2), telos (the shared workbench,
+live v0.2.0), index (2.9.0 beta, PyPI `index-graph`), gather (1.6.1 shipped,
+PyPI `gather-engine`), forum (1.13.0 shipped, PyPI `forum-engine`), crucible
+(1.2.0, PyPI `crucible-bench`), emet (1.1.0, PyPI `emet`, four independent
+language implementations), buildlang (v1.2.0, 1,605 passing tests), learn
+(v1.6.0), relay, plexus, mneme, studio-engine, build-color (1.0.2 beta,
+PyPI `build-color`).
+
+Six confirmed PyPI packages -- the site publicly corrected an earlier "19"
+over-count to this honest figure.
+
+### Tier 2 -- the private line ("public where safe")
+- **orca** (Runtime) -- security-engagement workbench, v1.0.0, 361 tests;
+  metadata-only, cannot reach the target it assesses. The most revenue-shaped
+  private asset.
+- **aleph** (Checkpoint) -- release/supply-chain coordination across docs,
+  contracts, MCP declarations, CI, receipts.
+- **kun** (Vault) -- path-only receipts, rotation discipline, value-free.
+- **behavior-transform.io** (Boundary) -- public repo; mode-aware
+  read/write/exec/fetch/input receipts; raw content stays in local adapters.
+- **seed** (Lab) and **sofer** (Ledger) -- private/proprietary; honestly labeled
+  as needing a public-safe split (credential-pattern corpora inside).
+- Adjacent private security cluster: offensive-platform, red-team-platform,
+  adversarial, protected-corridors, protected-research, legal-intl,
+  warden-reporting / warden-clp / warden-prefire-primitive (warden as a public
+  brand is retired into proof-surface), emet-internal, met-monorepo,
+  ai-safety-prefire, ai-safety-guardrail-manager, unified-engine, aurora,
+  apps, flywheel-host.
+
+### Tier 3 -- substrates
+proof-surface (652 tests, zero deps -- the shared proof-packet foundation),
+coherence-membrane, accountable-surface, accountable-engine (Forming; "the
+inward half is owed"), provenance-sensorium, reconcile, raw (D3D11 mid-frame
+graphics proxy), witnessing-spine (five regulated-sector adversarial
+steelmans, DOI 10.5281/zenodo.20778927), senses-and-sensibility,
+faithful-transpile, coherence and kernels lines (signal-kernels,
+anomaly-kernels -- C++23 header-only).
+
+### Tier 4 -- bricks (release + agent toolkit)
+secret-redact-io, release-surface-scanner, public-surface-sweeper,
+repo-proof-index, proof-surface-report, model-provenance-validator,
+gpu-trace-validator, agent-audit, agent-routing-kit, agent-hook-pack,
+context-curator-lite, workflow-harness-lite, consulting-template-kit,
+calibrate-pro, wol-pi.
+
+### Tier 5 -- research + receipts-in-the-wild
+Receipt repos aimed at other ecosystems: crewai-kickoff-replay-receipt,
+vllm-ci-forum-receipt. Research surface lives on the site: 3 theses,
+10 experiment packets, 6 DOI'd preprints (arXiv endorsement-pending),
+with a formal claim-status ladder (SOURCE_LEAD → … → LAW_CANDIDATE, plus
+UNVERIFIABLE).
+
+## The build/creative lines
+
+- **Build product line** (former "Quanta", renamed): build-engine, build-color,
+  build-oracle, build-ui, build-finance, build-universe, build-ecosystem,
+  buildlang + buildlang-vscode + buildlang-tmLanguage.
+- **Graphics/ENB line** (the consumer-traction record): elder-enb (900k+
+  downloads, 150k+ unique users, ~280 releases over ~2 years), truth-enb,
+  elder-weathers, enb-runtime-core, enb-validation-lab, enb-bridge-f4 (private),
+  brender-archival, skyrimbridge, skse/skse-playground (private), gaussian
+  splats (site page), engine-revival, studio-engine.
+- **Site/property repos**: HarperZ9.github.io (deployed output; canonical
+  source telos-v2, private), harpercompliance-site (private),
+  sendmyletter-site (private), freelance (private).
+
+## The fork belt (what it signals)
+
+55 forks cluster into: agent/eval infrastructure (pydantic-ai, deepeval,
+langfuse, inspector, agent-runtime-observatory, SmallHarness, runtrail,
+scorekeeper), MCP ecosystem curation (a dozen awesome-mcp-* lists, registry,
+python-sdk), core Python tooling (numpy, poetry, setuptools, isort, tomlkit,
+typeguard, grimp, datasette, llm), and agent-security reading
+(system_prompts_leaks, awesome-ai-agents-security). Read as a map: the
+account tracks exactly the frontier its products verify -- agents, evals,
+MCP, and the supply chain under them.
+
+## Engine-state cross-check (dev vs public)
+
+| Surface | Routes / lanes | Version |
+|---|---|---|
+| `flywheel` public repo gateway | 11 API routes | 0.1.0 |
+| `local-model` dev engine gateway | 96 API routes | dev (CC-1 era) |
+| Lane registry (identical in both) | 7 lanes | gather 1.6.1 · crucible 1.2.0 · index 2.9.0 · forum 1.13.0 · learn 1.6.0 · telos 0.2.0 · local-model bundled |
+| Marketplace | 17 built-in MCP connectors + unlimited user-registered plugins | |
+| SUPERPROJECT roster | 11 mission-tier flagships (5 live MCP spine + 6 declared) | curated 2026-07-06 |
+
+The freshest engineering proof point is CC-1 "Certified Commons"
+(STATE.md 2026-07-25): `flywheel gate` reaches rewitness=MATCH from a fresh
+clone with byte-identical digests; shown able to FAIL first (verdict tamper →
+DRIFT); 85 + 243 passing tests on that slice; CI across three OSes.
+
+## Reading the whole profile in one line
+
+One person, one discipline -- perceive, transform under a named criterion,
+carry a re-checkable receipt -- applied across model infrastructure, compilers,
+color science, graphics, security practice, and research, with the honesty
+mechanics (published nulls, self-caught over-counts, maturity tiers) intact
+everywhere. The portfolio's weakness is not depth or range; it is that no
+single surface assembles this into an offer a buyer can say yes to.

docs/deck/02-vertical-value.md

@@ -0,0 +1,82 @@
+# Vertical value assessment -- two industries, each tool alone and as the system
+
+> Companion to 01-github-profile-analysis.md. The claim discipline of the
+> portfolio applies to this document: where a value statement is a projection
+> rather than an observed result, it says so.
+
+## The shared premise (why either vertical pays)
+
+Machines made claiming nearly free; checking stayed expensive. Both verticals
+below are drowning in AI-produced output whose acceptance currently rests on
+confidence. The system's one operation -- propose, verify against an external
+criterion, keep a re-checkable receipt -- is the missing accept path. The
+pitch deck already names the wedge categories (originality-verification,
+accountable-agent-action, provenance) and is honest that they have no analyst
+TAM yet: that is the whitespace, and also the risk.
+
+## Vertical A -- sales & marketing (greenfield)
+
+Grep-verified: the current site has zero sales/marketing surface area. No
+positioning debt, no incumbent framing to unwind -- and no existing proof of
+demand either. Everything below is a mapping of shipped capabilities onto
+known agency/martech pains, not observed traction.
+
+| Asset | Alone, it is | In this vertical it becomes |
+|---|---|---|
+| **gather** + snapshot (citation freezing) | research intake with provenance receipts | Claims substantiation: every public sentence traceable to a frozen, hashed source that outlives the live web. Ad-claim defense files that assemble themselves. |
+| **forum** | agent orchestration on a replayable causal hash-chain ledger with human gates | Agency accountability: campaign automation the client can re-check instead of trusting a report. A deliverable format no incumbent offers. |
+| **flywheel** routing + receipts | one request shape over every provider, observed per-provider scoreboard | AI-spend governance: cost-honest routing, escalation receipts, the first "what did the model spend buy us" ledger. |
+| **studio-engine** + telos creative | seeded, replayable generative worlds/brand kits | Reproducible brand systems: same seed, same artifact -- campaign creative that is re-derivable, auditable, and provably original-to-seed. |
+| **learn** | witnessed courses; every graded step writes a receipt | Enablement proof: onboarding and certification that produce receipts, not completion checkboxes. |
+| **mneme** | agent memory where every memory carries provenance | Account knowledge that survives turnover: "who told us this, when, from which document" as a queryable property. |
+| **crucible** | falsifiable claim verification | Marketing-claims QA and competitor-claim testing: register the claim, measure it, publish the verdict packet. |
+
+**As the system**: an agency or in-house team runs intake (gather) → brief
+verification (crucible) → campaign orchestration (forum) → creative (studio)
+→ delivery with receipts (flywheel/desktop), and hands the client one chained
+record. The sale is not "AI tools"; it is *defensible delivery*.
+
+**Honest boundaries**: no marketing-specific integrations exist today (no CRM,
+no ad-platform connectors -- though 17 built-in MCP connectors plus the plugin
+registry are the extension path). First proof would need one design-partner
+agency.
+
+## Vertical B -- provenance-critical, privacy-bound industries (assembled, not built)
+
+The material for this vertical already exists across seven site pages and
+several repos; it has never been assembled into one buyer-facing argument.
+
+| Sector | Live requirement | Shipped answer |
+|---|---|---|
+| **Compliance / model risk (finance)** | Regulators require model origin and lineage to be auditable; incumbent platforms record lineage as editable database rows (site: writing.html; witnessing-spine, DOI'd, steelmans five financial-sector cases). | Merkle-logged receipts with offline inclusion proofs; hash-chained verifiable store with audit tail; attestation binding sign-off to exactly what was reviewed, overclaims tracked. |
+| **Legal / e-discovery** | "The summary is not the record." AI summaries entering matters without their sources. | Citation freezing (page bytes fetched, hashed, stored), content-addressed corpora, comprehension gating (a teach-back that pasting the diff cannot pass). |
+| **Healthcare administration** | Summaries must keep uncertainty attached (field-guide clinical lane). | Honest nulls as first-class UI; UNVERIFIABLE rendered as loudly as a win; escalation with the failed local attempt on record. |
+| **Newsrooms / media** | Every public sentence must find its source (field-guide media lane). | gather → forum → crucible chain; per-block source hashes; tamper caught on re-read (demonstrated in the recorded demos). |
+| **Security / red teams (gov, law enforcement, AI labs)** | Authorized adversarial work needs records that hold up afterward, and a lawful-basis gate. | The existing security.html practice + ORCA (v1.0.0, 361 tests, metadata-only engagement workbench). Already revenue-shaped. |
+| **Regulated AI operations** | Agent boundaries must be provable, credentials must not travel. | Boundary receipts (raw content never leaves local adapters), capability grants per run, presence-only credentials in the OS keychain, zero telemetry, fully local loop. |
+
+**The privacy story is architectural, not contractual** -- this is the
+differentiator dense-workflow industries actually buy:
+- The desktop client's only base URL is 127.0.0.1:8799; no outbound endpoint
+  exists in the codebase.
+- Credential *values* never enter the app, the gateway refuses to accept them
+  through the marketplace, and the UI shows presence + source only.
+- Proofs verify offline: a third party recomputes the Merkle root from the
+  leaf with no network, no account, no vendor.
+- Write/exec are grants per run; a verify stage without an exec grant reports
+  UNVERIFIABLE rather than pretending.
+
+**As the system**: the same loop, pointed at regulated work -- intake with
+receipts, judgment that fails closed, orchestration with human gates, sign-off
+bound to coverage, and an audit trail that a regulator, opposing counsel, or
+incident reviewer can re-derive without trusting the operator.
+
+## Value of each tool alone vs. the system
+
+Each lane is deliberately "a full product that also runs alone" -- gather,
+crucible, index, forum, learn are individually pip-installable with their own
+receipts. Alone, each competes in a crowded category on speed/cost and honesty
+(the battle map publishes THEY_LEAD rows). Together they occupy compositions
+with no incumbent: benchmark-claim escrow, accountable-agent-action,
+receipt-carrying creative, the concurrent-desktop agent. The system premium is
+the moat; the standalone tools are the doors in.

docs/deck/03-target-audiences.md

@@ -0,0 +1,60 @@
+# Target audiences -- where, how, and why these tools matter on today's frontier
+
+> The frontier condition in one line: autonomous systems now produce work
+> faster than any review lane absorbs it, provenance duties are hardening into
+> regulation, and public AI claims go unchecked by default. Every audience
+> below is someone for whom one of those three pressures is already a budget
+> line or a liability.
+
+## The three frontier pressures → the system's three answers
+
+| Pressure (where the frontier bites) | The system's answer |
+|---|---|
+| **Agents outran review** -- autonomous work lands faster than humans check it | The verified accept path: an external oracle decides; models propose, oracles dispose; UNVERIFIABLE is a legal, visible outcome |
+| **Provenance became law** -- lineage/origin duties moving into regulation across finance, health, media | Receipts as infrastructure: content-addressed, Merkle-proven, offline-verifiable, vendor-independent |
+| **Claiming became free** -- capabilities, benchmarks, and originality asserted daily, checked never | Claim escrow: sealed verdict packets a stranger, competitor, or court can recompute |
+
+## Audience matrix
+
+| Who | Where they live | Why now | How they enter | Value to them | Value to us |
+|---|---|---|---|---|---|
+| **Agency ops lead** | Campaign delivery, client reporting | Clients audit AI-made work; agencies have no proof layer | forum demo: the replayable run ledger | Defensible delivery -- the client re-checks instead of trusts | First design partner in the greenfield vertical |
+| **CMO / martech owner** | Brand, content, AI-spend budgets | AI spend unmeasured; claims risk rising | flywheel routing scoreboard + gather claim receipts | Cost + substantiation in one surface | The paying wedge account for vertical A |
+| **GRC / model-risk officer** | Banks, insurers, model inventories | Lineage regulation arrived; current audit trails are editable rows | Witnessing Spine findings + a receipts inclusion-proof demo | Tamper-evident lineage a regulator can re-derive | The regulated flagship reference |
+| **Legal ops / e-discovery lead** | Firms, corporate legal departments | AI summaries entering the record without their sources | snapshot (citation freezing) + attestation | Summaries bound to frozen records; sign-off bound to coverage | The dense-workflow proving ground |
+| **Newsroom standards editor** | Media, fact-check desks | Generated content floods intake; provenance IS the product | field-guide media lane: gather → forum → crucible | Source-checked publishing with per-block hashes | A public credibility case study |
+| **Red team / security lead** | Government, law enforcement, AI labs | AI attack-surface work needs lawful, replayable records | ORCA + the security.html engagement gate | Engagement records that hold up afterward | The already-revenue-shaped practice |
+| **Clinical admin / health-ops** | Hospital administration, payer ops | Summaries must keep uncertainty attached | field-guide clinical lane (gather + forum) | Honest-null summaries; UNVERIFIABLE stays visible | Entry into the highest-stakes provenance market |
+| **Builders & researchers** | OSS, eval/agent infrastructure | Already served -- the audience the site speaks to today | PyPI packages, recorded demos, DOI'd papers | Tools + receipts + a falsifiable research program | Contributors, field testers, standing -- not revenue |
+
+## Where each audience is reached
+
+- **Agencies / CMOs**: martech and agency-ops communities; the deck's Act II-A
+  is the first artifact that speaks to them -- nothing on the site does yet.
+- **GRC / model risk**: the Witnessing Spine paper is citable; model-risk and
+  AI-governance forums; the "editable rows vs. Merkle proof" demo is the hook.
+- **Legal**: e-discovery and legal-ops conferences; the citation-freezing demo
+  is self-explanatory to anyone who has fought over a dead link in a record.
+- **Newsrooms**: the field guide's media lane is already written as a script
+  with a break test -- publishable as-is.
+- **Security**: the practice exists with a lawful-basis gate; the audience is
+  reached by referral and credential, not marketing.
+- **Builders**: already arriving via PyPI, the demos, and the papers.
+
+## The honest split
+
+The audience that already responds (builders, researchers) is not the audience
+that pays. The audiences that would pay (compliance, legal, agencies, regulated
+ops) are named nowhere on the current public surface except behind the
+security page's authorization gate. Closing that gap is a positioning task,
+not an engineering one -- the evidence is already written, scattered across
+seven pages.
+
+## Why this matters to each side (the deck's dual promise)
+
+For the owner: each tool is independently sellable (pip-installable, own
+receipts, own demo), so every audience above is reachable without waiting for
+platform adoption. For the audience: the system never asks to be trusted --
+every claim it makes about their work arrives with the means to re-check it,
+which is precisely the property their regulators, clients, editors, and
+courts are beginning to demand.