Flywheel: the whole workflow, every tool its own product.
+One local platform that routes any model, verifies what it can, + and hands back a receipt a stranger can re-run -- presented tool by tool, then as the system.
+Prepared 2026-07-30 · every count traceable · ← → to move
+
+
+ 02 · The frontier problem
+"Machines made claiming nearly free.
Checking stayed expensive."
Generative systems now produce plans, code, summaries, campaigns, and citations faster than + any human review lane can absorb. Every industry is accumulating review debt: output + accepted on confidence, not on evidence.
+The frontier question has flipped. It is no longer "can the model do it?" -- + it is "who checks, against what, and can anyone re-check later?"
+"Ask for the witness, not the confidence."
+The verdict vocabulary
+MATCH re-derived, byte for byte
+DRIFT the record moved -- caution, honestly shown
+UNVERIFIABLE cannot be checked -- never rounded up to "fine"
+Never TRUSTED, by construction. The honest null is a first-class result.
+Source · no-receipt-no-accept release essay · presentation.html · CREDO.md
+03 · The practice -- range first, rigor as the floor
+Map the system → make the surface → test the claim → leave a usable artifact.
+One practice across model workflows, compilers, color science, real-time + graphics, research infrastructure, and clear writing -- verification is one strong lane inside it, not the cage around it.
+Source · list_repos census 2026-07-30 · catalog.html · toolkit.html · cover-letter.html
+04 · The system, in five honest tiers
+Fourteen engines lead. Four more tiers stand behind them.
+| Tier | What lives there | Standing |
|---|---|---|
| Public engines ×14 | flywheel · telos · index · gather · forum · crucible · emet · buildlang · learn · relay · plexus · mneme · studio-engine · build color | SHIPPED / LIVE / BETA |
| Private line ×6 | Checkpoint (Aleph) · Runtime (ORCA) · Vault (Kun) · Boundary (behavior-transform) · Lab (Seed) · Ledger (Sofer) | PUBLIC WHERE SAFE |
| Substrates | proof-surface (652 tests) · coherence-membrane · accountable-surface / -engine · provenance-sensorium · reconcile · RAW | SHIPPED FORMING |
| Bricks ×12+ | release toolkit: secret-redact-io, repo-proof-index, model-provenance-validator, agent-routing-kit, signal/anomaly-kernels … | PUBLIC |
| Research | 3 theses · 10 experiment packets · 6 DOI'd preprints · a formal claim-status ladder ending in UNVERIFIABLE | EVIDENCE-LED |
Maturity labels are kept on every slide -- the roster is honest about tier, because that is the product.
+Source · catalog.html · overview.html · SUPERPROJECT.md
+05 · One loop, every organ
+Each run leaves the next a better map, corpus, plan, verdict, route, or memory.
+Intake carries a provenance receipt. Judgment is a pure function that fails closed. Orchestration writes a + replayable causal ledger. The map carries file-and-line evidence. Memory carries its own provenance. + Creation is seeded and reproducible. Routing shows observed outcomes, not promises.
+The loop is closed: flywheel loop-status → loop closure 9/9 (100%). + The model is the replaceable half; the loop is the durable one. Proven live 2026-07-29: a local model proposed through a sealed boundary, a model-free oracled kernel disposed, and one receipt chain bound both, each member re-verified.
+Source · flywheel README · SUPERPROJECT.md (spine verified by test_superproject.py)
+06 · Each tool alone -- the working spine
+| Engine | In its own words | Standing |
|---|---|---|
| flywheel | "A companion for every model." Routes local or hosted, verifies what it can, escalates only the hard part, hands back a re-checkable receipt. | PUBLIC · DESKTOP v0.2.2 |
| gather | "Reach it anywhere. Record how you got it." Guarded APIs, JS-walled pages, scans, audio, papers -- every block carries a source hash. | SHIPPED 1.6.1 · PyPI |
| crucible | Register a thesis, steelman it, measure it against a substrate, track drift, refine the weakest axis. Verdicts fail closed. | VERSIONED 1.2.0 · PyPI |
| forum | An agent team with routing, quality checks, and a replayable causal hash-chain ledger. Human approval gates. | SHIPPED 1.13.0 · PyPI |
| index | Repo-inventory + dependency graph for many-repo workspaces; every edge carries file-and-line evidence. Fully offline. | BETA 2.9.0 · PyPI |
| emet | Re-derives bytes; reports MATCH / DRIFT / UNVERIFIABLE -- "never TRUSTED, by construction." Four independent language implementations. | SHIPPED 1.1.0 · PyPI |
| learn | Turns your own material into a witnessed course; every graded step writes a receipt. | SHIPPED v1.6.0 |
Source · catalog.html per-engine pages · lanes.py registry (versions cross-checked)
+07 · Each tool alone -- the range
+| Engine | In its own words | Standing |
|---|---|---|
| telos | The shared workbench: a person and a model shaping the same artifact; one MCP surface over the flagship family. | LIVE · v0.2.0 |
| buildlang | "What a function may do belongs in its type." Effects + lifetimes checked, lowered through a C backend; one sealed receipt schema now spans five computation modes and binds them into a tamper-evident chain. | SHIPPED v1.2.0 · 1,605 TESTS |
| relay | Zero-dependency coding agent for any endpoint: failover, gated tool loop, witnessed ledger. | PUBLIC |
| plexus | Capability discovery + auto-wiring: reads what each tool emits and consumes, wires producer to consumer with evidence per edge. | PUBLIC |
| mneme | Layered agent memory + hybrid retrieval where every memory carries provenance. | PUBLIC |
| studio-engine | Replayable creative worlds -- shaders, sound, motion -- from a seed plus a receipt. | PUBLIC |
| build color | A bounded color-science workbench: 15+ spaces, 12 HDR tone-mapping operators, ΔE, CAMs, ICC, LUTs. | BETA 1.0.2 · PyPI |
| local-model | "The oracle decides, the model proposes. No receipt, no accept." A trained 14B coder behind the verified accept path; a sealed model-boundary receipt (in review) witnesses each crossing. | DEV ENGINE · BUNDLED LANE |
Source · catalog.html · lane_identity.dart (each line stays in the product's own words)
+08 · The private line -- public where safe
+Six organs stay gated, and the site says so plainly.
+Runtime · ORCA
"Run the workspace, keep the record. And never touch the target." Security-engagement workbench -- metadata-only, cannot reach the thing it assesses.
v1.0.0 · 361 tests · private line
Checkpoint · Aleph
One checkpoint over docs, package contracts, MCP declarations, CI, supply-chain evidence, receipts.
private line · public overview only
Boundary · behavior-transform
Mode-aware read/write/exec/fetch/input receipts. Raw content stays in local adapters; hosts receive counts, hashes, decisions, redacted refs.
public repo
Vault · Kun · Lab · Seed · Ledger · Sofer
Path-only receipts and rotation discipline; a controlled-lab assessment engine; an agent workflow ledger. Honestly labeled: awaiting a public-safe split.
private / proprietary
Source · catalog.html organ index · orca.html · aleph.html
+09 · Flywheel Desktop -- one window over the platform
+"No browser, no terminal."
+A native desktop surface with 24 destinations -- Chat, Compare, Code, Plan, Workflows, + Studio, Memory, Graph, Receipts, Lanes, and more -- grouped by what you came to do: + START · DO · KNOW · ADVANCED.
+-
+
- Under every chat turn: the re-derivable receipt -- request hash, prompt hash, response hash, served-model disclosure. +
- Tap any hash anywhere → jump to Receipts → Merkle inclusion proof, re-walkable offline. +
- Ships as one installer with the engine frozen inside -- no Python, no PATH setup. +
"The trust is built in, never in the way."
+The accountability posture
+-
+
- Local-first: every call goes to 127.0.0.1:8799. Zero telemetry in the client. +
- Credentials: presence only, never values -- typed once, stored in the OS keychain, never displayed, logged, or echoed back. +
- Capability gating: write and exec are grants, never defaults. +
- Honest nulls stay visible: "a view that cannot show DRIFT is not shipped." +
Source · flywheel-desktop README + CLAUDE.md · keys_panel.dart · tokens.dart
+10 · The connection surface -- bigger than its roster
+Seven lanes is the floor, not the ceiling.
+Each count keeps its tier label -- live, declared, built-in -- because a roster + that rounds "declared" up to "live" would fail its own discipline. A down lane reports missing honestly; it never breaks the roster.
+Source · lanes.py · SUPERPROJECT.md · catalog.html · marketplace.py · gateway.py (dev)
+11 · Competitive position -- measured, not asserted
+"Lead on features, UX, demo, speed. Accountability is the closing multiplier."
+| Ours | Class leaders (July 2026) | Honest verdict | The niche nobody occupies |
|---|---|---|---|
| crucible eval harnesses | promptfoo (OpenAI-owned), DeepEval, Braintrust, LangSmith | WE LEAD speed/cost THEY LEAD features · UX · traction | Benchmark-claim escrow: sealed verdict packets anyone can recompute offline -- no incumbent escrows the claim itself. |
| telos agent end-to-end agents | Claude Code, OpenCode, Devin, OpenHands | WE LEAD native computer control THEY LEAD features · traction | The concurrent-desktop agent: a second pair of hands on your logged-in workstation while your keyboard stays free -- every action a replayable evidence packet. |
| flywheel routers & runners | every router, every runner | "Routers route. Runners run. None of them check the work." The category's missing verb is verify -- that is the whole wedge. | |
The battle map publishes THEY-LEAD rows on its own products. A tool that refuses to overclaim is a tool whose other claims you can trust.
+Source · BATTLE-MAP.md (whole-product sweep, 2026-07-07) · flywheel.html
+12 · Register shift
+Here the portfolio becomes an offer.
+Two markets are named below, and inside them the retired language comes back + on duty -- because in these rooms it is not branding, it is the specification.
+"You can't grade your own homework.
That's the anti-hallucination heart."
Source · pitch-deck.pdf slide 3 · models-propose-oracles-dispose.html
+13 · Vertical A · Sales & marketing
+The industry that ships the most AI output owns the least proof of any of it.
+Claims & brand safety
Every public sentence traceable to a frozen source: gather receipts + citation snapshots that outlive the live web. Substantiation for ad claims, on demand.
Agency accountability
forum runs campaign agent-work on a replayable causal ledger with human approval gates -- the client re-checks the run instead of trusting the report.
AI-spend governance
flywheel routes to the cheapest honest source and keeps escalation receipts -- the first per-provider scoreboard of observed cost and success, not promises.
Reproducible brand systems
studio-engine + telos: seeded generative campaigns, parametric typefaces, brand kits -- same seed, same artifact, forever re-derivable.
Witnessed enablement
learn turns playbooks into graded courses where every completion writes a receipt -- onboarding proof instead of completion checkboxes.
Account memory
mneme keeps client knowledge with provenance per memory -- "who told us this, when, from what document" survives team turnover.
HONEST NOTE this vertical is greenfield: the current site + has zero sales/marketing surface. That is a gap and an opening -- no positioning debt to unwind.
+Source · engine pages mapped against agency/martech workflows · site grep (no existing surface)
+14 · Vertical B · Provenance-critical, privacy-bound industries
+Where "who touched this, and can we prove it" is the workflow.
+| Sector | The live question | The system's answer |
|---|---|---|
| Compliance & model risk | Regulators require model origin and lineage to be auditable -- incumbent platforms store lineage as editable database rows. | Merkle-logged receipts, offline inclusion proofs, hash-chained store with an audit tail. The record cannot quietly move. |
| Legal & e-discovery | "The summary is not the record." | Citation freezing, content-addressed corpora, attestation that binds sign-off to exactly what was walked -- overclaims tracked explicitly. |
| Healthcare admin | Can a summary keep its uncertainty attached? | gather + forum lanes with honest-null rendering; UNVERIFIABLE stays visible instead of rounding up. |
| Newsrooms | Can every public sentence find its source? | gather → forum → crucible chain; every block carries a source hash; tamper is caught on re-read. |
| Security & red teams | Authorized offensive work needs records that hold up afterward. | ORCA engagement workbench (metadata-only, never touches the target) + a lawful-basis gate published in the open. |
| Regulated AI ops | Agents act faster than review; boundaries must be provable. | Boundary receipts (raw content never leaves local adapters), capability grants, presence-only credentials, zero telemetry, fully local loop. |
Source · writing.html (model-risk) · field-guide.html lanes · security.html · witnessing-spine (DOI'd)
+15 · Privacy is architecture here, not policy
+Nothing leaves the machine
The only base URL in the desktop client is 127.0.0.1:8799. No telemetry, no analytics, no outbound endpoint. Local model tiers are first-class; escalation is a choice with a receipt.
Keys are never collected
"Presence only, never values." A credential is typed once into an obscured field, handed to the OS keychain, and shown forever after as presence + source. The gateway refuses to accept values through the marketplace.
Proof works offline
A receipt is a Merkle leaf; the inclusion proof returns root + audit path so a third party recomputes it with no network, no account, no vendor.
Authority stays with the human
Write and exec are grants per run. Attestation binds sign-off to coverage. Comprehension gates mean you own what you can review, explain, and defend -- pasting the diff back cannot pass.
"Built to be used humanely -- and architected to resist being used any other way."
+Source · keys_panel.dart · gateway_streams.dart · receipts_view.dart · why.html
+16 · Target audiences · who, where, why now, how in
+| Who | Where they live | Why now | Enter with | They get / we get |
|---|---|---|---|---|
| Agency ops lead | Campaign delivery, client reporting | Clients now audit AI-made work; agencies have no proof layer. | forum demo -- the replayable run ledger | Defensible delivery / first design partner in the greenfield vertical |
| CMO · martech owner | Brand, content, AI-spend budgets | AI spend is unmeasured; claims risk is rising. | flywheel routing scoreboard + gather claim receipts | Cost + substantiation in one surface / a paying wedge account |
| GRC / model-risk officer | Banks, insurers, model inventories | Lineage regulation arrived; audit trails are editable rows. | Witnessing Spine findings + receipts proof demo | Tamper-evident lineage / the regulated flagship reference |
| Legal ops / e-discovery | Firms, corporate legal | AI summaries entering the record without their sources. | snapshot (citation freezing) + attestation | Summaries bound to records / a dense-workflow proving ground |
| Newsroom standards editor | Media, fact-check desks | Generated content floods intake; provenance is the product. | field-guide media lane, gather→crucible | Source-checked publishing / public credibility case study |
| Red team / security lead | Gov, law enforcement, AI labs | AI-attack surface work needs lawful, replayable records. | ORCA + the security.html engagement gate | Engagement records that hold up / the existing revenue-shaped practice |
| Builders & researchers | OSS, eval/agent infra | Already served -- the audience the site speaks to today. | PyPI packages, demos, papers | Tools + receipts / contributors, field testers, standing |
Source · field-guide.html · security.html · writing.html · audience analysis 2026-07-30
+17 · Why these tools matter on today's frontier
+Agents outran review
Frontier condition: autonomous work lands faster than humans can check it.
Answer: the verified accept path -- an external oracle decides, and UNVERIFIABLE is a legal, visible outcome.
Provenance became law
Frontier condition: lineage and origin duties are moving from best practice to regulation across finance, health, media.
Answer: receipts as infrastructure -- content-addressed, offline-provable, vendor-independent.
Claiming became free
Frontier condition: benchmarks, capabilities, and creative originality are asserted daily and checked never.
Answer: claim escrow -- verdict packets a stranger, competitor, or court can recompute.
The wedge categories -- originality-verification, accountable-agent-action, + provenance -- have no analyst TAM yet. That's the whitespace.
+Source · pitch-deck.pdf slide 7 · presentation.html (the runway)
+18 · Property one · harperz9.github.io
+An exceptionally honest site that cannot see its own visitors.
+-
+
- 72 sitemap pages, near-100% meta/OG coverage, reproducible social cards, tested (30+ unit tests, link checker, visual contract), accessible by rule. +
- A published self-audit that caught its own over-counts -- "the honest figure was six" -- and a published null against its own product's uplift claim. +
- Recorded demonstrations with transcripts, machine-readable evidence, and claim boundaries. +
-
+
- No custom domain -- free github.io hosting; brand split three ways (Harper / Project Telos / ZentropyLabs). +
- Zero analytics -- the site is unmeasurable, by grep-verified fact. +
- No conversion instrument -- every CTA dead-ends at a gmail address. No pricing anywhere. +
- Homepage leads with a compiler test count, not a value proposition. +
Asset verdict
+VERIFIED content depth, hygiene, honesty discipline
+DRIFT measurement, conversion, naming
+Value today: a world-class evidence library and credibility engine. + Value locked away: every commercial signal a buyer would need to say yes.
+Source · sitemap.xml · SITE-FIX-LIST.md · toolkit.html correction · site-wide grep
+19 · Property two · the Flywheel product surface
+A real product whose public face runs one lap behind the engine.
+-
+
- A downloadable installer (Windows x64, 20 MB) with a published SHA-256 and an honest unsigned-binary note -- "the receipt does the signature's job." +
- Desktop v0.2.2 releases built from version tags by a public pipeline that refuses to ship a mislabeled artifact. +
- OpenAI-compatible gateway: point any existing client at it and get receipts under the same API. +
-
+
- The public flywheel repo exposes 11 API routes; the dev engine has 102. The public surface undersells the shipped product by roughly a factor of nine. +
- Distribution is GitHub-releases-only; no product page outside the portfolio site. +
Route parity, public vs dev
+A sync cadence from dev to public is the single cheapest credibility win available.
+Source · flywheel.html · flywheel.html · public-repo route grep · dev gateway path census, local, 2026-07-30
+20 · Recommendations -- in verdict order
+| # | Move | Why it compounds |
|---|---|---|
| 01 | One name, one domain. Consolidate Harper / Project Telos / ZentropyLabs under a single registered domain; keep github.io as a mirror. | Every paper, receipt, and release currently builds authority for a domain you don't own. |
| 02 | Instrument honestly. Privacy-respecting, self-hosted analytics + one conversion instrument (a real work-thread form or calendar). | The portfolio's own thesis: no claim without measurement. The site cannot currently make a measured claim about itself. |
| 03 | Publish the two vertical pages. Assemble the scattered provenance material into one buyer-facing argument; write the first sales-&-marketing surface (it does not exist yet). | Vertical B's evidence is already written -- it lives on seven different pages. Vertical A is an open field. |
| 04 | Sync the public engine. A release cadence that keeps the public flywheel repo within one minor version of the dev gateway. | Closes the 11-vs-102 gap; the README's claims become checkable against the code behind them. |
| 05 | Name a price for one thing. The security/ORCA practice is already revenue-shaped; give it (or a receipts pilot) a package and a number. | A single priced offer converts the whole evidence library from portfolio into pipeline. |
Also on the shelf, unexplored: two private site repos (harpercompliance-site, sendmyletter-site) suggest adjacent property experiments already underway.
+Source · property assessment 2026-07-30
+21 · Honest boundaries -- what this deck does not claim
+Published nulls stay published
Flywheel's own uplift test: +0.100 with a 95% interval of [−0.236, +0.420] -- includes zero, so no capability uplift is claimed. The site prints this against its own product.
Maturity labels held
Live ≠ declared ≠ forming. The installer is not code-signed (the hash is the substitute). arXiv listings are endorsement-pending. Demos are recorded workflows, not scale benchmarks.
Traction is early
Six PyPI packages, one downloadable product, zero pricing to date. The 900k-download record is from the graphics line, not the platform.
A deliberate register divergence
BRAND-NARRATIVE.md retires verification-first language at site level. Act II of this deck re-arms it on purpose, as vertical sales language -- flagged here so the divergence is a decision, not a drift.
Source · flywheel.html honest null · README.md honest boundaries · BRAND-NARRATIVE.md retire list
+"Don't trust it; check it."
+Every number in this deck traces to a repo, a page, or a receipt. + That is the product, demonstrated by its own pitch.
+Work threads open -- bring it real work, then try to re-check it.
+github.com/HarperZ9 · harperz9.github.io · flywheel desktop v0.2.2
+