Staffless Co. — Dashboard

✋ Waiting on Jan

2026-07-13 — P1 v1 is a RESEARCH PREVIEW, not a sellable product. Please read.

**I need to walk back part of the business plan, on evidence.**

Driving our own API and MCP server found two defects the benchmark was *structurally

incapable* of finding:

1. **`Acme Corp.` vs `ACME Corporation` → 0.385 → "NOT a match."** (`Microsoft Corp` vs

`Microsoft Corporation` → 0.936, because `microsoft` is a token the model has *seen*.)

2. **The score depends on how many fields you send.** The same true-match pair scores

0.412 with 1 field, 0.797 with 2, 0.862 with 3 — it flips across the 0.45 threshold

purely by adding columns.

**Root cause:** the model exploits *distributional artefacts* of its training data (field

count, corpus token frequency) rather than learning name equivalence. Our sequestered-test

numbers (0.979 / 0.862 / 0.658) are **true** — for records shaped like that training data.

They do **not** transfer to two-column name matching, which is the single most obvious

entity-resolution use case.

**Therefore:** selling this as a general-purpose matcher would be a promise we cannot keep

(Hard rule 5). **P1 v1 is a research preview with a published, reproducible benchmark.**

The determinism, the cost ($0.0012/1M), the sub-millisecond latency and the benchmark are

all real and publishable. The general product is not real yet. That is a smaller claim than

the business plan assumed, and it is the true one.

**v2 needs a redesign, not a patch:** field-count-invariant aggregation; a

corpus-frequency-free similarity floor; and a company-/person-name eval set with a **fresh**

sequestered split (the current test set is closed).

---

2026-07-13 — Two blockers I cannot route around (both need you)

**1. Cloudflare token has no DNS scope.** `CLOUDFLARE_API_TOKEN` verifies as active but

sees **zero zones** — it was created for Pages only. I cannot point `kywio.com` at the VPS.

→ *Add Zone:Read + Zone:DNS:Edit for kywio.com to the token (or issue a new one).*

**2. I cannot open a public port.** `ufw` allows only 22/80/443; binding :80 needs root; and

`sudo`, `systemctl` and `ufw` are all on my deny list (correctly). So the REST API cannot be

exposed publicly, nor run as a durable service, without you.

→ *Simplest fix: `sudo ufw allow 8080/tcp`, then I'll route Cloudflare to origin port 8080.

For durability I'll also need a systemd unit installed — I'll write the file; you run

`systemctl enable --now`.*

**Not blocked by either:** the **MCP server works right now**, over stdio, no hosting needed:

`claude mcp add kywio -- /home/company/kywio-venv/bin/python /home/company/company/api/mcp_server.py`

2026-07-13 — Small ask: an ANTHROPIC_API_KEY, to measure our own headline claim (~$2-3)

**The problem.** Our central competitive claim is that we are orders of magnitude cheaper

and faster than LLM-direct matching. Right now we cannot *measure* that ourselves:

`ANTHROPIC_API_KEY` is absent. What we hold is `CLAUDE_CODE_OAUTH_TOKEN`, which drives

Claude Code but **cannot call the Messages API**.

**Why it matters.** A company whose entire brand is research rigor should not rest its

headline number on someone else's published figures. Every other number in our benchmark is

ours and reproducible; the LLM comparison would be the one exception.

**The ask.** Create an Anthropic API key and add `ANTHROPIC_API_KEY=...` to the VPS `.env`.

small model). Well inside the $10 per-transaction auto-approve.

**If you'd rather not:** two honest fallbacks, both worse.

1. Cite the literature (Peeters & Bizer) and label it **EXTERNAL / UNVERIFIED — not our

measurement**, everywhere it appears.

2. Drop the LLM comparison from the headline entirely and sell on determinism + cost, which

we *have* measured.

**Recommendation:** option (1) in the ask — it's ~$3 and it makes our loudest claim our own.

**If unanswered:** I proceed with fallback 1 (cite + label as external) and the benchmark

still ships.

2026-07-13 — T013: 5 company name candidates — ✅ SUPERSEDED, see below

*Resolved 2026-07-13: the CEO asked the question I failed to ask — whether the

company name should be the product name at all. All 5 candidates below were

entity-resolution names; I had silently assumed company == P1. He chose a branded

house (neutral company name + descriptive product names) and proposed **KYWIO**,

which verified clean on all four TLDs with no trademark footprint. The 5 below are

retained as **P1 product-name** options, not company names.*

Availability **verified by RDAP this tick**, not guessed — `.com` via Verisign,

`.io` via Identity Digital, `.dev` via rdap.org, each smoke-tested against a

known-taken and a known-free control first. (The first `.dev` endpoint I tried

reported `google.dev` as available; the controls caught it and I discarded it.)

Availability ≠ price: premium/reserved pricing is NOT checked and can be

surprising. Verify at the registrar before purchase.

| # | Name | .com | .io | .dev | Note |

|---|------|------|-----|------|------|

| 1 | **EntityKit** ← *my pick* | taken | **free** | **free** | Says exactly what it is |

| 2 | **SameFold** | taken | **free** | **free** | Ownable, serious, not dedup-narrow |

| 3 | **Samewise** | taken | **free** | **free** | Most memorable |

| 4 | **KinMatch** | taken | **free** | **free** | Descriptive + evocative |

| 5 | **DedupeLab** | **free** | **free** | **free** | Best domains — but see the caveat |

**1. EntityKit** — *Recommendation.* Descriptive to the point of being boring,

which is the point: our customers are agents doing semantic search, so a name

that literally states the capability is a discovery asset, not a missed branding

opportunity. Clean across trademark and search. **Downside, stated honestly:**

descriptive names are weak trademarks — we'd have a hard time stopping anyone

else using "entity kit". If you want a defensible mark, pick #2.

**2. SameFold** — Coined, so genuinely ownable. Evokes folding duplicates

together, and unlike the Dedupe\* names it covers the whole product (match +

resolve + dedup), not just dedup. Neutral, serious register that fits a

research-first brand. The strongest choice if trademark defensibility matters

more than instant legibility.

**3. Samewise** — The memorable one; "judging sameness" with a Samwise echo.

**Downside:** the pun cuts against a rigor-is-our-brand positioning, and it's the

one name here I could see us regretting in a paper's author line.

**4. KinMatch** — Kinship = records that are the same entity. Clear and broad.

Slightly generic; no strong objection, no strong pull.

**5. DedupeLab** — The only candidate sweeping **.com + .io + .dev**, and "Lab"

matches the business plan's "research excellence is the brand." **But the caveat

is real and I'd weigh it heavily:** `dedupe` is an existing Python library *and*

a commercial service (Dedupe.io) — and it is **one of the baselines we intend to

benchmark against**. Naming ourselves after a competitor's library invites brand

confusion and looks careless in exactly the audience we're trying to impress. I

would take worse domains over that. Same objection applies to `Dedupa`

(.com/.io/.dev all free) — noted as an alternate, not recommended.

**Rejected during the check (worth knowing):**

(Akin AI PBC, NASA JPL contract), plus Akin Gump in legal. A name collision

inside the AI space is disqualifying.

biomedical domain we chose as our showcase. Disqualifying.

Different trademark class so low legal risk, but it owns the search results,

which is bad for a company whose marketing is a published benchmark.

**What I need:** your pick (or a name of your own). No domain is purchased under

this task — the purchase is T010 and comes back to you with a price first.

**If unanswered:** nothing happens; T010 and the registry listings stay blocked

on the name, but T007 (the spec) proceeds in parallel and is not blocked by this.

Task board

Recent journal

honest repositioning).**

Built and drove `api/mcp_server.py` over real stdio JSON-RPC. It works **today, with no

hosting**: `claude mcp add kywio -- .../python api/mcp_server.py`. Tools: `match_records`,

`dedupe_records`, `kywio_info`. The AI disclosure and the known limitations are returned

in the tool descriptions **and in every tool result** — an agent that can't see the caveats

will happily build on a model that doesn't work for its data.

**Driving it immediately exposed NEG-005.** The MCP call naturally sent *one* field instead

of the benchmark's three, and the same true-match pair came back **0.430 → "not a match"**

where REST had said 0.862. Quantified:

| fields sent | score | verdict (thr 0.45) |

|---|---|---|

| 1 | **0.4124** | **NOT a match** |

| 2 | 0.7973 | match |

| 3 | 0.8617 | match |

**The same pair flips across the threshold purely by adding columns.**

**Root cause, now unmistakable across NEG-004 and NEG-005: the model exploits

distributional artefacts of its training data — field count and corpus token frequency —

rather than learning name equivalence.** Its threshold is only meaningful for records

shaped like the training set.

**So I am walking back part of the business plan, on evidence.** Our sequestered-test

numbers (0.979 / 0.862 / 0.658) are *true* — for records of that shape. They **do not

transfer** to two-column company-name matching, which is the single most obvious ER use

case. Selling this as a general-purpose matcher would be a promise we cannot keep (Hard

rule 5). **P1 v1 is a RESEARCH PREVIEW with a published, reproducible benchmark — not a

sellable general product.** The determinism, the $0.0012/1M cost, the sub-millisecond

latency and the benchmark are all real. The general matcher is not real yet. That is a

smaller claim than the plan assumed and it is the true one.

v2 needs a **redesign, not a patch**: field-count-invariant aggregation, a

corpus-frequency-free similarity floor, and a name-matching eval set with a **fresh**

sequestered split (the current test set is closed).

**Two blockers I cannot route around, both surfaced to Jan:** the Cloudflare token is

Pages-scoped and sees **zero zones** (no DNS control over kywio.com), and I cannot open a

public port (`ufw` allows only 22/80/443; binding :80 needs root; `sudo`/`systemctl`/`ufw`

are all correctly denied to me). The REST API therefore cannot be exposed publicly without

him. **The MCP server needs neither and works now.**

Three ticks, three defects, all found by *using the product* and none findable by the

research. The pattern is no longer a coincidence — it is the finding.

Recent ticks (cost)