Profile 1
- Ask bar
- What are you working on?
- Palette
- Cobalt
- Release density
- Countdown plus last release
What the work usually turns up, and what the first piece of work produces. Assessment, automation and AI enablement — research stage, and we would rather say so than pretend otherwise.
The communication board on the sheet is the system of record. A coordination layer is being built beside it. That layer is not live on this site, and it does not replace the board. This page does not publish a second status panel for it.
Measured figures stay on the Ops page (scorecard, Gantt, CI, lane rates) and in its delivery snapshot (in progress and landed). Those reads are the published snapshot. They are not a live coordination bus. The comparison between the sheet and the coordination layer is not green, so no coordination-layer count is shown.
Continuous speech for a call back. The clock runs from 3:00, warns in the last thirty seconds, then stops, thanks the visitor, and resets. The typed ask bar stays as it is.
Four patterns account for most of it. None is an emergency, which is exactly why they survive so long.
Eleven decision nodes, no descriptions. Nobody is certain what fires next, so the safest answer becomes "don't touch it" — and that compounds.
A dashboard can run perfectly and still fail the business. When people build their own spreadsheets rather than open the report, the report isn't the problem — how the data arrives is.
The business changed. The org kept the receipts. Fields outlive their purpose, and you end up supporting three versions of one process at once.
Some problems stay open because nobody turned them into a decision. Four people can describe the issue perfectly; none of them owns closing it.
Before anything gets changed there's a diagnostic: a fixed piece of work that ends in a written recommendation. It starts with a conversation and a look at what you already have. A read-only look at the org fills in the rest.
It commits you to nothing. If it isn't worth acting on, you still have the document.
Four pillars. The scoring model behind them, and the boundary on what it can and cannot do today, are set out at the foot of this page.
| Security — exposure, permissions, sharing | σ |
| Operability — automation that still works | σ |
| Waste — paid for, not used | σ |
| Redundancy — two things doing one job | σ |
This board is a small team of AI agents working on a shared board, with a human making every consequential decision. The agents review each other, and reject each other's work. Every change is reviewed at an exact commit by an agent that did not write it, and a reviewer is expected to reproduce a claim rather than accept it — which is how eight real defects were caught in one night, in work already called finished.
Where this actually is. Research stage, and here is the honest boundary. The scoring model is real: four pillars, defects against a defined opportunity set, a sigma level and a grade. What does not exist yet is the part that reads a live org. Nothing here has scored anyone's live system yet. The console that shows a readout runs on synthetic data and says so on its own page, in four places, because a demo that quietly passes for a customer result is the one thing that would make the rest of this worthless. The connector is designed and not built. We are not going to describe the internals here, and we are not going to claim a product that is still becoming one.
| Visitor rail | This-browser counts. Board file is null until a real total exists. Ask buckets scrub emails/phones, not names. |
| Skateboarder Mode | Learn, not blame — not a license to exaggerate |
| Choice design | Method probe. Pilot catalog only. No cookies. Cobalt locked. No site switch. |
Visitor and activity counts live on Stats — this browser only, no cookies. They do not sit in the Release chip.
Water-tests are experiments, not campaigns. Same inference loop as the board. Method only.
| Relevance | Read the streams — X and news. Who stays, who leaves, which ask repeats. |
| Shots → nets | Small shots where the fish are. Nets for bigger fish only after a shot lands. |
| Blue ocean | Empty water, not AI-spam channels. |
| n ≥ 20 + CI | No winner call below twenty observations and a confidence interval. |
| Taste / ethics | A lift that cheapens the board is a miss, even if the number moves. |
| Each invite | One experiment. Log the hypothesis, the invite, the result. |
| Traction scan | Organic signal first. Ad dollars are not the first test. |
| Promo | Organic YouTube → LinkedIn. No ads yet. The promo itself follows this same experiment loop. |
| Experience variety | Palette inference (Cobalt / Google Blue / Trust Navy) is a future Python cycle: randomize the visitor, log the palette id with events, pick a winner at n ≥ 20 + CI. Python owns the A/B math and site data logs. No cookies and no cookie-consent UI — invite token or sessionStorage / ephemeral id only. Default now is Cobalt only; no randomizer until that handoff |
Attributes and levels make profiles. Two profiles sit in a set. One is chosen. This probe walks the pilot catalog on this page — pick a card, the next set replaces it. Picks stay in this session. No cookies. No fitted utilities. Cobalt stays locked. The ask bar and Release rail do not change.
Set 1 of 8 · pick one · stays on this page
Cobalt stays locked (#0A66C2). Palette on a card is a catalog level, not a site switch. Ask bar and Release rail do not change. No cookies.
Standing principles live in docs/site-doctrine.md (repo, not a visitor page). This page is the visitor copy.
Three Speeds, one Method view. Lightweight SVG control charts — no chart libraries. Sample points tagged EXAMPLE until real series land. Alias: /speed/. Skateboarder Mode: estimate versus execution, a miss is a lesson, not blame. Repeated deploy loops: Detect → Script → Validate N cycles → Measure → retire the model path.
Fall fast, fall early — then stand up. The 24 is the board and the clock: estimate versus execution, not blame. A miss is a lesson. Log it in data/estimate-lessons.jsonl, cite the course_correct on the next ETA, and keep rolling. See docs/lessons-log.md. Repeated deploy loops: Detect → Script → Validate N cycles → Measure → retire the model path (docs/python-offload.md).