Roadmap

Recommended roadmap (2026-10-12, updated 2026-10-16)

This is a recommendation, not a commitment. Each horizon ends at a gate, and the gate decides whether the next one starts. The IDs refer to BACKLOG.md.

Horizon	Goal	Key work	Gate to pass
Now (≈ 6 weeks)	Build v01 and see it work	Close the sourcing gaps (B-01–B-07). Print and assemble the kit (B-14). Ring encoder and audio pipeline (B-08, B-09). Measure power and heat on the real board (B-10, B-40). Start the model-efficiency research track (B-41).	One kit records a meeting. The ring shows each person's level. A phone reads the note by NFC.
Next (≈ 1 quarter)	Show that people trust it	Ring cold-read test (B-15). Phrase collection in EN/ES (B-16). Speaker-attribution error rate (B-11). Accessibility signal (B-18). Counsel review (B-19). Decisions on R1 retention and copying (B-20). Two or three real meetings and one small event with consent. Design partners in three segments (B-87). The mod framework and its "only lower" check (B-81, B-82).	Strangers read the ring. Lowering works in real rooms. Counsel clears the default-on policy for at least one setting (e.g. internal meetings, EU).
Later (≈ 6–12 months)	Make the 72-hour version	v02 low-power capture architecture (B-21). Custom PCB at EVT (B-22). Publish the Resolution Protocol draft (B-23). One anchor pilot (B-24). The first three packs: German, Research, Facilitator (B-84–B-86).	72 h on one charge on the bench. A conformance test suite that a second device could pass.

Why this order: the riskiest assumptions are human, not technical: can people read the ring, and do they trust it? So v01 goes into real rooms before money is spent on a custom board. Since ADR-0014, v01 runs about 3 h on its cell and indefinitely on a USB-C charger, so a long session or a plugged-in workshop day works now. The 72 h stream on battery alone needs a different architecture (B-21), and energy per transcribed minute is a research problem in its own right (B-41): how efficiently the models run decides battery life, heat and the size of the device more than the cell does.

Business track (runs alongside):

Publish the protocol early (option A in v01/research/market.md §6). Amazon has bought Bee and Meta has bought Limitless, so an open consent layer is the position no platform can credibly take.
An enterprise or event pilot (option B) produces the evidence a licensee would ask for.

Mods track (ADR-0023): add-on packs that make the device better for a particular room, installed over USB-C, with no cloud.

Why: as phones and room systems get on-device AI, computing stops being the reason to buy Goldfish. What stays is a shared object nobody at the table owns, a light wired to the microphones, no radio, and the protocol. Packs let it keep improving for one kind of room without asking anyone to trust a server.
The rules: a mod can only lower, never raise or keep longer. The note names every active mod. No emotion or sentiment inference. Signed packs with an offline licence. The firmware and the checks stay open source.
Order:
1. The framework first: signed install, the hostile-pack check, the note listing mods (B-81–B-83).
2. Then three packs, each tied to a segment:
  - German (works councils): a speech model plus the spoken phrases. Released when recall on real German phrases is reported.
  - Research (UX and qualitative researchers): personal details removed, insight tags. Released after one ethics board has reviewed it.
  - Facilitator (facilitators and teams): a talk-time balance light, decision capture, a parking lot. Released after a cold-read test of the balance light. It is also the clearest demo of "better sessions".
3. Later: more languages and domains (clinical huddles, legal, boards, retros, education), and certified packs from partners or the community.
Gate: a pack ships only with its evidence, and only if speech recognition stays real-time on the board it runs on. Heavier packs wait for v02.

Customer outreach (runs alongside, starting at Next): learn from a few rooms before talking to many.

Segments, in order of how sharp the need is:
1. Works councils, Germany first. They already block AI notetakers. Goldfish answers their objection by design, so this is the best wedge.
2. Qualitative and UX researchers. Ethics boards and GDPR push for keeping less. Per-person consent and forgetting by default are what their protocols ask for.
3. Professional facilitators. They are the landing page's audience and the best word of mouth. Each buys one or two devices, so they are evangelists more than a market.
4. Sensitive internal sessions: HR, boards, legal, clinical teams, the public sector. Reached later, through the first three.
Where to meet them:
- Works councils: the unions' technology advisers (ver.di, IG Metall), works-council trainers, works-council conferences.
- Researchers: ResearchOps and UX research communities, university ethics boards and research-data offices.
- Facilitators: IAF chapters, INIFAC, the Liberating Structures and Art of Hosting communities, Group Works.
- Dialogue: civic-dialogue groups, and Cortico and the MIT Center for Constructive Communication, the closest neighbour (B-90).
- Builders: open-hardware communities, for the kit.
How:
- A design-partner program: about 10 organisations across the first three segments. Kits on loan, as libraries lent Cortico's Hearth. A consent script and a short session plan. A debrief after every session (B-87).
- Paperwork that removes the blocker: a one-page privacy sheet for IT, and a draft works agreement for works councils, reviewed by counsel (B-88, ties to B-19).
- Proof before promotion: with consent, publish real session notes and what people said about trusting the lights. Use the open protocol and the public repo (B-78) as credibility, not as the pitch.
- Segment the waitlist: the sign-up's "What kind of sessions do you run?" answers decide which segment hears first. Landing-page variants for each segment once the sign-up is live (B-89, needs B-76).
- Talks and workshops at facilitation and research events. Run a real session with Goldfish on the table, not a slide deck.
What each pilot must answer:
- Did people trust the lights and speak at least as freely?
- Were spoken requests understood?
- Were the notes useful to the facilitator?
- Which pack would they ask for first?
- What blocked the organisation from saying yes?

Stage gates

Standard hardware stage gates, adapted. Each gate has an exit test; don't move on without it.

Phase	Name	What gets built	Exit test
0	Concept	Brief, ladder + forgetting curve, simulator, phrase collection, ring cold-read	Strangers read the ring correctly; lowering phrase set collected
1	Feasibility	Works-like rig (dev boards) + looks-like stone with a real LED ring + Wizard-of-Oz facilitated session	Lowering and intent spikes pass; 72-hour power closes for A or A′
2	Proof of concept	First custom board, hand-assembled	A full 72-hour stream on one device, through decay to R1 on an accelerated clock
3	EVT	~10–20 units, near-final electronics	All S-xx properties verified; EMC pre-scan
4	DVT	~50–200 units, production-intent enclosure	Certification, forensic decay test, field trial with facilitators in real sessions
5	PVT	Pilot run	Yield, test fixtures, packaging
6	MP	Mass production	—

v1 scope

In:

A continuous stream.
Each participant sets the level for their own words (ADR-0008). Room-wide lowering affects everyone.
Lowering by plain English, including the retroactive "forget what I just said".
Raising your own words directly. Raising the room by name plus an objection window.
The forgetting curve R3 → R2 → R1.
The light ring at the seam, one segment per person, and the single eye light (ADR-0009).
Reading only by physical handling, with a passive NFC hand-off (ADR-0007).

Later: close-of-meeting renegotiation; the 72-hour stream on one charge (v02 architecture).

v01 kit (built 2026-10-12)

The works-like + looks-like kit is designed in v01/. It moves phase 1 forward: order the parts, print the shells, then benchmark speech and power on the real board.

Phase 0 — next steps

Recover the Claude.ai Goldfish discussions into 00-sources/ (see the note there).
Play with 07-simulator/index.html — the Stream and Forgetting tabs model the conference scenario.
Decide: how long outcomes (R1) last; whether anything can be copied off (ADR-0004); 72 h with or without overnight charge.
Collect lowering phrases from real people, in English and Spanish.
Build a ring prototype (5 LEDs + dot on a microcontroller) and run the cold-read test.
Pick dev boards for the phase-1 power spike.

# Roadmap ## Recommended roadmap (2026-10-12, updated 2026-10-16) This is a recommendation, not a commitment. Each horizon ends at a gate, and the gate decides whether the next one starts. The IDs refer to [BACKLOG.md](BACKLOG.md). | Horizon | Goal | Key work | Gate to pass | |---|---|---|---| | **Now** (≈ 6 weeks) | Build v01 and see it work | Close the sourcing gaps (B-01–B-07). Print and assemble the kit (B-14). Ring encoder and audio pipeline (B-08, B-09). Measure power and heat on the real board (B-10, B-40). Start the model-efficiency research track (B-41). | One kit records a meeting. The ring shows each person's level. A phone reads the note by NFC. | | **Next** (≈ 1 quarter) | Show that people trust it | Ring cold-read test (B-15). Phrase collection in EN/ES (B-16). Speaker-attribution error rate (B-11). Accessibility signal (B-18). Counsel review (B-19). Decisions on R1 retention and copying (B-20). Two or three real meetings and one small event with consent. Design partners in three segments (B-87). The mod framework and its "only lower" check (B-81, B-82). | Strangers read the ring. Lowering works in real rooms. Counsel clears the default-on policy for at least one setting (e.g. internal meetings, EU). | | **Later** (≈ 6–12 months) | Make the 72-hour version | v02 low-power capture architecture (B-21). Custom PCB at EVT (B-22). Publish the Resolution Protocol draft (B-23). One anchor pilot (B-24). The first three packs: German, Research, Facilitator (B-84–B-86). | 72 h on one charge on the bench. A conformance test suite that a second device could pass. | **Why this order:** the riskiest assumptions are human, not technical: can people read the ring, and do they trust it? So v01 goes into real rooms before money is spent on a custom board. Since ADR-0014, v01 runs about 3 h on its cell and indefinitely on a USB-C charger, so a long session or a plugged-in workshop day works now. The 72 h stream on battery alone needs a different architecture (B-21), and energy per transcribed minute is a research problem in its own right (B-41): how efficiently the models run decides battery life, heat and the size of the device more than the cell does. **Business track (runs alongside):** - Publish the protocol early (option A in `v01/research/market.md` §6). Amazon has bought Bee and Meta has bought Limitless, so an open consent layer is the position no platform can credibly take. - An enterprise or event pilot (option B) produces the evidence a licensee would ask for. **Mods track (ADR-0023):** add-on packs that make the device better for a particular room, installed over USB-C, with no cloud. - **Why:** as phones and room systems get on-device AI, computing stops being the reason to buy Goldfish. What stays is a shared object nobody at the table owns, a light wired to the microphones, no radio, and the protocol. Packs let it keep improving for one kind of room without asking anyone to trust a server. - **The rules:** a mod can only lower, never raise or keep longer. The note names every active mod. No emotion or sentiment inference. Signed packs with an offline licence. The firmware and the checks stay open source. - **Order:** 1. The framework first: signed install, the hostile-pack check, the note listing mods (B-81–B-83). 2. Then three packs, each tied to a segment: - **German** (works councils): a speech model plus the spoken phrases. Released when recall on real German phrases is reported. - **Research** (UX and qualitative researchers): personal details removed, insight tags. Released after one ethics board has reviewed it. - **Facilitator** (facilitators and teams): a talk-time balance light, decision capture, a parking lot. Released after a cold-read test of the balance light. It is also the clearest demo of "better sessions". 3. Later: more languages and domains (clinical huddles, legal, boards, retros, education), and certified packs from partners or the community. - **Gate:** a pack ships only with its evidence, and only if speech recognition stays real-time on the board it runs on. Heavier packs wait for v02. **Customer outreach (runs alongside, starting at Next):** learn from a few rooms before talking to many. - **Segments, in order of how sharp the need is:** 1. **Works councils**, Germany first. They already block AI notetakers. Goldfish answers their objection by design, so this is the best wedge. 2. **Qualitative and UX researchers.** Ethics boards and GDPR push for keeping less. Per-person consent and forgetting by default are what their protocols ask for. 3. **Professional facilitators.** They are the landing page's audience and the best word of mouth. Each buys one or two devices, so they are evangelists more than a market. 4. **Sensitive internal sessions:** HR, boards, legal, clinical teams, the public sector. Reached later, through the first three. - **Where to meet them:** - Works councils: the unions' technology advisers (ver.di, IG Metall), works-council trainers, works-council conferences. - Researchers: ResearchOps and UX research communities, university ethics boards and research-data offices. - Facilitators: IAF chapters, INIFAC, the Liberating Structures and Art of Hosting communities, Group Works. - Dialogue: civic-dialogue groups, and Cortico and the MIT Center for Constructive Communication, the closest neighbour (B-90). - Builders: open-hardware communities, for the kit. - **How:** - **A design-partner program:** about 10 organisations across the first three segments. Kits on loan, as libraries lent Cortico's Hearth. A consent script and a short session plan. A debrief after every session (B-87). - **Paperwork that removes the blocker:** a one-page privacy sheet for IT, and a draft works agreement for works councils, reviewed by counsel (B-88, ties to B-19). - **Proof before promotion:** with consent, publish real session notes and what people said about trusting the lights. Use the open protocol and the public repo (B-78) as credibility, not as the pitch. - **Segment the waitlist:** the sign-up's "What kind of sessions do you run?" answers decide which segment hears first. Landing-page variants for each segment once the sign-up is live (B-89, needs B-76). - **Talks and workshops** at facilitation and research events. Run a real session with Goldfish on the table, not a slide deck. - **What each pilot must answer:** - Did people trust the lights and speak at least as freely? - Were spoken requests understood? - Were the notes useful to the facilitator? - Which pack would they ask for first? - What blocked the organisation from saying yes? ## Stage gates Standard hardware stage gates, adapted. Each gate has an exit test; don't move on without it. | Phase | Name | What gets built | Exit test | |---|---|---|---| | 0 | **Concept** | Brief, ladder + forgetting curve, simulator, phrase collection, ring cold-read | Strangers read the ring correctly; lowering phrase set collected | | 1 | **Feasibility** | Works-like rig (dev boards) + looks-like stone with a real LED ring + Wizard-of-Oz facilitated session | Lowering and intent spikes pass; 72-hour power closes for A or A′ | | 2 | **Proof of concept** | First custom board, hand-assembled | A full 72-hour stream on one device, through decay to R1 on an accelerated clock | | 3 | **EVT** | ~10–20 units, near-final electronics | All S-xx properties verified; EMC pre-scan | | 4 | **DVT** | ~50–200 units, production-intent enclosure | Certification, forensic decay test, field trial with facilitators in real sessions | | 5 | **PVT** | Pilot run | Yield, test fixtures, packaging | | 6 | **MP** | Mass production | — | ## v1 scope **In:** - A continuous stream. - Each participant sets the level for their own words (ADR-0008). Room-wide lowering affects everyone. - Lowering by plain English, including the retroactive "forget what I just said". - Raising your own words directly. Raising the room by name plus an objection window. - The forgetting curve R3 → R2 → R1. - The light ring at the seam, one segment per person, and the single eye light (ADR-0009). - Reading only by physical handling, with a passive NFC hand-off (ADR-0007). **Later:** close-of-meeting renegotiation; the 72-hour stream on one charge (v02 architecture). ## v01 kit (built 2026-10-12) The works-like + looks-like kit is designed in `v01/`. It moves phase 1 forward: order the parts, print the shells, then benchmark speech and power on the real board. ## Phase 0 — next steps 1. Recover the Claude.ai Goldfish discussions into `00-sources/` (see the note there). 2. Play with `07-simulator/index.html` — the **Stream** and **Forgetting** tabs model the conference scenario. 3. Decide: how long outcomes (R1) last; whether anything can be copied off (ADR-0004); 72 h with or without overnight charge. 4. Collect lowering phrases from real people, in English and Spanish. 5. Build a ring prototype (5 LEDs + dot on a microcontroller) and run the cold-read test. 6. Pick dev boards for the phase-1 power spike.