/01 · in flight
deep tier + cited rationale
policy-tuned llm path for the hard 1%. every block ships with a citation pointing at the policy clause that triggered it.
confidence · high62%
/roadmap · revision 03 · live
what we're
building,
roughly.
a working document. dates are best guesses. items get added, dropped, and shuffled the moment we learn we were wrong about them.
nothing is finaleverything is negotiabletell us when we're wrong
in the works
this week · deep tier · model fine-tune/this week · rationale citations · prototype/this week · flash tier · benchmarking/this week · echo tier · eval harness/this week · slang corpus · v3 training/this week · ts sdk · scaffolding/this week · docs site · drafting/this week · deep tier · model fine-tune/this week · rationale citations · prototype/this week · flash tier · benchmarking/this week · echo tier · eval harness/this week · slang corpus · v3 training/this week · ts sdk · scaffolding/this week · docs site · drafting/
04
horizons
17
items tracked
00
shipped · pre-launch
06
in flight now
Q2
now
spring 2026
pre-launch · private beta opens
we haven't shipped the api yet. q2 is about getting all three tiers, the rationale system, and the sdks far enough to put a key in early hands.
eta · jun
/02
monthlyslang corpus v3
discord, twitch, reddit. first training run before launch, monthly cadence after.
/03
jultyped sdks
tspy· zero deps
/04 · in flight
junflash tier
deterministic regex + lexicon path. targeting ~1ms p99 on the obvious 95% of traffic.
/05 · in flight
junecho tier
small classifier for the 4% the rules miss. targeting ~30ms p95.
Q3
next
summer 2026
post-launch · scope, not survival
assuming q2 lands, q3 is the first round of things we want right after the api is live: multilingual, batch, and the surfaces that make the api easier to live with.
/01 · first thing post-launch
multilingual
native, not translated. each language gets its own slang corpus and tuned thresholds. english ships at launch; this is the very next thing we pick up.
esptfrdeja
confidence · med55%
/02 · early target
jul · alphabatch endpoint
10k messages, one call, async return. one of the things we want out the door pretty early after launch.
conf · med60%
/03 · planned
augpolicy editor (web)
tune category thresholds. preview against history before deploying. timing depends on how launch lands.
conf · med50%
/04
sepwebhooks
signed deliveries. retries with backoff.
/05
sepaudit log
searchable. every decision, every revision.
/06
prepsoc 2 prep
controls in place. audit window opens q4.
Q4
later
winter 2026
scope-uncertain · scale + new modalities
things we want, but where the timeline is honestly fuzzy. anything here can move into next, slip into someday, or get cut entirely.
/01
image moderation
same shape, different payload. nsfw + violence first. not sure how fast we can get this real — depends on how text lands.
/02
eu + apac regions
data residency. gdpr-shaped. for scale plans.
/03
more languages
second wave after the q3 set. driven by what waitlist teams actually need.
/04
soc 2 type ii
audit window opens. report on request after.
27+
someday
things we chew on
no dates. no commitment. these are conversations we keep having internally and would love to hear opinions on.
exploring
— · —audio moderation (live + recorded)
streamed transcripts route to the text path. raw clips run through a separate model for nsfw flags. open question: latency budget for live streams.
exploring
— · —customer-tuned models
drop in 1k flagged examples. we fine-tune your slice without leaking it back into the base. open question: pricing model.
exploring
— · —on-device flash tier
wasm bundle. zero server hop for the obvious 95%. open question: keeping the corpus fresh without a chatty update channel.
exploring
— · —moderator console
because eventually a person has to look at the hard 1%. queue, hotkeys, decision audit. open question: build vs partner.
/anti-roadmap
things we are
not doing.
a roadmap is more useful when it includes what's off the table. these are not “yet” — these are no.
- 01noa moderator-as-a-service product
we sell the api. you build the workflow. we won't review your queue for you.
- 02nevertraining on customer content
not now, not later, not "anonymized." the privacy policy says it; the company is structured to keep it true.
- 03noa chat-based "ask the moderator" interface
moderation is structured output. chat surfaces lose the structure. wrong tool.
- 04nevergovernment surveillance contracts
not what we built this for.
/feedback
roadmaps shift in response
to real teams.
tell us what you'd build with us if it existed. we read every reply, and the loud ones move things up the list.