Stop guessing temperature
Know exactly what to change. Prove it worked by tomorrow.
Describe your workload and get the settings to paste — temperature, memory, alert sensitivity, commit threshold — plus one named follow-up test that verifies the fix within a session. Every number traces to registered, published research. Monitoring tools show you the problem; this hands you the fix and the proof.
No account. About a minute. Free to rerun.
The free check takes about a minute and is directional. The diagnostic adds a written memo: what to change, in what order, and the test that proves it worked. Every number traces to a derived law — including the checks that failed.
live demo
Try one workflow in under a minute.
Pick a preset, run RPCS-1, and see the status, configuration, language mode, and next check.
Best next check
Running the live demo now...
Receiver profile — estimate → detect → commit
Estimate
Detect
Commit
Recommended settings — paste these
New — SendRight
Say it your way. Send it right.
Type a prompt like you'd say it out loud. SendRight shows you what your words actually say — forks, contradictions, tangled asks — then opens your own AI app with the clear version filled in. No signup, no keys, free.
The problem
Three ways agents fail. You’ve seen at least one this week.
Oscillation
It flip-flops — rewrites its own work, reverses decisions, chases every new signal.
Usually: Agility and Gain set too high for how noisy the task actually is.
Overload
It drowns — context stuffed full, every alert firing, quality collapsing as the session grows.
Usually: Memory holding too much, Trigger set too loose.
Freeze
It stalls — hedges forever, asks for confirmation it doesn’t need, never commits.
Usually: Commit bar set too high for the actual stakes.
These aren’t metaphors — they’re the three failure modes of any bounded decision system, and which one you’re near is computable from your workload. See the derivation →
Exactly what you get
Know what you're buying — and what it changes in your workflow
Free tuner
$0 · no account- →Your agent’s five-dial profile — Memory, Gain, Trigger, Agility, Commit — for one described workload
- →Which failure mode it’s closest to: stable, oscillation, overload, or freeze
- →Runtime settings to paste: temperature, max_tokens, context and tool-use strategy
- →The reasoning behind every number — nothing is a vibe
In your workflow: a 60-second check before you ship a config change. Directional, deterministic, free to rerun.
Founding diagnostic
first 3 free · then $99- →Five dials measured for one configured agent, with a failure-risk score
- →The runtime settings to change before rollout, in priority order
- →One named follow-up test — so you know within a session whether the fix worked
- →A written memo your team can act on: a checklist, not a research project
In your workflow: apply the settings, run the named test, know by tomorrow — instead of guessing across a week of prompt edits.
The model
Five dials, not fifty flags.
Every agent runtime — any model, any framework — reduces to five settings. RPCS-1 reads your workload and sets each one.
Memory
TI · Temporal Integration
How much history your agent holds onto
Too much: stale answers in a fast-moving task. Too little: it forgets what it just learned.
Agility
UE · Update Elasticity
How fast it changes its mind when the world changes
Too high: it thrashes on noise. Too low: it keeps acting on last week’s reality.
Gain
SG · Signal Gain
How loud incoming signals get before it reacts
Too high: everything looks like an emergency. Too low: real problems slip past.
Trigger
FT · Firing Threshold
How much pressure it takes to notice something
Too tight: false alarms. Too loose: it sleeps through the fire.
Commit
AR · Action Rule
How much evidence it needs before it acts
Too eager: premature actions it has to undo. Too cautious: it stalls while the window closes.
RPCS-1 Translator
The same engine, pointed at people.
You say it once. It lands the way each reader needs to hear it.
- →Untangle a message that could mean three different things
- →Turn fragmented notes into something you can actually send
- →Split a mixed request into its separate asks
- →Rewrite anything for the specific person receiving it — their profile, not a generic “style”
How do you want this explained?
Pick a reading profile — change it anytime, up top.
Profiles change the explanations only — pricing, deliverables, and limitations are identical in every register. The values shown are the typical calibration for readers who prefer that register, not a measurement of you: axes a register says nothing about sit at the neutral prior (50). Want your own numbers? Take the 60-second calibration.
See the difference
Same product. Same deal. Your language.
This is the translation engine doing to our own site what it does to your content. The statement below is identical in meaning in every profile — only the register moves.
Your profile · Technical
RPCS-1 maps your workload’s entropy, stakes, and change rate onto five runtime dials via derived receiver laws, and returns the nearest failure mode plus a falsifiable follow-up test.
Compared with
You describe what your agent does. We tell you what to change, and give you one test that shows whether the change worked.
What never moves: the free tuner is $0 with no account, the founding diagnostic is first-3-free then $99, and the deliverables lists are word-for-word identical in every profile. We translate the explanation. We never translate the deal.
You can trust the numbers because we publish the misses.
Every recommendation traces to a law that was checked numerically against criteria fixed before the data was generated. Three registered checks failed during development — one corrected and re-run, one cut from the claims entirely, one traced to an artifact. All three are reported in full, because a scorecard you can trust has to include the misses.
The research & the full scorecard →Is your agent stable enough to ship?
Most teams don’t need a bigger theory. They need to know what to change and which test confirms the fix.