How LongChat Guard estimates long-conversation risk without an official quota
LongChat Guard does not know OpenAI's private remaining-conversation allowance. It measures the current chat locally and compares it with a historical chat that you personally confirmed reached the conversation-length limit.
The three core local signals
L — current conversation load
A local estimate of how much conversation material is present. It is an internal signal, not a user-facing official token count.
R — empirical failure reference
R is established from a historical chat that the user explicitly confirms reached the conversation-length limit and that can be read completely and reliably. It is local evidence, not an official OpenAI maximum.
G — whole-turn growth reserve
G is learned from reliable conversation-load changes from before a user turn to after the assistant completes it. It represents whole-turn growth rather than only the assistant reply.
Optional diagnostics
Parser and measurement schema versions protect measurement compatibility. A readable model label may be retained for diagnostics, but it does not determine R quality, risk level, calibration validity, or whether recalibration is required.
How the warning bands work
When R and G are both usable, the engine compares the current load—including an unsent Composer draft—with the room needed for roughly the next one, two, or three typical whole-turn growth reserves. The user-facing states are Lower risk, Long, Near risk, and High risk.
If G is still being learned, LongChat Guard does not invent warning bands; it uses R as the empirical high-risk reference until enough reliable whole-turn samples exist.
Why LongChat Guard uses an empirical reference
ChatGPT's web product does not expose a simple, universal public “conversation remaining” meter that a browser extension can truthfully present as official. A fixed magic threshold would therefore look precise while being weakly grounded. LongChat Guard instead anchors warnings to a user-confirmed local failure reference and the conversation's own measured growth.
This is also why the visual track is a local risk indicator rather than an official OpenAI quota meter.
Measurement compatibility and recalibration
Parser and measurement schema versions define the measurement ruler. If that ruler changes, the previous R becomes a stale prior and a fresh calibration is required before determinate risk is shown again. A model label, when readable, is diagnostic metadata only and does not invalidate an otherwise compatible calibration.
Privacy and retained evidence
Monitoring starts only after affirmative consent. Raw user messages, assistant replies, composer drafts, and attachment body text are not persisted. Local storage contains pseudonymous fingerprints/identifiers, estimates, empirical reference metadata, whole-turn growth samples, and reminder controls. Successful history scans leave no diagnostic record; a failed history scan may keep only a bounded structural diagnostic locally for up to seven days, without raw chat text or the conversation URL.
What this method can and cannot tell you
It can provide a browser-local signal for when a long working thread is becoming risky based on evidence from your own use. It cannot certify OpenAI's exact token count, model context window, remaining quota, billing limit, or a universal conversation maximum.