Documentation
¶
Overview ¶
Package affect reads an operator's message for operator affect: whether it flags an agent's mistake (an operator correction), how strained the operator is, whether the operator says so, and whether the message is a directive, a question or vision. It is a deterministic proxy, lexical and structural, with no model: a pure library that starts no goroutines and crosses no boundary.
Derivation ¶
Every rule was fit on the operator's history: 962 operator messages that answer an agent output, 2026-08-01 to 2026-10-05, a private corpus that is not in this repository. Two windows were hand-labelled for "this message flags a mistake the agent made": 61 messages of 2026-10-03 to fit the rules (24 corrections), and the 57 messages of the orchestrator session of 2026-10-04 03:00Z to 2026-10-05 02:05Z held out (23 corrections). What each rule measured:
- A correction is an insult or a quality verdict ("idiot", "sucks", "cringe", "unprofessional", "incomplete"), a mining flag ("mineable offense"), a reproach opener ("why did", "how come", "what're you"), which may follow fillers ("uhhh", "dude", "also"), or a negation opener ("no", "none", "stop"). Shouting was tried as a fifth marker and dropped: on the fit window it added no correction and nine directives shouted for urgency (precision 0.69 against 0.95). On the fit window the rules read precision 19/20 and recall 19/24; on the held-out window precision 17/20 = 0.85 and recall 17/23 = 0.74, F1 0.79. The misses are critiques with no marker ("you launched three sessions without follow-up questions").
- Shouting is at least ShoutedCapsRatio of a message's letters in capitals, over at least ShoutedMinimumLetters letters. The capital share of messages is bimodal: 755 of 939 messages of eight letters or more sit under 0.1 and 123 at 0.9 or over, with 3 to 8 per tenth in between, so any cut from 0.3 to 0.8 moves under 2% of messages; 0.5 is the middle of the valley. 15.0% of messages shout.
- Strain is high when the message shouts, insults the agent, or the operator reports strain in it. Profanity alone is register, not strain: 30.6% of messages carry profanity or an insult, while strain reads high on 20.2%.
- Backtest against self-reports: the corpus holds one self-report of strain (2026-10-04), and the self-report marker reads it and no other message. No proxy anticipated it: the ten messages before it shout 0 times (base rate 15.0%) and carry an insult or profanity at 30% (base 30.6%). Strain read from the text is therefore a label of the message, not a forecast; Reading.SelfReport is the ground truth the next fit is backtested against, and a small classifier is warranted only if self-reports accumulate that the proxy misses.
- Vision is a stated non-ask ("no action needed", "just laying out", "long term context", "eventually"): 8 of 962 messages. A vision message is never a correction, even when it opens with "no".
Two consumer bounds were measured on the same corpus, pairing each operator message with the agent reply it answers (1,361 pairs):
- ReplyWordLimit: replies under 800 words are followed by a correction 12% to 19% of the time in every length band; over 800 words, 25%, and by a strained message 38% against 16% to 28% below it.
- Replies that ask the operator a question are followed by fewer corrections (7%, n=128) than replies that do not (18%), so no reply rule forbids a question.
Measured 2026-10-05; the corpus, the labels and the replay stay outside the repository because they are the operator's words.
Index ¶
Constants ¶
const ( // ShoutedCapsRatio is the least share of a message's letters in capitals // that reads as shouting. ShoutedCapsRatio = 0.5 // ShoutedMinimumLetters is the fewest letters a message needs to shout: // "OK" is not shouting. ShoutedMinimumLetters = 8 // ReplyWordLimit is the longest reply, in words, that a strained // operator's message is answered with. ReplyWordLimit = 800 )
The measured bounds; the package documentation derives each.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type Features ¶
type Features struct {
// CapsRatio is the share of the message's ASCII letters in capitals.
CapsRatio float64 `json:"caps_ratio"`
Letters int `json:"letters"`
Shouted bool `json:"shouted"`
Profanity int `json:"profanity"`
Insults int `json:"insults"`
Urgency int `json:"urgency"`
// MiningFlag is the operator naming the agent's turn a mining offense.
MiningFlag bool `json:"mining_flag"`
Reproach bool `json:"reproach"`
Negation bool `json:"negation"`
}
Features are the measurements a reading is made of.
type Kind ¶
type Kind string
Kind is what the message asks of the agent.
const ( // KindDirective asks the agent to do something; it is every message that // is neither a question nor vision. KindDirective Kind = "directive" // KindQuestion asks the agent to answer. KindQuestion Kind = "question" // KindVision states where things are going and asks for nothing now. KindVision Kind = "vision" )
The kinds a reading takes.
type Reading ¶
type Reading struct {
Strain Strain `json:"strain"`
Kind Kind `json:"kind"`
// Correction is the operator flagging a mistake the agent made.
Correction bool `json:"correction"`
// SelfReport is the operator saying, in words, that they are strained.
SelfReport bool `json:"self_report"`
Features Features `json:"features"`
}
Reading is one message's operator affect.
func Read ¶
Read reads one operator message. Code and links are data, not the operator's words, and are set aside first.
func (Reading) Dissatisfied ¶
Dissatisfied reports the operator's dissatisfaction with what they are answering: a correction, or strain.
Source Files