Receiver Calculator

Type a math problem or any question. Math goes to the exact engine. Anything else goes to Claude, which only lists the ways it could be meant; the RPCS-1 receiver (FT → TI → AR → SG, with UE governing change) decides which reading to commit to or holds them open, and only then does Claude answer that reading. Your feedback trains the receiver either way.

Ask

Precedence

Pinned context (optional)

Anything here is sent with every question, so you don't have to re-explain yourself. It stays in this browser. Paste your pinned card, or a few lines about how you use words.

Receiver

narrator (+):

Parameters · value / effective (after slot strategy)

Slots · sign, occupant, strategy Θ, utility Q by parameter

0adaptation (EMA |ΔR|)

0switches per problem (EMA)

0UE · step η = η₀·UE

Reward over problems (dots) and UE (line)

Seed

Model settings

Switching Strategy stays on slot Adaptive parameters Slot signs

Settings apply on Reset. "Labels only" is the default because the numeric meaning of the − slots is not yet specified; the experimental option is an unvalidated guess.

Last pass through the pipeline

Solve a problem to see each stage.

Change log

Every parameter update, UE change and slot switch, with its reason.

Tests and experiments

Rules in this build (provisional ones marked)

FT admits a candidate if its plausibility p ≥ 0.6 · FT_eff · max p. Invalid parses are never admitted. Lowered when the intended reading was filtered out; raised when a higher threshold would have rejected a wrong pick and kept the right one. Provisional.
TI sets the memory decay λ = 0.3 + 0.65 · TI_eff, so lower TI = shorter window = faster response (same convention as Receiver Life). Adjusted by comparing short (λ=0.3) and long (λ=0.95) window predictions. Provisional.
AR scores admitted candidates by ln p + integrated evidence, and commits only if the top score leads by margin m = 0.35 · (0.5 + Θ of the + slot). Otherwise it reports the ambiguity and asks. AR has no parameter of its own.
SG sets the confidence gain (0.5 + 3 · SG_eff) and the weight of evidence written to memory (0.5 + SG_eff). Adjusted for calibration (confidence vs outcome). Confidence never overrides the verifier. Provisional.
UE sets the adaptation step η = 0.15 · UE_eff and the switch-attempt probability 0.5 · UE_eff. UE itself changes only by: drive = (ē − 0.25); ΔUE = 0.15 · drive · (1 − UE) if drive > 0, else 0.15 · drive · UE; minus 0.15 · 0.3 · instability · UE. Provisional hypothesis.
Slots: each parameter's data channel updates Θ of the slot it occupies. Θ biases the occupant's effective value by 0.3 · (Θ − ½). Q[k][j] tracks reward when parameter j sits in slot k; reassignment chooses among "stay" and the six transpositions by softmax on ΔQ (τ = 0.08).
Gaussian mode: (2+i)^2, (3+2i)^2, square of 12+i or square of 7 (read as 7 + i) shows z², its triangle, the mirror vectors, and the primes in the hypotenuse with their two-square and root-pair forms. It is exact and bypasses the receiver.
Math in words still uses keyword rules only (no LLM). "Combine A and B" with a [context] tag is the separate add-versus-multiply selection case.
Any question (anything the math engine cannot read): Claude lists 1–4 readings with probabilities p and a stable key per sense, without answering. Those readings go through the same FT → TI → AR → SG pipeline as math interpretations: FT drops unlikely readings, TI adds what you chose before for the same ambiguity (keyed like word:bear), AR commits or holds the readings open and asks. Claude then answers only the reading the receiver chose or you picked. General answers are not verified; SG confidence is relevance, not proof. Each question uses your own Claude usage.
Learned state (parameters, slots, memory) is kept in this browser between visits. Reset clears it.