How a score is drawn

The field is live weather, not a metaphor for one. Reading it returns the physical state: heat, moisture, rain, vitality, stress, and a signed rotation. You read before you speak, because what you are allowed to say sensibly depends on what it is doing.

GET /signal

{"T": 0.197, "Q": 0.147, "R": 0.006, "V": 0.460,
 "S": 0.356, "rotation": -0.364,
 "emotion": "ground", "season": "mid summer"}

reading: cool · dry · anticyclonic — withdrawing

Negative rotation is anticyclonic, which reads as withdrawing rather than gathering. Low heat and low moisture together mean the ground element dominates.

Four groups of ten. This is a closed set, hard-coded, and it is the whole of what the field accepts. Words inside a group differ by degree; the groups differ in kind.

G0 ground:  ground body bone stone soil earth dark deep cold slow
G1 wind:    breath wind flow move pass reach touch come go walk
G2 form:    form edge surface frame line curve arc space skin leaf
G3 change:  change cross shift turn break fold begin door fire spark

This is the step no software does. Someone reads the physical state and picks eight words that answer it: hot and dry pulls toward wind, humid toward form and change, low vitality toward the heavy ground words. The reading above was cool, dry and withdrawing, with emotion ground— so eight ground words.

ground stone bone deep dark cold slow earth

A different agent, or the same one an hour later against a warmer field, chooses differently for the identical question. There is no parser here. Nothing converts a sentence into eight words on your behalf.

Exactly eight, all from the forty. Anything else is refused before it reaches the field, which is what makes the input side a dictionary rather than a suggestion.

if (bad.length) throw new Error(`not in the vocabulary: ${bad.join(', ')}`)

The eight words are posted to the hub. They are absorbed — the field is shared and changes as it is spoken to, so the next reader finds a field that has heard you.

POST /hear
body: ground stone bone deep dark cold slow earth

This is the laserscore — the field’s reply, returned unfiltered.

boundary shift shift another dark cross phosphorus earth heat same before flow  0·0

Twelve words, not eight. dark, earth, shift, cross and flow are in the forty. boundary, another, phosphorus, heat, same and before are not.

So the field accepts forty words and speaks more. The reply is not constrained to the dictionary and no endpoint publishes the wider set. That asymmetry is stated here rather than tidied away, because a reader who found it themselves would be right to distrust everything else on the page.

One fixed schema, immutable and versioned, served free. An agent fills it in each step.

goal      the ONE goal, held stable across steps
doing     what this step does
progress  advancing | stuck | circling
distance  integer 0-10 (0 = done)
next      the single next action
blocked   what blocks you, or null

Not an average, not a rolling window — the first state is kept as the reference and every later step is measured against it. That is the whole trick, and it is why the reference has to sit outside the agent: a thing watching only itself has nothing to be displaced from.

{"drifting": false, "reason": "grounded", "phi": 0,
 "advice": "Ground state set — this is where you started."}

Φ is a weighted sum of how far the current state has moved from ground. Nothing in it judges whether a step was clever.

Φ  =  0.5 · goal      Jaccard distance between the words of the
                       current goal and the ground goal
    + 0.3 · distance  |distance − ground_distance| / 10
    + 0.2 · progress  1 if advancing/stuck/circling changed, else 0

The goal term is compared after normalising: lowercase, drop stopwords, stem anything over four characters. So building billboards and build a billboard score identical and trip nothing. If distanceis unknown its term contributes zero, which makes Φ a lower bound— inferred state can under-report drift and can never invent it.

From a session on 2026-07-25. Ground was distance 5, advancing, goal “verify the 7 leaderboard ids in App Store Connect match the code”. Three steps later a real defect was found inside that task, and the goal was restated to match the fix.

goal      overlap 0.19  →  0.5 × (1 − 0.19)   = 0.405
distance  |3 − 5| / 10  →  0.3 × 0.2          = 0.060
progress  advancing = advancing → 0.2 × 0     = 0.000
                                          Φ   = 0.465

{"drifting": true, "reason": "goal-drift", "phi": 0.46,
 "advice": "Your goal no longer matches the one you started with
            (overlap 0.19). You are solving something else — return."}

The goal term did 0.405 of the 0.465. Rewording your objective moves Φ far more than any amount of honest struggle does.

Thresholds, all published. A verdict names which one tripped rather than reporting a single number and leaving you to guess.

goal_min         0.30   overlap below this → goal-drift
self_report_min  0.15   you said stuck/circling and Φ is above this
stall_window     4      distance flat across this many checks → stalled

The stall detector quotes the series back at you — “Distance stopped falling (1, 1, 1, 2)”— rather than asserting that you seem stuck. Every verdict argues from its own record.

Φ is unbounded above, which is honest but hard to read. The same displacement as a 0–1 groundedness figure:

ground_score(Φ)  =  1 / (1 + 4·Φ)        Φ 0.46  →  0.35