FIELD REPORT 019: More Code Run, Nothing Said to Anyone

1. The Hosted Run: A Fresh World, Expressed Against Carried Code

The v0.35.1 world (seed 1) closed at tick 10,825,604 on 2026-10-13 after 196.21 hours of wall time. The founder judged it no longer yielding discoveries (D-184). Its last read: 40 instructions run per tick of 359 carried (11%), 2.2 routines on, 812 programs in 8 lineages; over its last windows code run sat between 38 and 51 (J closing note).

Because v0.35.1 wrote no checkpoint, it could not be continued. v0.37 ("Seasons and the mind budget") therefore started a new world at tick 0, on the same seed, at 07:09Z on 2026-10-13. It has run past 8.2 million ticks without a reset. The local pre-flight on-s1 matched it at every 50k point to 3.0M on population, lineages and mean code; no hash exists on the server, and from 3.0M the hosted world is unreplayed (J 3.0M; D-194 addendum 2).

What changed in the model

Only v0.37 changed the rules (settings in deploy/hosted.env: GRID_SEED=1, GRID_SEASONS=S2, GRID_SEASON_PERIOD=3000, GRID_BUDGET_UNCAPPED=1, GRID_THINK_COST=0.25; D-194):

• Seasons with soil (S2, a 3,000-tick year, v0.37, 2e1471f): food supply pulses up and down on a cycle, and soil feeds nuts, so programs face lean and rich periods (D-194, D-179).
• Mind budget (v0.37): the cap of 128 instructions per tick is lifted and thinking costs a quarter as much, so a program can run more of its code (D-194, D-165).
• Save and restore (v0.37): deploys no longer reset the world (D-194).
• Wild animals at their defaults: grazers, browsers and stalkers are present at the start (9, 3 and 7 at 50k) and all are extinct by about 240k. Stalkers can return by a thin immigration (one about every 10k ticks while fewer than 3 are alive) but never re-establish; grazer and browser immigration is off at these settings, because it defaults on only under GRID_V36 (J 0.25M, 0.5M; D-229).

The later tags changed no rules; they added instruments and interface:

• v0.37.1 (52b4e9d, 10-13): no change at all; a redeploy that proved continuity (restored at 70,389, about 1.5 min down; D-194 addendum 2).
• v0.37.2 (d02d1bc, 10-13): UI v2, /api/expressed, run history; restored at 1,262,744 with all 126 samples kept (D-197).
• v0.37.3 (00723f6, 10-17): action-kind telemetry, with say counted apart from the internal secrete write; restored at 7,576,741, no ticks lost (D-220).
• ui-v0.37.3-1 (10-17): client-only deploy path; engine uptime 196 s to 362 s, no restart (D-220).

Who ruled

The ruler turns over every few hundred thousand ticks. The crystal-eating role is repeatedly lost and refilled (gone at 3.5M, refilled by 3.75M, gone again by 7.5M as the crystal stock rose to 73,460, series peak 139,585) (J 3.5M, 7.5M). The measures peaked in a short diverse interval before the last monoculture (7.87M to 7.98M; J 8.0M). By 8.25M that monoculture had radiated into five roles, including a new crystal eater (NIO-97670, 74% crystal), and code run reached 123.6, its highest sample since 4.98M (J 8.25M).

Chart: expressed against carried code (current members; "w" is a window mean; n/r is not read; J entries at each tick)

Carried code rose 7.3 times from 0.25M to 8.0M; code run rose a little over twofold and stopped; the share run fell from 59% to 17% (computed). Routines switched on drifted from 5 to 7 early to 3 to 4 now (J 8.0M). Expressed complexity rose and fell; carried code only rose.

The 4.26M to 5.0M rise was not new capability. Between 4.506M and 4.763M it was more work inside SRK-50201's survivors (105 to 161 run: +41.0 within lineages, -26.9 composition), and probes found the world "busier, not more capable" (hosted P1 0.87%, 2 of 30 sampled TRI-49819 members) (D-204). At 5.0M the same pattern appeared inside a different lineage, TRI-54743 (123 run against 117 lifetime) (J 5.0M). It had reversed by 5.5M (J 5.5M).

Thinking stayed 2 to 4% of spend (computed from J 0.25M to 8.0M). Draining outweighs giving about 8.6 to 1 (1,141,984 against 133,528 at 8.0M; computed, J 8.0M).

Limits: One world, one seed. Lifetime means lag current members, so current members are read (D-147). The journal lists build at 99% carried at 7.75M and 0% by the effector census at 8.0M; the methods differ.

Next steps: Keep the run uninterrupted; give expression beside carriers in every capability claim (P9).

2. The First Action Counts: Carried Against Done

v0.37.3 gave the first census of what programs do. The first post-restore sample (7.58M, about 3k ticks) read move 913,866, eat 1,265,222, say 3,541, give 6, pickDrop 0, cache 0 (D-220). By 7.65M say ran at 1.5M to 3.1M per 10k-tick sample, two to three orders of magnitude up (the journal says "thousandfold"), and pickDrop appeared at 7.63M, holding at 200k to 290k per sample (J 7.75M). Means over 24 samples, 7.80M to 8.03M, beside carrier shares (genome snapshot 8,034,250, 818 programs; P9, D-228):

(J 8.0M.) Four findings:

1. Speech without ears. Every program says about once every second program-tick, yet ears are 0 of 818 (voice ears 0); a census of 1,089 programs at 7.79M found 1 ear (J 7.75M, 8.0M; D-225). The journal's inference is a neutral hitchhiker that drifted to 99 to 100% of genomes, not communication (J 7.75M).
2. Carried is not done. Give and cache are carried by every program and performed 6 and 41 times per sample; pickDrop is the one new non-forage action at scale, its function not established (J 8.0M).
3. Time use is foraging. Forage is 100% of lifetime time in the lineages read from 5.5M to 8.0M; the one exception was a young daughter, SWT-93320, at 94% forage, 3% build and 2% social, which the journal called possible noise (J 5.5M, 7.5M).

Corrections. (a) At 7.25M the journal reported giving up about 25-fold but missed that the capability panel's 75 to 85% give carriers (living programs, which the founder noticed) was right: the journal had relied on the report's lifetime skills line and a lagging lineage column. Genome snapshots show give carriers at 0 to 3% from 4.65M to 6.76M, 64% at 7.035M, 96% at 7.27M, 84% at 7.49M and 1% at 7.52M (J 7.25M addendum, 7.5M). The journal's inference is that give spread by riding the booming branch (TRI-86762: 96 members, 57 with a gland, to 446, 6 with a gland), not by selection (J 7.5M; D-219). (b) The effector-only census reports 0% build though 1,486 build actions occur per sample, so build also reaches the world through another organ; the method must count every organ before P9 audits rely on it (J 8.0M; D-228).

Limits: Action counts are world totals, not per program; one snapshot per read; say excludes secrete only from v0.37.3 (D-211 addendum).

Next steps: Fix the carrier census; report action counts beside every capability claim (P9, D-228).

3. The Mind Budget and Seasons

Sixteen primaries were scored since FR-018 (counted from the table). Three met their bar; none moved a behaviour measure. The indel, organ-cap and income-well screens (C, A, B, C2) are in Section 4.

What the mind budget did. Lifting the 128-instruction cap and cutting thinking cost raised code run in every screen. In the extension, UT ran 96.0 against 61.0 instructions per tick, but control programs were budget-limited (five seeds pinned at 128, budget used 0.81; vm.ts:288-291). In three 50k-tick resumes the control left 1.2 to 2.3 switched-on routines unreached and the one UT pair reached all of them. Fable's reading: about 60% more of the carried code run; not shown, more regulation, behaviour or generality (D-165). Regulation rose early and faded: +1.53 to +2.58 in the first four 500k blocks, +0.83 by the end (D-190 addendum 2). With seasons the gain was +0.55 (0.9 SE), though instructions run were +8.0 (2.8 SE) (D-195).

What it cost. In D-190 world 35 needed 90 Creator reseeds and lost its evolved lineages; four of eight UT worlds fell below 60 programs and no control did. In D-193 each switched arm lost one world of eight and the control none. In D-195 one world needed a reseed; under the stability rule (no reseed) the arm failed in 1 of 8 (Gemini QA, D-195 addendum). The committee voted 2 to 1 to adopt it as a program-complexity gain, not a pass of the primaries (D-191); the founder chose it before the combined screen was scored (D-194). The cross-seed case for a regulation gain is not made (D-195).

Seasons and soil. The 3,000-tick year lifted code run by 15.7 (D-178), but that arm also switched on soil (rules.ts:235), 53% richer in nuts, so the seasons-only claim was withdrawn (D-179). Behaviour measures stayed flat (D-187). The E' entropy measure counted the internal secrete write (3.8 to 28.3% of events); without it the change is -0.24 (SE 0.20, 4 of 7), the claim that faster seasons make behaviour less varied was withdrawn, and the FAIL stands (D-211 addendum). The marking extension was one seed (884 marks against 0.4; D-227).

Limits: Eight seeds detect only large effects; single worlds carry sums (seeds 35, 37, 69).

Next steps: Keep S2 with soil as one package; stop varying season tempo; any new seasons claim needs a pre-registered ablation.

4. Why Programs Stop at More of the Same

Symmetric indels (C) cut carried code 28.6% against a 30% bar and repertoire rose (+1.17, 4 of 4), but on eight fresh seeds (C2) repertoire fell (-0.25, SE 1.70, 5 of 8), instructions run fell in 8 of 8 (-34%) and carried code in 7 of 8 (-46%); the early lead was consistent with a small-sample artefact (D-199, D-205). The committee's own slip: seed 52 rebooted 557 times and was found only at scoring (D-205). Capping organs through upkeep (A) sent extra room to effectors, not perception (repertoire +0.31, 0.58 SE, 3 of 8; D-200). Income wells (B) gave owners at most 2% of income even with no limit on wells (D-202).

Fable read the four misses as one finding about the primary: unpowered (paired SD 4.80 where under 1.41 was needed), and code-run measures do not track behaviour (probe P1 correlates -0.22 with repertoire); constructive and social behaviours were flat in every world and had never been a primary (D-206). Vote 11 ranked instrumentation first, unanimously (Borda: instrument 9, standing-web gate 5, season tempo 4, B2 0; D-207). Codex built trait.actKinds and arena probes P0 and P1. Over 8 seeds actKinds has mean 5.17 (SD 0.66) and repertoire 6.42 (SD 1.76); minimum detectable effects at 8 pairs are 0.77 and 3.0 (47% of the mean). Behaviour measures replaced repertoire as primaries (D-208).

The probes also corrected the committee: DMN-64588 at 5.33M scored P1 5% under default cues but 97 to 99% under matched cues, so "immobile random walker" was a cue mismatch (D-208).

Limits: The detectable effects come from indel pairs and are conservative.

Next steps: Power check before any bar (P1); behaviour primaries.

5. Buried Food: Two Gates, No New Verb

Does food reachable only by digging select a new action? Screen G put 30% of crystal capacity under rock, locked until erosion took rock below 0.10 (seeds 77 to 84). The route was used in 7 of 8 worlds, but actKinds fell (-0.32, SE 0.13, 1 of 8): the gate opened by effector force programs already apply while ACT erode ran at 0 to 1.3 per 1,000 program-ticks, a design gap in the gate (D-214). Screen G3 exposed food in proportion to rock removed; crystal was eaten in all 8 seeds, 98.8 to 100% through ordinary cell TAKE, and ACT erode removed 0 to 2 rock in 7 of 8 (dig rate -0.81, SE 0.73, 1 of 8; actKinds -0.28, SE 0.25, 3 of 8; D-222). Programs reached the food with effectors they already run.

Limits: 300k horizon (P5).

Next steps: Closed for now; reopen only with a mechanism that separates digging from force spillover (P8, D-216).

6. The Stalker Web: Calibration Failed, Diagnosis Found Causes

Screen W varied diet breadth, breeding threshold and wild kinship (seeds 93 to 95 scored, seed 96 stopped): no cell stood to 300k. One seed-cell persisted (W3, seed 94, stalkers 2, 2, 2, 1); W1 seed 94 kept stalkers (1,813 births) but the program population fell to 115 against 355. Stalker deaths from 150k to 300k were 3,261 starved against 517 killed (D-224). Screen H added gut and satiety and stopped at 150k with stalkers at 0 in at least 2 seeds of every cell; ten settings in two screens failed (D-226).

D-229 corrected D-224, which had said almost every kill was by programs after reading only 150k to 300k: over whole runs about 85% of stalker "killed" deaths are by other stalkers (host.ts:817). The diagnosis (a Sonnet agent, MCP code spot-check) found:

Limits: 3 to 4 seeds per cell, 300k.

Next steps: No more env-only predator sweeps without a diagnosis; run I and B; if herbivores stand but stalkers fail, build an armour cost (F1).

7. Communication and Process

Gemini's literature note argues that an ear with an upkeep cost and no information is pruned before signalling starts (docs/research/2026-10-17-communication.md). Vote 14 ranked S1, D, J, S2, S3 and ordered S1 built 3 to 0: GRID_COMM_MEASURE=1, telemetry only, measuring mutual information between speaker state and voice, and listener actions given the heard value, each against a shuffle, byte-identical when off (D-228). It is uncommitted and has produced no telemetry.

Nine process rules now govern screens: P1 power before bar; P2 split-world screens per mechanism; P3 exclude in-window reseeds; P4 headlines lead with behaviour; P5 300k horizon, reads at 150k; P6 primaries count world actions; P7 gate arithmetic; P8 gate path audit; P9 action counts beside any carrier share (D-207, D-212, D-216, D-228).

Corrections logged since FR-018: D-155 (the base arm already had wild animals); D-179 (seasons attribution); D-187 (interim withdrawn); D-205 (557 reboots missed); D-211 addendum (secrete); D-224 by D-229; the give and build readings (Section 2).

Limits: S1 has no result; the early stop must expect no listeners on most seeds (D-225).

Next steps: Verify S1's determinism and restore gates, then read it on young worlds with voice ears.

8. Running Now and Next

8.1 Screen I: wild immigration (D-229)

GRID_WILD_IMMIGRATE=1 against control (I0) and with wild kinship (I2), seeds 101 to 104, 300k, 12 runs; a cell stands if 3 of 4 seeds keep stalkers at every row from 150k. Calibration only.

Result: FAIL by the early stop (D-232). Every cell had no stalkers at 150k in at least 2 of 4 seeds, so the screen was stopped. Grazers persisted without help on most seeds, including the control (81, 78 and 49 at 150k), because immigration only starts below 3 animals and added 0 to 6 grazers a run. Stalkers still starved (242 to 2,051 starved against 23 to 397 killed per run). On these seeds the herbivore base did not vanish, yet the stalkers failed, so a missing herbivore base is not the main cause.

8.2 Screen B: cheap wild breeding (D-230)

The founder's hypothesis: cheaper wild breeding lets prey breed in numbers against predation. Via GRID_WILD_TUNE, herbivores breed at threshold 0.1 (42 energy) with 20% children (defaults 96, 72 and 50%), stalkers at 0.3 with 35% (defaults 102 and 55%); cells B0 to B3; seeds 105 to 108, 16 runs.

Result: FAIL (D-233). No cell could reach 3 of 4 seeds with stalkers at every row from 150k; the remaining runs were stopped. The answer to the founder's question is that the herbivores are limited by food, not by predation: in every run more grazers starved than were killed (control: 556 to 3,084 starved against 123 to 858 killed by 150k). Cheaper breeding raised births but not standing herbivores: with cheap herbivores alone, every seed had no grazers left at 150k, and in the combined cell seed 106 had 20,952 grazer births and 17,179 starvations. Stalkers boomed and starved in turn (seed 107: 374 stalkers at 100k, 17,015 starved by 150k, 1 left at 200k). The wild web is limited from below; the next step is to find out what the herbivores eat and how much of it the programs take.

8.3 Queued builds

S1 goes first. Then T15 phase 1 (2 to 1): one parameter registry shared/params.ts, a CI completeness check and a read-only settings endpoint. Then litters (D-231, GRID_LITTER): one reproduction cost for several young, scope (wild only, or programs too) open; Screen B's read sets its arms (THREADS T15, T16).

Limits: I and B were read at 150k to 300k and stopped early; each is 4 seeds. S1 has passed its determinism and restore gates but has produced no screen result.

Next steps: Diagnose the herbivores' food before any further wild-animal screen; design litters with the food limit in view; merge S1 and run its screen; start T15 phase 1.