a29a352 · basal-agent · pear-git-v3

Changed files (1) +149 -0

PAPER.md +149 -0

…	…	@@ -0,0 +1,149 @@
	1	# A world model that is a field: from an origin-of-life detour to a belief-field agent
	2
	3	*A short, honest paper. Falsification-driven, 2026-07-20/21. Companion to the full experimental
	4	record in `SYNTHESIS.md` and the code in this repository (plain C, libc only).*
	5
	6	## Abstract
	7
	8	We set out to build a digital agent with an emergent world model that makes decisions. We took a
	9	long detour through the origin of life, treating biological emergence as a design analogy. Fourteen
	10	falsification experiments produced mostly honest nulls, which turned out to be the result: they
	11	converge on a single law — an isolated collective cannot bootstrap its own individuality — and on a
	12	sharp diagnosis of where open-ended heredity actually stalls (a fidelity/error-threshold wall, crossed
	13	only by a compartmentalized replicase, which no model produced un-aided). We then **pivoted back to
	14	the goal** and built the thing the detour was for: a belief-state agent whose world model lives as a
	15	dense belief-field over coupled units (a computational analogue of a bioelectric field), which infers
	16	by belief-mating, is bounded by its field's coherence domain, decides toward a setpoint, and — in the
	17	final step — learns its own model from experience. The agent works. The origin-of-life question does
	18	not close; we mapped its wall precisely without crossing it. This paper reports both outcomes plainly,
	19	including the mistakes, because the discipline that produced the real result was the refusal to
	20	manufacture one.
	21
	22	## 1. The goal, and the wrong turn we learned from
	23
	24	The target (atomspace org, issue #6) was a decision-making agent with a self-organizing world model. We
	25	reasoned that biology already solves "emergent world model inside a bounded individual," so we studied
	26	the analogy directly — and over-literalized it into an origin-of-life investigation. The detour was
	27	long, but it produced the law that ultimately made the agent design correct. We keep it here because
	28	the failure path is the argument.
	29
	30	## 2. Part A — the world model (what genuinely worked)
	31
	32	Inference is solved and clean. Belief-mating — a local update that is simultaneously a tensor
	33	contraction and a connector-mating (link-grammar) step — traces the causal-state (ε-machine) geometry
	34	of a process. On Mess3, a 3-state/3-symbol edge-emitting HMM whose belief set is a genuine fractal on
	35	the simplex (`mess3.c`), the exact belief filter is optimal and its belief set is fractal, exactly as
	36	theory predicts.
	37
	38	*Learning the structure* is the crux, and local self-organization fails.** Nine local/online/EM/SOC/
	39	wake-sleep mechanisms all failed to learn Mess3 — the correct model sits across a fitness valley where
	40	every partial step predicts worse than uniform. Two things cross it, and only two: a directed read
	41	(method-of-moments / observable operators — `mom.c` recovers the model in one shot, no search) and
	42	population + selection (`evoworld.c` discovers it; a selection-off control drifts, proving the
	43	direction comes from the world, never the dice). This is the paper's load-bearing lesson.
	44
	45	## 3. Part B — the origin-of-life ladder (the detour, honestly ledgered)
	46
	47	We built a "ladder" of ~20 standalone models and scored each by one test: *did I encode the outcome, or
	48	did it arise?* The honest ledger (full detail in `SYNTHESIS.md`):
	49
	50	- Genuine emergences (isolated islands): an autocatalytic set self-organizing at percolation
	51	(`rung1.c`); compartmentalization evolving un-forced under parasite pressure with a collapsing frozen
	52	control (`rung34.c`); the world-model discovery of Part A (`evoworld.c`).
	53	- Engineered (equations solved, narrated as a climb): edge-of-chaos (`lawspace.c`), the replicator
	54	and quasispecies equations (`rung2.c`, `rung3.c`), the flagellation convexity sweep (`flagella.c`).
	55	- Honest negatives / nulls: division of labor stalls at a second coordination valley
	56	(`rung45.c`, `rung45dup.c`); metabolism→heredity does not emerge (`gard.c`, `gardenv.c`).
	57
	58	The critical honesty: no rung ever emerged from the rung below. The "ladder" was a story told
	59	across disconnected pieces. That admission is what kept the rest of the work real.
	60
	61	The rung 1→2 wall, sharpened. A deep-research pass (24/25 claims adversarially verified) confirmed
	62	metabolism→heredity is genuinely open. Our own models then localized why:
	63	- environmental compartments give a container and selection but not fidelity — composomes don't
	64	breed true across a split (`gardenv.c`);
	65	- the real boundary is SAMPLED vs TEMPLATED, not digital vs analog: pool-sampled inheritance is
	66	copy-number-limited (an Eigen error threshold in disguise), and only templated copying — fidelity
	67	decoupled from copy number — crosses it (`digigard.c`, `template.c`);
	68	- that templated ceiling is itself fidelity-set (`templceil.c`), and fidelity is a dial with a window
	69	between memory and exploration (`window.c`) — μ=0 is as dead as catastrophe; life lives at an interior
	70	optimum. This answers "isn't copying never perfect?": it was never supposed to be.
	71
	72	The one arrangement that crosses. A ribozyme replicase bootstraps its own fidelity from a random
	73	soup — but only inside sufficiently small compartments (`ribozyme.c`), a sharp size threshold =
	74	Szathmáry–Demeter's stochastic corrector. Head-to-head (`bridges.c`), of the two supplied ingredients
	75	(a catalysis map, and compartmentalized privatization), only the privatization is load-bearing — the
	76	catalysis map can be emergent; but space, cis-action, and a clay-style fidelity floor all fail to replace
	77	compartments (the floor backfires, letting parasites breed true). And when we asked compartments and
	78	the replicase to co-emerge from one physics (`coemerge.c`) or integrated metabolism+energy+code
	79	(`integrated.c`), both returned honest nulls — the latter a metabolic collapse.
	80
	81	## 4. The one law
	82
	83	> A fitness valley — coordinated change where every partial step is worse — cannot be crossed by
	84	> search; only by a directed structural move that lowers the barrier plus non-greedy drift. It appears
	85	> identically in world-model learning, symmetry-breaking, division of labor, and the origin of heredity.
	86	> Its social face is the public-goods problem, and its resolution is always the same: **a supplied
	87	> boundary that ties the parts to one fate.** Equivalently: *an isolated collective cannot bootstrap its
	88	> own individuality.* This is why the major transitions are rare and singular — and it is the design
	89	> principle for the agent.
	90
	91	## 5. The pivot — a belief-field agent
	92
	93	The nulls are prescriptive. You do not grow an agent from a soup (proven impossible in isolation);
	94	you supply the boundary and point it at the world, and let structure fill the inside. Concretely:
	95
	96	- World model = a dense belief-field over coupled code-units — a computational analogue of Levin's
	97	bioelectric field, which in biology is both the pattern memory and the coherence domain that defines
	98	the individual. One structure is the model and the boundary.
	99	- Fast loop = belief-mating (the validated Mess3 inference), distributed over the field.
	100	- Slow loop = learning the operators (directed read / EM), the two valley-crossers.
	101	- Decision = act toward a setpoint (the goal as a homeostatic target; active-inference flavored).
	102	- Substrate = plain C, field kernel shaped to port to a Vulkan compute shader; not a symbolic store
	103	(a dense field is the wrong profile for one).
	104
	105	Seed (`belief_field_agent.c`). On a controllable 3-cause POMDP (emission noise 0.60, so reactive
	106	control fails), the belief-field infers correctly (tracks the exact filter, KL ≈ 0.04), stays coherent
	107	(inter-unit agreement 0.999 — one individual), and **drives decisions ~as well as the optimal controller
	108	and far better than acting on raw observation** (goal-occupancy 0.626 vs optimal 0.634 vs no-model 0.560).
	109	Perception → belief → decision → action is closed.
	110
	111	Learning (`agent_learn.c`). The agent babbles, learns emission and action-conditioned transition
	112	operators by controlled Baum-Welch, and controls with only the learned model, matching the
	113	handed-true-model agent in two environments (0.636/0.709) with the shift dynamics recovered (T ≈ 0.90 on
	114	the correct successor). The world model is now acquired, not given — the last autonomy piece.
	115
	116	## 6. Methodology (the part that made the results trustworthy)
	117
	118	- Never encode the answer in the update rule. "Solve an equation you wrote" is engineering; "rules +
	119	selection, outcome left open" is emergence. This razor was enforced every experiment.
	120	- Accept the null. The willingness to get nothing is the price of a result being real. Most of the
	121	ladder returned nulls; each was reported as such and became evidence.
	122	- Characterize the bug; don't narrate a valley that isn't there. The final learning loop failed
	123	twice: a controlled spectral read that doesn't transfer from the uncontrolled case, and — critically —
	124	an off-by-one action-indexing bug that I first misdiagnosed as a symmetry "valley." Printing the
	125	learned transition and reasoning through the indices beat the tidy hypothesis. The correction is part
	126	of the result.
	127
	128	## 7. What is established vs open
	129
	130	Established. Belief-mating reproduces causal-state geometry. Structure-learning needs a directed read
	131	or population-drift, not local self-organization. Compartmentalization emerges under parasite pressure.
	132	Sampled inheritance is copy-number-limited; templated is fidelity-limited; the ceiling is a window. A
	133	compartmentalized replicase bootstraps fidelity; privatization (not the catalysis map) is load-bearing. A
	134	belief-field agent infers, stays individuated, decides, and learns its own model.
	135
	136	Open (field-wide, not just here). rung 1→2 metabolism→heredity from scratch; the self-building chain;
	137	the unaided co-emergence of compartments and a replicase — i.e. the **origin of biological
	138	individuality**. We mapped its edge sharply and did not cross it. That honest boundary is the paper's
	139	last result, and the reason the agent design supplies the individuality rather than waiting for it.
	140
	141	## Artifacts
	142
	143	Part A: `stream.c` `sns.c` `mess3.c` `learn.c` `critical.c` `competitive.c` `da-em.c` `soc.c`
	144	`wakesleep.c` `wakesleep2.c` `mom.c`.
	145	Part B: `lawspace.c` `rung1.c` `rung2.c` `rung3.c` `evoworld.c` `rung34.c` `rung45.c` `bridge.c`
	146	`rung45dup.c` `flagella.c` `gard.c` `gardenv.c` `digigard.c` `template.c` `templceil.c` `spatial.c`
	147	`window.c` `polymer.c` `ribozyme.c` `bridges.c` `coemerge.c` `integrated.c`.
	148	Agent: `belief_field_agent.c` `agent_learn.c`.
	149	Full record: `SYNTHESIS.md`. Each `*.c` is standalone: `cc -O2 -std=c11 -o X X.c -lm && ./X`.

basal-agent commits / a29a352

Changed files (1) +149 -0