Razi cells: open directions

Laws Of The Cell: What a cell feels: energy, price, need

Known: cells stop at the first passing answer, pick the cheapest model when they can see prices, and never divide (0 of 12+). Colony shape is not the lever (SP2). Hiding the price list changes the choice, not the reach (SP3).

• Need sensor (SP3) tested or paused, set aside. Tested. One inward number: my body's score against the best my lineage ever recorded. With the price list and the number both shown, reach was highest (2.03 against 1.48 with the list alone), on 4 runs per arm, smallest p 0.21. 6 of 12 guesses held. A lead, on one model family.
• ★ Escalation after a failed try ready to run, tooling exists. A failed round on one model does not count as 'no headroom' until one round on the next model up has also failed. Won in the toy network (GR4, lowest worst-case regret). Never tried in cells.
• ★ Exchange rate: thought price vs win reward ready to run, tooling exists. A mid step costs 50 to 100, a +10% win pays 38, so cheap and early is rational (K18). Move the rate and see if reach moves. Possibly the real lever.
• Rest is not free (upkeep) ready to run, tooling exists. I2a. A resting body keeps paying a small living cost. Today rest is free and safe, which may cost reach (K16).
• Red Queen kit ready to run, tooling exists. I2c. The kit gets harder as the colony's best rises, so a body that stands still loses ground.
• Inheritance by fitness ready to run, tooling exists. I2b. Surplus energy flows in proportion to scored gain, not equally.
• Reproduction as the only store ready to run, tooling exists. I2d. Energy above a level cannot be held, only spent on a daughter. Gives division a cause.
• Daughters born different ready to run, tooling exists. K1/K3. A copy gives the gate nothing to judge (seen in networks and cells). Daughter starts with another model, effort, or slice of the need.
• Machine load as price needs new tooling first. K4. Energy cost per step follows real CPU and memory load, so cells feel the machine through the law, not a prompt line. The sensor-exposure idea from the project lead.

Which Mind: Model families and mixtures

Known: the effect of the variants tool flips with the model (F5). hy4 never attempted improvement. Only Sonnet/Haiku have run on the valid instrument.

• ★ Same laws, second family ready to run, tooling exists. I2e. Run the energy laws on Mistral Large 4 and Argon. No finding counts until it holds on two families (flag rule).
• Mixed colony: strong builds, cheap runs ready to run, tooling exists. Strong model for tool-building steps, cheap for routine. Metric: tokens per improvement, against all-cheap and all-strong.
• hy4 with the hold rule ready to run, tooling exists. F7. hy4 scored 0 because it never tried, not because it could not. Does a hold lift it?
• Ask vs rule, same model ready to run, tooling exists. F3. A nudge did nothing (Argon), a refusal did a lot (Sonnet). Never compared on one model and one kit set.
• Model fingerprints in vector space ready to run, tooling exists. E5. Do different models' code variants occupy separate regions? Does a mixed colony cover more area?
• Argon reruns on the Gemini key needs new tooling first. Cells are not yet wired to the Gemini route. Small wiring job, then the held Argon runs.

What Is Inherited: Memory, reflexes, lessons from losers

Known: one error type recurred 15 times in the same cell. The prose library is read but changes little. Losers' code is archived and unused.

• ★ Division at birth (T-L2) next to run. Each birth is a pair: one daughter keeps the inherited body, the other must change it before it may rest. The gate keeps the better. Follows from L1, where 65 of 70 heirs changed nothing.
• Heredity across lives (L1) tested or paused, set aside. Tested. A body passed down generations, an energy store that pays for births, a gate on the germline. Reach was level with blank newborns (2.27 against 2.28). Heirs spent 31% less energy and died less, but did not compound. 5 of 8 guesses held; the three that mattered missed.
• A need longer than one life (T-L3) needs new tooling first. A need that cannot be finished in one life, so that compounding is the only way up.
• ★ Dormant reflexes ready to run, tooling exists. T-J2, from Weft discussion 105. A cell that recovers from an error leaves a trigger and response. Costs nothing until the trigger recurs. Colony inherits it.
• Nearest lessons only ready to run, tooling exists. E4. Give each cell the top few library entries closest to its current code, not the whole file. None / all / top-k.
• Losers passed on as vectors needs new tooling first. H2b. The direction from losers to winners steers the next attempt (F1 steering, F2 toward/away test).
• Tool commons needs new tooling first. R5/A1. Cells publish tools to a shared shelf. Watch which spread (cultural evolution).
• Seeded library ready to run, tooling exists. Start a colony with the library of an earlier one. Is progress sequential across colonies or reset each time?
• Structured genomes needs new tooling first. Corpus idea (s-expression genomes): memory as structure the harness can cut and recombine, not prose.

Beyond Text: Fitness and signals that are not words

Known: a text-parsed score was once fooled by a warning line (179.0). Bend interaction counts are exactly repeatable on a pinned runtime.

• ★ Deterministic fitness (Bend/HVM) ready to run, tooling exists. G1/H2a. Score by interaction count, not wall time. Removes noise gate, repeats and bench lock. First question: can models write valid Bend at all?
• An evaluator as the need needs new tooling first. G3. The task is a lambda-calculus evaluator, which could later run the next generation. Gains could compound.
• Lean arm vs Bend arm needs new tooling first. G5. Same task, same kit, one colony writing Lean, one Bend.
• Predict before benching ready to run, tooling exists. E3. Nearest-neighbour guess of a variant's score from past variants. Skips likely losers, cuts the serial bench.
• How wide should a split be ready to run, tooling exists. E1/E2. Distance between daughters against gain. A novelty floor so every A/B is a real experiment.
• One basin or many ready to run, tooling exists. E6. Across colonies, do winners for a need collapse to one point or several?

Proof: Selection as mathematics

Known: Lean 4 and Bend are installed. Nothing proven yet. Lean cannot verify Bend code in our stack, only Lean code or a log.

• ★ Needs that are theorems ready to run, tooling exists. R2. Fitness is 'the proof compiles'. No timing noise, no fake wins. First need: the Weft gate invariant (Paper 1, Prop 3).
• Selection rule as a theorem needs new tooling first. R3/F3. State the keep rule in Lean, prove fitness never falls across generations, generate the check from it.
• Proof-carrying wins needs new tooling first. R1. A win is kept only with a certificate: differential test first, Lean equivalence proof later.
• Anchor to generation zero ready to run, tooling exists. W2. Every winner must still match the first body's outputs. Measures silent drift (Weft's compounding-bound point).
• Hash-chained lineage ready to run, tooling exists. W3. Each cell's history as a replayable committed record, like Weft Paper 2.

Growing A Model: The same laws one layer down

Known (toy networks): blind division plus a gate plus gradients sizes a network to its need. A fixed gate cannot tell satisfied from stuck. Tinker works end to end (5 of 5).

• Adapter genome on Inkling 2 Small tested or paused, set aside. T-J3. Cell experience as training data for a small adapter; fork, train two ways, gate keeps the better. Incomplete: the plain arm did not beat the untrained base (test loss 0.904 against 0.674, median of 3 seeds) and the gated arms stopped on a storage limit. Set aside until the plain arm works.
• Sensor proposes, escalation confirms ready to run, tooling exists. GR5. The combined gate law on the toy network. Only if the cell pilots need it.
• When need is far out of reach ready to run, tooling exists. K9. On the hardest toy task no gate wins and ungated growth does. Restraint costs most where need is highest. Open.
• Weight genomes on local models needs new tooling first. H2c. Small local model, genome is a weight delta, a split branches the delta. Direct bridge to Weft Paper 1.
• Do both layers obey the same laws needs new tooling first. H2d/J1. Compare acceptance rate, bold-win share and gate binding between weights and code on matched budgets. The central hypothesis of the program.
• Model reshapes its own chip out of reach today. The FPGA-like layer from the Weft roadmap. No such hardware here.

Colony As Organism: Structure, agency, scale

Known: doubling cells halved wins per 100 calls. A stem that feeds parts costs 1.6x tokens and ends level. Width needs a problem that splits with cheap checks.

• ★ A problem that truly splits needs new tooling first. J2/J7. Width only pays on needs with many unlike parts, each cheaply checkable. We have not yet built such a need. Without it 'swarm vs one agent' is untestable.
• Cells that rewrite the mitosis needs new tooling first. A2/R4. A cell whose body is the runtime itself, scored by the wins of a small child colony. Bench and sandbox stay outside.
• Adversary cells needs new tooling first. R6. One writes code, a sibling writes inputs that break it. The kit grows from the attacker's finds.
• Swarm vs one strong agent needs new tooling first. T-J1 remainder. Colony against a single strong agent at equal tokens.
• Micro-harness swarms, many data types needs new tooling first. J3. Modular harnesses spawning swarms across model families. Needs the splitting need first.
• Stem cell feeding tested or paused, set aside. SP1/SP2. Need-directed feeding emerged without a rule, but ended nowhere better at higher cost. Parked, code kept.
• Mortal law (no rest) tested or paused, set aside. mp3. More attempts, but half the energy went on idling at end of life. Biology keeps rest and adds a wake signal.

Instrument: Things that keep results honest

Known: three instrument faults found and fixed so far (stderr score, cell-owned inputs, own-baseline gain).

• ★ Kits with real headroom ready to run, tooling exists. compress has no headroom and fibmod carries most effects. Results lean on 4 uneven tasks. More kits, each with a known ceiling.
• Benchmark carryover ready to run, tooling exists. G6. Does one cell's bench leave the machine hot for the next? Unchecked.
• Do cells drift their own test ready to run, tooling exists. F1. Count post-assay benchmark edits per model on the old cell-owned protocol. A finding about agents, not just a bug.
• E07 record corrections tested or paused, set aside. Weft PR 101. Numbers recomputed by script. The corrections are not yet written into the record.