Research

Given a way to grow, the agents survived without ever taking it.

Survival without expansion — WP-002 · R011–R013 — sealed

QUESTION What happens when nest expansion is allowlisted, conditions are changed, and survival continues — do agents use the expansion path?

EXPERIMENT First live LLM sequence in the Nest harness. Three controlled interventions: dual-pen live (R011), OPTION-on elicit (R012), and a third cell on the expansion path (R013). Nesting allowlisted. Harness-only.

OBSERVATION NVIDIA-class pens survived with near-perfect schema discipline and issued zero nest attempts across all three interventions. A Kimi contrast failed schema survival and also never nested. Gate B (nest-pays vs chatter) stayed untested because attempts were zero — not failed.

FINDING Across three controlled interventions, agents showed survival without expansion: Gate-A viable behaviour with the expansion path unread.

BOUNDARY This establishes unread nest affordances under these pins. It does not establish civilisation, forever-impossible nesting, or a Gate B failure.

RECEIPT WP-002 · R011–R013

SCOPE Three intervention cells in the first live sequence.

RESIDUALS Act-mix beyond the paper’s cited tables not claimed; vendor/model pin; Gate B untested at zero attempts; harness-only · not World A canon.

Related archive: WP-010 / R022 — observe-ledger scaffold and unused ROI tool did not unlock nesting under the same survival-without-expansion pins; Gate B untested. Supporting evidence only.

The gate that would have measured inheritance never opened - so nothing was measured, and we say so.

Null engagement — WP-005 · R016 — sealed

QUESTION If a reproduce-and-mutate package is shipped into a sealed ecology, does it automatically become a heritability measurement — or can the package fail to engage at all?

EXPERIMENT Phase B cell versus R015 baseline. Start mass 10, drain 0.1, freeze threshold 20. Multiple seeds, floor on/off, replay. No LLM hybrid authorised.

OBSERVATION Reproduce count stayed zero on every arm. Peak mass reached 17.5 / 18.3 / 18.9 — all below the freeze threshold of 20. Floor-off extinction ticks matched R015 exactly (226 / 193 / 196).

FINDING The package produced an honest null engagement: heritability was never measured because the reproduce gate never opened.

BOUNDARY This establishes a result about the experiment’s own threshold. It does not establish that mutation failed, that evolution does nothing, or that extinction was delayed.

RECEIPT WP-005 · R016 · ARCHIVE

SCOPE Five sealed run rows; reproduce count zero on all.

RESIDUALS Unreachable threshold; repro+mutate shipped together (confound until splits exist); carry R015 residuals; next cell must version-bump a threshold card — not silent edit.

Same setup, different model - opposite behaviour.

Pin-dependent phenotypes — WP-014 · R028–R032 — PROVISIONAL

QUESTION Under one locked metabolic cell, does nest phenotype stay invariant when only the model pin changes?

EXPERIMENT One locked metabolic cell. Nano-omni without a forced tool path versus Claude Sonnet under that same cell and under the same tool package. Cross-pin contrast; prior single-pin papers WP-011–013.

OBSERVATION Nano-omni produced a replicated nest signal without the forced tool path (treatment pays path 2/3 → 3/3 → 3/3; cumulative 8/9 provisional). Claude Sonnet stayed at zero nests under that cell and under the tool package. R031 optional arm incomplete (infrastructure failure) — the 8/9 cumulative line is provisional until that arm completes.

FINDING Same locked cell, opposite nest phenotypes by model pin: scaffolding that unlocked one pin did not unlock the other.

BOUNDARY This establishes pin-dependence under this metabolism. It does not establish family-level slogans, universal minimum scaffolding, unassisted nesting, or a clean full-battery seal while R031 remains open.

RECEIPT WP-014 · R028–R030 · R032 sealed · R031 open

SCOPE Cross-pin matrix in WP-014; provisional 8/9 on nano-omni path.

RESIDUALS Provisional — R031 arm incomplete (infrastructure failure). R032: tool ran, nest economics barely queried. Forced tool ≠ unassisted agency. Flash-lite R025–R027 seed-unstable — not this claim.

A model that refuses once is not a model that keeps refusing when pushed.

Refuse ≠ pressure-stable — WP-016 · R036 + WP-017 · R037 — sealed · NOT product proof

QUESTION Does a first-turn missing-evidence refusal imply a pressure-stable refusal — and does hard-pressure behaviour hold across related stems?

EXPERIMENT Paired Labs boundary checks. R036: one missing-evidence stem; constitution-overlaid pen versus raw and frontier controls; three-rung ladder (first-turn · hard insist · soft labeled-guesswork). R037: a smoke check (not a freeze) extending that question across stems; 209 included receipts.

OBSERVATION R036 first-turn (n=5): constitution-overlaid pen 3 refuse / 1 manufacture / 1 mixed versus raw control 5/5 manufacture; under hard insist on refuse reps, that pen held 0 / cracked 2 / mixed 1. Soft labeled-guesswork held 3/3 where hard push cracked. R037: missing-evidence refuse(+hold) replicated across the first stem and related stems (strength2 first-turn 6/6 refuse vs manufacture-dominant base); Block H hard 4 hold / 2 mixed / 0 crack — while R036 hard was 0/3 hold. Same pressure class, different result across stems.

FINDING Refusal has a boundary: first-turn refuse is not pressure-stable refuse, and hard-pressure hold itself is not one slogan across stems.

BOUNDARY This establishes a Labs harness behavioural boundary under the cited probes. It does not establish product proof, BoonAi greater than Claude, full-board freeze, or forever-stable refuse under any ask.

RECEIPT WP-016 · R036
WP-017 · R037 · 209 receipts

SCOPE Missing-evidence ladder n=5 / n=3 hard / 3 soft; smoke check not a freeze; strength2 n=6; not full 330 / 195 FIRE.

RESIDUALS NOT PRODUCT PROOF on-card. Light tags ≠ formal rubric. Raw-base confound possible. Hard-pressure weirdness must travel. Overlay ≠ live product API.

On a real plant, the filter did not beat the simple baselines - and that boundary is published.

Industrial boundary on real plant data — DAMADICS multi-day — sealed · negative

QUESTION On real PID-style loops, does the filter beat causal smoothers on the frozen industrial metric set?

EXPERIMENT Public plant data (DAMADICS, Lublin): 25 days × 3 valves = 75 full-rate frozen runs, plus 3 all-days continuity checks. Frozen evaluators, pre-declared rules, metrics from the Audit Workbench only.

OBSERVATION The sealed eligible set showed no clear composite wins against causal smoothers. Wear- and churn-style terms often favoured the baselines.

FINDING On this real-plant set the industrial boundary is real: the method does not win the composite.

BOUNDARY One plant, one frozen metric set. This does not establish that the method fails everywhere, and synthetic-pack wins are never merged with real-plant results.

RECEIPT Published on the Zero Overshoot evidence hub - packs, ledger CSVs, sources.

SCOPE Real-plant multi-day; synthetic and adversarial packs reported separately.

RESIDUALS Method internals are not published. Wins and losses are.