# Operational Organism-Intelligence Causal Audit **Date:** 2026-07-10 **Status:** preserved V3 predecessor; superseded as current evidence by the verified [V4 Phase-A report](./ORGANISM-INTELLIGENCE-V4-RESULTS-2026-07-11.md) **Binding design contract:** [ADR-0013](../adr/0013-operational-organism-intelligence-2026-07-10.md) **Machine-readable result:** [`organism-intelligence-causal-benchmark-v3.json`](./assets/organism-intelligence-causal-benchmark-v3.json) (`contentSha256` `96495964f03eca1860ba7f869cbd35e929bbdfdd75f801eb4479e6239dfd1b9f`) ## Verdict V3 uses a fresh deterministic 30-seed family disjoint from V1 and V2. It is a fixed evaluation set, not an externally preregistered or untouched validation set. On this family: - the synthetic `+x` goal controller beats exact legacy by `0.009081553179910407`, with bootstrap 95% CI `[0.008224136063586919, 0.009901016159411519]`; - the corpus-conditioned controller beats the goal-preserved substrate-disabled control by `0.008926803443918892`, with CI `[0.0074709186839338375, 0.010524275559338788]`; - reversal adaptation improves by `6.121266720602802%`, clearing the script-declared `5%` threshold; - all `17/17` integrated repository rows remain causal on every evaluation seed, all `5/5` excluded rows remain exactly inert, 10,000 forced numerical-fault steps remain finite and bounded, all ten named living-system classes have matched counterfactuals, and all 30 performance batches across three fresh processes remain under `3 ms` incremental cost. The complete claim gate still fails. Enhanced behavior does not separate from the uniform random-action baseline; rotating all four exposed aggregate semantic channels does not establish mapping specificity; replacing the final exploration signal with a uniform surrogate is also tied and is not an entropy-matched or physical-quantum control. Therefore V3 authorizes no additional A-Life, capability, consciousness, or sentience score uplift. The statistical task is a synthetic goal-field response—not a flora/resource simulation. This is bounded deterministic task evidence—not evidence of general intelligence, phenomenal consciousness, sentience, physical quantum entropy, quantum advantage, or cryptographic security. ## Fixed V3 protocol - 30 fresh deterministic evaluation seeds, disjoint from V1/V2, plus 10,000 paired bootstrap samples. - The seed family is fixed in the benchmark source but was not externally preregistered. - Synthetic-goal controls: goal-preserved substrate disable, exact legacy/no-goal context, a cyclic rotation of all four exposed aggregate semantic channels, a uniform replacement of the final composed exploration value, and a uniform random-action baseline. - The aggregate rotation is not described as a 22-repository permutation because the live consumer API exposes four aggregate channels. The uniform exploration surrogate is not described as entropy matched. - Goal task: one-step x-axis response to the same synthetic `+x` goal field; no flora/resource environment is constructed. - Adaptation task: 240 steps with the synthetic goal direction reversed after step 120; online trace versus frozen trace. - Corpus causality: every external row is ablated separately on every evaluation seed. - Numerical safety: 10,000 forced steps with NaN, positive/negative infinity, and extreme finite inputs. - Performance: three fresh processes, each with ten batches of seven samples per branch at 50,000 entities; branch state is isolated and first-run order is counterbalanced. The worst shared-field p95, median process increment, and all-30-batch stability are reported. - Consumer coverage: the receipt executes and records the exact ten-file targeted Bun test gate; only full-class matched counterfactuals pass. ## V3 results | Test | Exact measured result | Gate | | ---------------------------------------- | ----------------------------------------------------------------------------------------------------------------- | ------------------------------------ | | Enhanced synthetic-goal response | mean `0.013675299144226055` | context | | Goal-only synthetic response | mean `0.004748495700307164` | context | | Exact legacy response | mean `-0.0043330574796032395` | context | | Uniform random-action response | mean `0.009705811435589566` | context | | Goal-only − legacy | `0.009081553179910407`; 95% CI `[0.008224136063586919, 0.009901016159411519]`; dz `3.7930432385228356` | **pass** | | Enhanced − goal-only | `0.008926803443918892`; 95% CI `[0.0074709186839338375, 0.010524275559338788]`; dz `2.0584821748050905` | **pass** | | Enhanced − random | `0.003969487708636489`; 95% CI `[-0.010601483800813, 0.0189792125751305]`; dz `0.09414009637952211` | **fail/tied** | | Enhanced − aggregate rotation | `-0.000029724700357763793`; 95% CI `[-0.00034302817854577695, 0.0002708306663364767]`; dz `-0.033895147462430615` | no aggregate-mapping-specific uplift | | Enhanced − uniform exploration surrogate | `0.00017427102104637573`; 95% CI `[-0.0005910144811096056, 0.0010254502857670654]`; dz `0.0754619380605088` | no surrogate-specific uplift | | Post-reversal median | enhanced `1.2743610106013112`; frozen `1.2008535611966096`; relative `0.06121266720602802` | **pass** (`6.1213% > 5%`) | | Post-reversal enhanced − frozen | mean `0.15978181829699184`; 95% CI `[0.11788071166290177, 0.20429248191903368]`; dz `1.3134377665377908` | **pass on fixed family** | | Worst reversal seed | relative loss `-0.0027290855917749624` | safety bound pass | | External-row causality | `17/17` integrated changed velocity on each of 30 seeds; `5/5` excluded were exact zero on each seed | **pass** | | Numerical fault run | 10,000/10,000 revisions; every signal finite and in `[0,1]` | **pass** | | Shared-field worst-process p95 | `0.10890000000000555 ms` | **pass** | | 50,000-entity process-median increment | `1.3857499999999163 ms` across three isolated, order-counterbalanced processes | **pass** | | 50,000-entity repeat stability | worst of 30 batch medians `2.8494000000000597 ms`; `30/30 < 3 ms` | **pass** | | Every-consumer counterfactual | targeted ten-file run: `101` tests, `0` fail, `377,438` assertions across all ten named classes | **pass** | ## Evidence inventory and complete named-consumer coverage The JSON field `consumerCoverage` binds the receipt to a targeted `bun test` run (`101` pass, `0` fail, `377,438` assertions) over these exact evidence files; the full repository gate reruns them: - `tests/operational-organism-intelligence.test.ts` - `tests/alien-flora.test.ts` - `tests/nhi.test.ts` - `tests/glyph-brain.test.ts` - `tests/wilderness.test.ts` - `tests/tsotchke-facade.test.ts` - `tests/shoggoths.test.ts` - `tests/puppet-masters.test.ts` - `tests/titans.test.ts` - `tests/leviathans.test.ts` Those tests cover ordinary entities, alien flora, NHIs, glyph beings, wilderness fauna, primordial digital biologics, shoggoths, puppeteers, titans, and leviathans. The four specialist seals hold seed/config/signal values fixed, toggle only `enabled`, assert deterministic replay, and require full-class action or trajectory divergence. Archon/apex and Mechalogodrom paths are pre-existing deep systems outside this named-consumer gate. ## Why earlier receipts are superseded V2 removed both the corpus signal and the explicit goal in its disabled arm, confounding goal-controller and corpus effects. The first V3 draft fixed that control but reused the V2 seeds and overstated a four-aggregate-channel rotation as a repository permutation and a final exploration replacement as entropy matched. The current V3 receipt replaces that draft before publication: it uses a disjoint seed family, honest control names, a uniform-random-action difference, isolated and order-counterbalanced three-process/30-batch performance stability, an executed ten-file consumer gate, and a SHA-256 of the exact benchmark source. V1/V2 remain tracked as historical evidence. ## External repository ledger The live registry contains exactly 22 external repositories: 15 under the `tsotchke` user and seven under `Tsotchke-Corporation`. The corrected classes are `8 deep`, `7 wired`, `2 harvest`, `4 fenced`, and `1 meta`. Thus `17/21 = 0.8095238095238095` non-meta repositories are integrated. The internal `classical-contrast` experiment remains operational but is not an external repository. | Class | Count | External repositories | | ------- | ----: | -------------------------------------------------------------------------------------------------------------------------------------- | | Deep | 8 | `eshkol`, `moonlab`, `tensorcore`, `libirrep`, `spin_based_neural_network`, `quantum_geometric_tensor`, `quantum_rng`, `classical_rng` | | Wired | 7 | `simple_mnist`, `asteroids`, `PINN`, `PIMC`, `ulg`, `logo-lab`, `quantum-quake` | | Harvest | 2 | `homebrew-eshkol`, `Quantum-RNG-API` | | Fenced | 4 | `gpt2-basic`, `llm-arbitrator`, `SolanaQuantumFlux`, `OBLITERATUS` | | Meta | 1 | `.github` | `OBLITERATUS` is an AGPL-3.0 refusal-removal LLM toolkit and is fenced by both license and the non-LLM simulation mandate. `gpt2-basic` and `llm-arbitrator` are mandate-fenced; `SolanaQuantumFlux` is proprietary. Fenced and metadata rows are exactly inert. ## Eshkol, QRNG, and receipt provenance The Eshkol reference is `v1.3.2-evolve` commit `8443ddaeecec579c60ac858348a23cf1912d7a78`. Cosmogonic implements an order-0-through-8 Float64 Taylor-jet analogue; it is not native Eshkol parity, exact-rational arithmetic, unlimited order, or full R7RS. The Quantum RNG reference is `v3.0.1` commit `a00ad483cbbef31ea7536f09ae99409d81c9a823`. Cosmogonic uses a seeded deterministic classical state-vector adaptation. It is not a CSPRNG, SP 800-90B certification, hardware entropy source, physical Bell experiment, or quantum-advantage result. A simulated CHSH value near `2√2` is classical model conformance. The upstream audit also observed a current ARM64 duplicate-output test failure, so no production/security-readiness claim is made. The exact measured runtime is commit `517c1a41c0d5fbcbdf791a318056c7b28be5013c`. The exact benchmark source has SHA-256 `542e44bd6a502788f511ecb9e2be46414d17cfefae13fb416faeacfb3360ea68`. ## Public 9-axis profile and 113-system comparison The current canonical code-grounded vector is: `[4.0, 2.2, 3.2, 3.8, 3.9, 4.5, 4.3, 3.5, 4.0]` The independently gated live soup-selection and ordinary-entity paths support the bounded ecology/cognition increases already represented here. V3 authorizes no further increase. - breadth `3.711`, rank `#1 / 113`; - population z-score `2.881`, peer-only z-score `3.008`; - Mahalanobis distance `10.245`; - on the nine-dimensional Pareto front, dominated by `0` and dominating `22` systems; - profile evenness `0.992`; peer maturity remains `1.5 / 5`. These are curated comparative scores, not independent scientific validation. The consciousness-theory axis measures instrumentation breadth, not consciousness. ## Claim gate Allowed now: - synthetic goal-controller improvement over exact legacy on the fixed V3 family; - corpus-conditioned improvement over a goal-preserved disabled control; - fixed-family reversal adaptation above the script-declared 5% threshold; - per-repository causal reach for all 17 integrated rows on all 30 evaluation seeds; - exact inertness for four fences plus metadata; - 10,000-step bounded numerical safety; - all 30 incremental-cost batch medians below `3 ms` across three fresh processes; - matched counterfactual coverage for all ten named living-system classes. Not allowed now: - independent/preregistered replication or generalization beyond the fixed seed family; - superiority to the uniform random-action baseline; - repository-identity, substrate-specific, or quantum-specific behavioral uplift; - additional numeric A-Life, capability, consciousness, or sentience uplift from V3; - phenomenal consciousness, sentience, general intelligence, physical quantumness, or security claims.