# Phase-B Mechanism Development V3 — Falsification Result

**Generated:** 2026-07-11 · **Status:** development-only · **Advancement:** withheld

This milestone adds executable temporal-memory and NHI closed-loop tasks, exact neural checkpoints,
structured production action outcomes, and byte-stable evidence artifacts. It does **not** establish
learning, adaptation, consciousness, sentience, general intelligence, or a successful neural-scaling
result. The temporal configuration failed every advancement criterion. The NHI task produced two
narrow positive lane diagnostics, but its paired surface-conflict response declined and no broad
four-action neural interpretation is allowed.

- [Canonical compact JSON](./assets/phase-b-mechanism-development-v3.json)
- [Long-form metric CSV](./assets/phase-b-mechanism-development-v3.csv)
- [Accessible result SVG](./assets/phase-b-mechanism-development-v3.svg)
- [Deterministic artifact generator](../../scripts/organism-intelligence-phase-b/mechanism-evidence-artifacts.ts)
- [Artifact integrity test](../../tests/phase-b-mechanism-evidence-artifacts.test.ts)

![Phase-B temporal and NHI mechanism result](./assets/phase-b-mechanism-development-v3.svg)

## Outcome

| Question                                                                                         | Executed result                                                             | Honest conclusion                                             |
| ------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------- | ------------------------------------------------------------- |
| Can the 101-input V3 temporal predictor recover a hidden delayed cue under terminal-input twins? | **No.** All eight advancement criteria failed.                              | Do not advance this configuration to confirmatory evaluation. |
| Does the NHI closed loop execute bounded, target-valid material actions before acknowledgement?  | **Yes**, for the synthetic development task and production action boundary. | Functional closed-loop evidence only.                         |
| Do semantic neural inputs help every declared NHI action?                                        | **No.** Only HUNT/resource and SPAWN/social are interpretation-eligible.    | Withhold a broad semantic-neural effect claim.                |
| Does performance improve after the surface mapping conflicts?                                    | **No.** Mean service declined by `0.013306`.                                | Paired conflict response, explicitly **not adaptation**.      |
| Is consciousness or sentience established?                                                       | **No.** Those claims are prohibited by both harnesses.                      | Proxy mechanisms remain mechanisms, not phenomenal evidence.  |

## Temporal-identifiability gate

The temporal task uses counterbalanced twins with byte-identical terminal inputs and labels determined
by a cue observed 2, 8, or 16 steps earlier. Present-only controls therefore cannot solve it. All 80
configured task seeds produce distinct seed-derived cue/query/neutral profiles, all 46,080 configured
rows are retained, and only 36,864 locked validation rows enter the gate.

The primary V3 H8 identity-history model measured mean SSE `0.168091`, mean cross-entropy
`0.710260`, mean twin margin `0.000118`, and ordering rate `0.5` across 3,072 locked
validation rows.

Raw model execution remains IEEE-754 binary64. The temporal schema-2 rows, derived statistics, gate
decisions, digests, and combined schema-4 artifact apply one fixed-decimal `1e-6` evidence law before
publication. The same law covers temporal and NHI numeric fields. Its quantum is 10,000 times finer
than the smallest non-zero temporal threshold (`0.01`). Every observed metric in this sealed gate is
also more than one quantum from its own comparison boundary, including the strict zero bootstrap
boundary, so canonicalization did not change this sealed verdict; a future near-threshold result must
re-establish that margin. SSE and cross-entropy are computed from each raw prediction before their
evidence leaves are canonicalized, so rounding cannot snap a near-endpoint forecast to `0` or `1`
before metric evaluation.
Windows x64 and Ubuntu/WSL x64 on Bun `1.3.14` reproduced the same temporal seals and the exact same
JSON, CSV, and SVG bytes; artifact `--check` passed on both platforms.

| Advancement criterion             |   Required | Observed worst/minimum | Result   |
| --------------------------------- | ---------: | ---------------------: | -------- |
| Mean SSE gain over every control  |  `>= 0.04` |            `-0.008091` | **FAIL** |
| Median model gain                 |  `>= 0.03` |            `-0.007411` | **FAIL** |
| Every-delay mean gain             |  `>= 0.02` |            `-0.014588` | **FAIL** |
| Mean twin margin                  |  `>= 0.20` |             `0.000118` | **FAIL** |
| Correct ordering rate             |  `>= 0.75` |                 `0.50` | **FAIL** |
| Maximum Holm-adjusted one-sided p |   `< 0.01` |                  `1.0` | **FAIL** |
| Minimum bootstrap 99% lower bound |      `> 0` |            `-0.010691` | **FAIL** |
| Worst model gain                  | `>= -0.01` |            `-0.017918` | **FAIL** |

This failure is configuration-specific. It falsifies advancement of the recorded V3 H8 training
configuration; it does not prove that recurrent memory cannot solve the task.

## NHI closed-loop development result

The NHI harness commits each intent before hidden opportunity resolution, then applies bounded HUNT,
SPAWN, MANIPULATE, or DOMINATE semantics and returns only narrower `factSupported` evidence. The mind
then decides whether that evidence causes an actual GOAP transition. Cue, hidden outcome, mind birth,
policy, surrogate, and fault streams are domain-separated. The pre/post
surface-conflict phases reuse paired exogenous trials and reset environment, mind, and policy state at
the boundary. A balanced semantic derangement has zero fixed points, and the action yoke has zero
invalid current targets.

All 41,472 configured rows were retained. On locked validation rows, the full arm measured mean service
`0.060771` and success rate `0.211589`. Its aligned pre-conflict mean was `0.067424`;
post-conflict mean was `0.054118`; the paired response was `-0.013306`. That decline is not
evidence of adaptation or reward learning.

| Validation diagnostic                          | Full-minus-control mean service | Interpretation                           |
| ---------------------------------------------- | ------------------------------: | ---------------------------------------- |
| Balanced fixed-point-free semantic derangement |                     `+0.020168` | Development diagnostic; not confirmatory |
| Target-rebound action yoke                     |                     `+0.037909` | Development diagnostic; not confirmatory |
| HUNT/resource neural-semantic ablation         |                     `+0.018743` | Narrow lane interpretation allowed       |
| SPAWN/social neural-semantic ablation          |                     `+0.017038` | Narrow lane interpretation allowed       |
| MANIPULATE/explore neural-semantic ablation    |                     `-0.001696` | Unsupported diagnostic only              |
| DOMINATE neural-semantic ablation              |                     `+0.007298` | Unsupported diagnostic only              |

## Production boundary hardened

- `NhiMind` and `NhiSystem` now have versioned, JSON-safe, fully validated exact state snapshots;
  malformed restores are atomic and the 32-mind runtime cap is enforced.
- NHI world actions return `effectApplied`, `factSupported`, affected-body count, and exact energy
  transfer separately. The orchestrator validates action-specific evidence; the mind then reports
  whether its internal GOAP facts actually transitioned, so repeated evidence is never mislabeled.
- One percept/think/apply/acknowledge exception cannot starve later minds; even a failing diagnostic
  callback is captured in the exceptional report while the population beat advances.
- NHI identity, policy, and direct action sampling use dedicated deterministic streams. Material
  organism creation intentionally still advances the ordinary EntityManager physics/genome streams;
  the code and launch receipts no longer claim otherwise.
- Failed launch attempts consume a monotonic attempt id and do not pretend RNG can rewind. Rollback
  attempts to remove mind, maps, target, wallet, body, and spawned entity without fabricating a death
  event or decrementing mutation accounting; its receipt now publishes cleanup failures, individual
  absence postconditions, and `logicalRollbackComplete` instead of assuming success. Scene/GPU cleanup
  is explicitly `resourceCleanupStatus: best-effort-unverified`, never folded into the logical claim.
- Swarm births isolate each child attempt: an internally rolled-back failed spawn cannot erase earlier
  material children or suppress their aggregate acknowledgement. Titan procreation likewise reports
  attempted, succeeded, and failed births rather than equating attempts with births.
- Organic mitosis, sparse recovery, behavior-driven births, and NHI swarms now share the ultra-tier
  `512`-organism frame budget. Null/throwing spawns refund their reservation; at tiers capped at 5,000
  or below, the wrapper preserves the direct spawn trace and next RNG draw exactly.
- The exact-current spatial index performs zero work at `M=0` live NHIs and one `O(N)` rebuild for any
  `M>0`, independent of whether `M=1`, `8`, or `32`. The operation-count test pins 1k/10k/50k rows;
  a non-gating Windows-local probe measured 50k medians `0.507–0.514 ms` (maximum p95 `0.683 ms`).
- Alien visual bodies construct transactionally, delete logical ownership before no-throw disposal,
  and never free shared geometry during attempt rollback. DOMINATE and RETREAT sanitize hostile numeric
  velocities before bounded impulse arithmetic and distinguish repair from causal action evidence.
- Optional audio, HUD, audit, and same-frame grid indexing are guarded so their failure cannot erase
  an already-material action. Social-audio cooldown is committed before its external sink runs, and
  ordinary spawn audio cannot turn a completed entity insertion into a half-transaction.
- NHI bodies are excluded from ordinary SuperHunt and DomeFeeding predation paths; focused tests seal
  both immunities.

## Integrity receipt

### Temporal task

- Seed-family SHA-256: `a14179926f8fd43041b773790dc3bd62dcc5dc5bafc41db4a86fe4ea03794688`
- Configuration SHA-256: `e9d41c9ed838375f848867fd05e71ee895868feb8b5d64044a9015bf6ec73479`
- Row-matrix SHA-256: `241f61fee25f4d48462135083cecedd55cd60f36178dacd56a843de6121226c4`
- Task-profile-family SHA-256: `06ebbf60d95c3d09277f350b3ea768fdd0dfb97b86fea02a66efe7946fc3e6a2`

### NHI task

- Selected-family SHA-256: `ab86f694abcf079ed0ab44aaf97c5cf3fc4ae709d1d8b0580ed4f87dadfc3408`
- Global 22-family firewall SHA-256: `469f79e59c29639034afb4aea2bf6b0e3a82f2a3f3303b7fb3c9efa7ec443b8a`
- Configuration-protocol SHA-256: `bfff6581b8e3e032c596a114ef3bfc86c7d5067537efe24043d5d731a09e0f0c`
- Protocol-hash scope: configuration and declared laws, **not** source-blob closure; the Git tree is
  the source provenance, while schedule and row hashes bind the executed outputs.
- Schedule SHA-256: `8acb60a51bbef06bbf238f8b365f8b9f2d61314850909ff109cbbf70514f617f`
- Row-matrix SHA-256: `43f26e8b224449db588dd56ba5dd16c7f37c579e396a5689421e5545ee35db06`
- Fault-probe SHA-256: `2ba8bb51fa2c5d212cc51afd58883b4a8ecc0a62c0ad1f784ebd85b83c61670c`

### Compact evidence artifacts

- JSON SHA-256: `a509e241d9393bc807312fe994815f1fe988590112fc31736854992ede93ec83`
- CSV SHA-256: `eafef6d180c350b103f7a4fd33f57325a4acdb650ccd7aaea6490f33dd7f476c`
- SVG SHA-256: `bf7bb362d472aee5e848c2ef9d13b44c56a44c6af86e27e1568943a07f84af47`

For each displayed NHI validation comparison and action contrast, schema 4 preserves the source-study
receipt under an explicit `source*Sha256` field and adds a separately recomputable artifact-local hash
over the rounded displayed material. The artifact-local law is
`phase-b-rounded-material-canonical-json-sha256-v1`; both receipt fields are excluded from its scope.

## Reproduce

```sh
bun scripts/organism-intelligence-phase-b/mechanism-evidence-artifacts.ts --check
bun test tests/organism-intelligence-phase-b-predictor-temporal-development.test.ts
bun test tests/organism-intelligence-phase-b-nhi-closed-loop-development.test.ts
bun test tests/nhi-system.test.ts tests/nhi-target-effects.test.ts tests/entities-death.test.ts
bun test tests/nhi-world-production-boundary.test.ts tests/spawn-budget.test.ts
bun run bench:nhi-grid # wall-clock reference only; operation counts are the gate
bun run typecheck
```

## Claim boundary

This milestone authorizes only: exact checkpoint state and continuation when resumed with the same
caller-owned decision-RNG state (the RNG is not serialized); bounded production NHI action effects and
lifecycle isolation; a failed temporal advancement gate; and narrow synthetic-task
HUNT/resource plus SPAWN/social development diagnostics. It authorizes no adaptation, reward-learning,
cross-family neural scaling, consciousness, sentience, NHSI, AGI, physical-quantum advantage,
production-readiness, or independent-replication claim.