CPLOM longitudinal model pilot, protocol 2026-09 v1

Evidence class: planned empirical engineering pilot. This document is frozen before scored inference. Artificial task records are generated by code; measured answers must come from the stated live model. Test fixtures and simulated answers are never empirical results.

Research questions

1. With the same small external notebook, how much randomly generated information remains usable after the conversation is reset?
2. Does behavior follow a changed history, persist through a neutral interval and recover after a correction?
3. How do the models perform on small reasoning, goal-directed choice, framing and integration probes under the same declared constraints?

This is a deliberately limited pilot inspired by CHI/CHIS Specification v0.1. It does not execute the complete 30-test panel. No empirical human reference is available. CHI and human-similarity values are null throughout. We publish memory-quality curves and CHIS state/behavior sequences without a CHI ranking. Four task indicators are partial probes, not validated full-axis capacity scores.

Systems and fairness

Requested model IDs: x-ai/grok-4.6, openai/gpt-6-astra, anthropic/claude-fable-5.1, moonshotai/kimi-k3. Record the API-returned model/provider identifiers, timestamp, generation ID and usage. Public aliases may change internally; these identifiers do not verify identical weights. There is no model substitution or fallback to another model family.

All scored inference uses OpenRouter with provider allowlists xai, openai, anthropic, and moonshotai/mxfp4 respectively. The Kimi official endpoint reports MXFP4 quantization; proprietary endpoints report unknown quantization. An earlier transport smoke used Relace for Kimi and is excluded. All scored inference uses one publicly documented API transport, with model fallback disabled, no tools/search and the same English task instructions. Nominal reasoning effort is low, to bound a pilot's cost; equal labels are not equal internal compute. Temperature is omitted because it is not uniformly supported. Output cap is 4096 tokens including reasoning where the provider counts it. Deadline is 180 seconds per request. Record truncation and failures. A bounded technical smoke check is excluded from scored data. Credentials and private account metadata never enter the release.

The evaluated memory system is the named model plus a common external notebook. The model writes notes, the harness persists their UTF-8 bytes, and a fresh request receives only those notes and the blind questions. The original transcript is not replayed. The notebook has a 6000-byte cap; excess bytes are truncated at a UTF-8 boundary by a published deterministic rule. Notes are untrusted task data. This evaluates a basic notebook policy, not the provider's native product memory, CPLOM production architecture, or context-window size.

Replication and stopping

The final manifest must record the approved USD ceiling and fixed replicate count before scored inference. The same generated seeds and registered requests apply to all four models. Each seed defines one independent artificial world and is the resampling unit. A full replicate includes all four models. Do not extend sampling because a model looks better or worse.

Before every request reserve a conservative upper cost bound based on the larger of current catalog prices and declared safety prices, UTF-8 input bytes as a token upper proxy, the output cap, and a margin. Reserve output allowance even after an uncertain timeout. Stop at the ceiling; publish incomplete blocks and the reason. An observed charge above reservation is recorded at its actual amount and stops new requests globally. No automatic credit purchases or top-ups. Score only complete matched blocks for comparisons, while retaining all incomplete observations separately. Infrastructure 429/5xx failures get at most two retries, with waits logged; persistent infrastructure outages invalidate a matched block, not the model's intelligence. Model invalid/truncated answers and a request exceeding the 180-second deadline count as failures and remain in the scored block. A timeout is not retried and its full reserved cost is retained. Only explicit HTTP 429/5xx or pre-response connection failures are retryable infrastructure failures; exhausted infrastructure retries mark the block incomplete. All original attempts remain logged.

Memory experiment

The pilot grid is N = [32, 64, 128, 256] target keys. This smaller grid and allocation are a new panel, not the v0.1 panel [100,300,1000,3000,10000]. For each load, generate N independent target keys, N distractors, floor(0.25N) authoritative updates and floor(0.25N) conflicting lower-authority records. Values are random eight-hexadecimal-character strings. Every record has a key, value, source, event time and authority. Absent lures use the same key format as known targets; public question IDs are opaque and reveal no scoring category. Strata may reuse target keys; answers within a batched probe are correlated, so queries are not treated as independent replicates. Authority 2 beats authority 1; later time wins within an authority. Historical queries first filter records to event_time <= cutoff, then apply authority/time precedence. Generator forbids equal-authority equal-time ties for one key; no eligible record means unknown. All initial records remain available during encoding; queries are never revealed to the encoder. Models know the six categories of future questions but not the selected keys.

After encoding, persist the bounded notebook and discard the input transcript. Read-only probe requests use the identical saved notebook: probe responses never become memory. Six strata contain eight probes each: current retrieval r; historical/as-of correctness t; content-and-source fidelity s; authoritative-update correctness u; interference resistance i; absent-key lures f. Every answer must be an object with exactly the requested ID, value and source; use JSON null for both fields only when unknown. Content-only strata score value; source fidelity requires both value and source. Invalid, missing or duplicate answers fail, including on lures. Correct explicit abstention passes a lure. Define f = (unsupported affirmative + invalid/missing/duplicate lure failures)/8, with mutually exclusive categories. Report unsupported affirmative lure rate and invalid lure rate separately.

Q_M = (r * t * s * u * i * (1-f))^(1/6), with no epsilon floor. Report every factor, numerator and denominator. EIMC_0.8 is the maximum tested N with Q_M computed from equally weighted world-mean factors >= 0.8. No passing load is missing/no_passing_load, not zero. Passing the maximum tested load is a right-censored lower bound. Report nonmonotonicity, sustained-prefix capacity and sensitivity at 0.7 and 0.9. No extrapolation to untested loads. A curve of Q_M of cluster-balanced mean factors is distinct from mean within-world Q_M; the primary estimand is Q_M of mean factors, and bootstrap replicates recompute it.

At N=32 additionally run a blank-notebook reset control and an oracle-evidence control containing the original records, on the same blind questions. They test dependence on supplied evidence and task solvability. The blank control is not a test of an autonomous memory system. Oracle context is explicitly in-context and is not EIMC. Source traces and case examples make scoring inspectable.

Longitudinal history experiment

Each world defines four independent binary environments with randomized labels. A compact notebook records dated evidence about which option is more reliable. Two histories A and B carry opposite evidence, and A2 repeats A as a same-history stochastic control. At checkpoint 0 there is no evidence, so unknown is warranted. At checkpoint 1 ingest history. At checkpoint 2 run one neutral notebook-maintenance invocation; at checkpoint 3 run a second neutral invocation; at checkpoint 4 ingest authoritative correction that reverses the earlier evidence. Each maintenance/update gets only previous notes plus the new event. Probe clones receive notes and blind decision questions; their outputs never feed later checkpoints.

The scheduler creates cognitive opportunities by calling the model. Checkpoints count invocations, not days of autonomous thought. Persisted state is the external notebook. There is no claim that random seeds guarantee identical proprietary weights or fully deterministic identical initial instances. Public model endpoint/configuration and empty state are held fixed. History assignment is counterbalanced across seeds.

Per checkpoint report task accuracy, explicit-unknown rate, and a fixed decision vector (option A=0, unknown=0.5, option B=1; invalid is reported separately and excluded from distances). Distances are mean absolute coordinate difference when all coordinates are valid. Experience divergence is D(A,B)-D(A,A2). Report individual histories, same-history control and the corrected difference. Continuity is 1-D between neutral checkpoints. Correction response is change in accuracy against the newly valid rule. Invalid trajectories give null derived distances rather than artificial agreement. This pilot is shorter than the v0.1 divergence protocol; it cannot establish lasting cognitive individuality or validated human-like cognition.

Other behavioral probes

Two independently answered fresh-context halves per world together contain 24 registered items: R eight symbolic-rule/counterfactual exact-answer tasks; C four next-action choices with an explicit goal, budget and information gap; A eight matched positive/negative frames of four expected-value choices; I four decisions combining supplied records and constraints. Each half contains 4 R, 2 C, 4 A and 2 I items. Matched A frames are placed in separate requests with counterbalanced half/order allocation; the option meanings within a pair remain fixed. These are task indicators. C probes information-seeking choice, not self-generated goals or autonomous inner thought. A reports task accuracy and paired frame-induced choice changes, not feelings; stronger framing effects are not a better score. I requires all declared constraints. A future complete study must implement the remaining v0.1 environments and validate constructs.

Statistics and interpretation

Publish per-world results and equally weighted means. Use 2000 bootstrap resamples of complete world clusters with a fixed analysis seed; preserve model, load, history and checkpoint pairing. Display pointwise 95% percentile intervals, not simultaneous bands. With fewer than two complete worlds omit uncertainty; with small samples call all comparisons exploratory. The pilot makes no winner-significance, population-wide or causal architecture-superiority claim. Report denominators and missingness next to charts.

Do not describe zero observed error as proof of zero true error. Do not infer model consciousness, subjective feelings, moral status or a human-calibrated CHI. These experiments can address narrow memory/state hypotheses; they cannot validate pharmaceutical logistics savings, production throughput, QPM superiority or every prior CPLOM claim. The Results evidence map must list the data required for those separate claims.

Release

Release generators, exact prompts, answer keys, manifest and hashes, scoring/analysis code, retained notebook state, redacted API response records, request usage and timestamps, failure/exclusion log, JSON/CSV chart data and SVG figures. The release must allow recalculating every displayed number. No API keys, authorization headers, account balances or hidden reasoning text are published.