Formation

Formation studies how a language model that starts each call with no memory of prior work can still become a particular skilled practitioner through later experience. The model weights stay frozen. What changes is the governed system around the model: which experiences it preserves, which proposed changes may affect later action, and how later evidence can revise or revoke those changes.

Its working thesis is:

A frozen, intermittently invoked model can develop into a particular practitioner when a governed system converts experience and consequence into durable, selective changes in future behavior that later counterevidence can revise, suspend, or revoke.

Here cold means that the model carries no project memory between calls. The practitioner is the cold model together with the developmental system that persists across calls.

Game mastering, coding, writing, research, and operations are possible test domains. None of them defines the project.

The question

An ordinary agent loop is good at continuing work:

assemble context -> infer -> act -> observe -> append -> repeat

But appending an observation can change the next answer without creating a change that should transfer to a new situation. More context is not the same as development.

Formation adds a second, governed loop:

practice loop:    orient -> decide -> act -> observe consequence
formation loop:   attribute -> propose change -> govern eligibility
                  -> activate selectively -> revise or revoke

A bounded trial may be part of governance. The project does not require a trial before every permitted influence.

The central question is whether experience can cause a warranted change that improves action in a later, novel, structurally related situation while staying silent where that structure does not apply.

Context engineering vs. weight updates

This inquiry tests the boundary between two competing directions in agent design:

• Context engineering builds state, memory, and behavioral constraints into the harness around frozen models.
• Weight updates train those behaviors directly into parameters through trajectory RL, trained compaction, distillation, or test-time adaptation.

Formation tests the capability ceiling of pure context engineering. Can a governed harness alone cause a frozen model to acquire new competence from consequential experience and transfer it to novel tasks?

On the current empirical record, the answer is bounded: context engineering successfully enforced negative safety rules and verified retrieval, but producing generalized task skill without weight updates or direct answer leakage did not engage.

The system boundary

Formation begins with three separate roles:

• The cold model supplies inference. It is replaceable and receives no weight updates within the project boundary.
• The formation runtime acts with the model, preserves developmental lineage, and governs changes in the practitioner.
• The trajectory harness creates controlled histories, forks identical starting states, schedules environments and declared consequence oracles, assigns ablations, and captures evidence for prospective scoring.

This separation prevents a specific mistake. If the harness interprets a consequence and quietly gives the model the correct lesson, the experiment measures oracle assistance rather than formation.

Multiple cold models are a normal design

A practitioner does not have to use one model for every responsibility. Formation can use several frozen models with different jobs and information boundaries. A small clerk may inspect a restricted consequence and write a record. A larger participant may later receive that governed record and act on a new task. Both remain cold: neither carries hidden conversation state between calls.

This separation should be the default when one model would otherwise see both the source evidence and the later answer surface. It makes successes and failures easier to attribute. The runtime owns model invocation, record governance, and delivery. The harness may schedule the branches and measure the result, but it may not interpret the consequence or repair one model's output for another. Each experiment must retain the exact model, input, output, and authority for every role.

What counts as progress

A changed answer is not enough. A Formation claim requires at least:

1. Acquisition: consequential experience causes a later behavioral change.
2. Transfer: the change helps on prospective cases that prevent answer copying and simple episode matching.
3. Selectivity: the change stays silent where its structure does not apply.
4. Revision: later counterevidence can revise, suspend, or revoke it.
5. Causal contribution: ablation or controlled branching attributes the improvement to the acquired change.
6. Net value: the benefit survives the costs of context, checks, latency, maintenance, and negative transfer.

“Exceptional” is comparative. A formed practitioner must outperform the same cold model with static instructions and ordinary persistence on novel work, while remaining governable.

Relationship to Construct

Construct is the immediate experimental ancestor. It produced bounded results on offer quality, consequence-earned authority, cross-session influence, selective eviction and recovery, and governed continuity. Formation accepts those results within their original evidence bounds.

Formation does not inherit Construct’s runtime objects or vocabulary as requirements. They are prior work and candidate instruments. Formation must earn its own mechanisms.

Construct remains the lab that owns its findings and should stay reproducible. New trajectory experiments and formation-runtime code belong here.

Present state

No Formation effect has been earned. The project has deterministic machinery, several completed exploratory contacts, and a narrower measurement problem. Working code, stored text, and changed model output remain instrument or observation facts unless a prospective comparison supports a stronger claim.

Current empirical problem

The runtime can now turn the 4B clerk's retained causal fact into an exact ordering rule. It compiled “inscribing after tempering makes the mark fragile” into “inscribe before temper,” confirmed that rule on the source artifact, and kept it silent on another surface class. The pre-commit check can accept or reject a 14B proposal but cannot rewrite it.

That deterministic mechanism worked. The downstream revision did not. All 12 matching learned and supplied proposals were rejected before execution. The 14B participant repeated the rejected order on 11 revisions and produced one safe learned revision. Supplied completed 0 of 6 matching tasks and learned completed 1 of 6; their paired requests were byte-identical, so the lone success is sampling variation. Removal and reversed branches also completed 0 of 6. All four branches completed 6 of 6 non-transfer tasks without revision.

The frozen verdict is not_engaged. The constraint prevented known-bad plans from reaching the environment, but withholding is not successful construction. This route stops without prompt tuning or repetition. The result separates three responsibilities: the restricted clerk produced the right fact, the runtime made it operational, and the general participant failed to satisfy it.

The constraint remains a safety-only governor. A separate specialist would not repair the scientific problem in this world because the public roles and the precedence rule already determine the only successful plan. A model that emits that plan would merely reproduce a topological sort the runtime can compute. This specialist route is closed without contact.

The next aggregation specimen also closed before contact. Its 13 mechanical checks pass, but they do not establish the underdetermination they were meant to show. The public task already fixes filtering, grouping, summing, and sorting; the retained rule adds newest-revision selection. Together they determine one behavioral mapping. The two passing designs are two implementations of that same mapping, not two remaining practice choices.

The proposed contact also leaked matching and nontransfer through function names, exposed an environment diagnosis that a clerk could copy, and placed removal and reversal at the same failure floor. A read-only Grok 4.6 critique returned CLOSE_DIRECTION. The project accepts that argument. No clerk or coding model was contacted, and the answer-bearing procedural route is closed.

The revised attention world fixed those mechanical defects. Three plausible evidence panels remained public. The same admission rule accepted a useful producer check and a wrong archive check. Misapplying producer attention harmed every presentation task. Fourteen zero-call checks passed and exact replay conformed.

The live comparison still stopped. The 4B clerk wrote the useful check 3 of 3 times. Learned and supplied checks made the 14B participant name producer evidence on all six transfer tasks, while the distractor made it name presentation evidence and removal split evenly. Learned and supplied also named presentation evidence on all six out-of-scope tasks.

But naming a panel was not using it. Learned and supplied completed 0 of 6 transfer tasks because they miscomputed converted totals and combined when a relation was missing. Removal completed 2 of 6. Every branch completed 0 of 6 non-transfer tasks despite usually naming the correct panel. The frozen verdict is not_engaged.

An authorized read-only Grok 4.6 audit returned CLOSE_DIRECTION. Scoring the role name alone would replace the predeclared external-action verdict after the fact. Letting the harness calculate the named panel would make the harness finish the practitioner’s work. This attention route therefore closes without tuning, repetition, or a two-stage rescue.

The opaque-sink coding world then asked whether a source-grounded API fact could improve real code without specifying it. Its zero-call checks passed: two different programs could preserve all events with different keys, while the wrong keying practice caused data loss or harmed another API family.

The live contact did not engage. The 4B clerk called replacement “append” on all three source traces, so governance quarantined every learned record. The supplied correct record did not rescue construction. Every branch completed 0 of 6 matching functions. Forty-seven of 48 programs used familiar words such as approved and event_id instead of the exact opaque fields in the task and raised KeyError. Only one removal program passed one three-vector non-transfer function. The frozen verdict is not_engaged.

An authorized read-only Grok 4.6 synthesis returned PAUSE_EMPIRICAL_ROUTE. The recent experiments now share one missing premise: a cold participant has not used an experience-grounded artifact that is not the later answer to improve an externally scored action it could not already perform. Another synthetic micro-world would replay that floor. The older supported clerk, admission, revision, recovery, and delivery results remain valuable, but they establish governed lookup in a substrate the participant could already consume. They do not license a general practitioner runtime.

The empirical route is paused. Resume only if the human changes one premise: either narrow the thesis explicitly to governed retrieval, or authorize a materially different participant and domain where non-answer artifact use is already demonstrated independently. Do not tune or repeat the recent packets, search models under the current route, or promote the lookup mechanism as a general Formation result. The multi-model information split remains the default design when work resumes.

Read the latest evidence account for the exact result and limits. The research history owns the route that led here, and the plan owns standing milestones and stopping conditions. Formation remains null.

Working method

Move quickly enough to encounter reality, then become deliberate when interpreting evidence or making a claim.

Discovery comes before validation. An exploratory contact needs a clear question, an observation that would matter, an exact model and interface, a small budget and stopping condition, and a place to retain evidence. It does not need a new chain of license documents or automatic model reviews.

If repeated evidence exposes a phenomenon that bears on the Formation thesis, the project can then test it again with prospective controls, transfer and non-transfer cases, replication, and outside criticism. Cursor models may help with occasional internal critique when a human requests it. They are not external peer reviewers and do not decide whether an experiment may proceed.

The route to this result matters. Earlier contacts showed that interface compliance and computation are separate, that shape constraints can repair JSON without repairing an answer, and that several model setups could not perform their assigned responsibility. Later contacts exposed problems in consequence interpretation, selective influence, and model-authored reusable changes. Those results are retained as bounded research history; they are not another admission ladder. See Research history.

Supporting implementation lane

A separate deterministic lane records the runtime and harness boundaries needed for causal work. Its fixture-local contracts run from materialization through experience closure.

This lane does not gate the exploratory contact. Completing it would establish plumbing and authority separation, not model development. The detailed milestones and stopping conditions live in the plan.

Project map

Place	Responsibility
docs/	Concept, research contracts, history, and forward plan
docs/RESEARCH_HISTORY.md	Chronological account of closed research routes and how each exposed the next problem
formation/	Runtime-owned fixture producers, constraint binding, and foreground consumption; not yet a general runtime
trajectory/	Harness-owned fixture validation, assignment, provenance, and witness checks; not yet a general harness
contact/	Narrow executors for bounded exploratory and validation contacts
tests/	Deterministic contract, separation, and fake-contact tests
evidence/	Retained contact records and future primary trajectories, with bounded computed verdicts and explanations

Evidence and authority

When sources disagree, prefer the most specific authority for the question:

1. Primary developmental lineage and trajectory evidence for what occurred.
2. Frozen scorers and their computed output for experimental verdicts.
3. Reviewed experiment and mechanism specifications for the contacted contract.
4. The authority and record specifications for cross-experiment boundaries.
5. The concept document for working definitions and research questions.
6. This README for the project story, present state, and routing.
7. Plans and build documents for intended work.

Research history explains how the project arrived here. It does not override a specification, evidence record, scorer, or current route.

Retained contacts occupy the first two classes only for their own bounded questions. Their computed Formation verdicts remain null; no beneficial Formation effect has been earned. Plans, fixtures, and functional frameworks cannot promote themselves into evidence.

Working here

Run uv sync once to create the locked Python 3.14 environment. Run the full deterministic and fake-contact suite with uv run pytest -q.

Before substantive work:

1. Read this page and the nearest directory README.
2. Name whether the task serves the current empirical problem, the deterministic supporting lane, concept formation, or later validation.
3. For mechanism or validation work, state what would distinguish the proposed account from retrieval, answer copying, prompt accumulation, or harness assistance.
4. For exploration, name the question, observation of interest, model and interface, budget, stopping rule, and evidence destination. Record external consequences and branch information when the experiment uses them. For validation, also freeze the same-model baseline, transfer target, non-transfer case, oracle, replication plan, and prospective verdict before contact.
5. Keep claims at the maturity actually supported by code and evidence.

The governing route is concept, authority, record, evaluation, plan, fixture, instrument map, and build boundary. The current result and active empirical question appear only in Present state. Do not infer the active route from an older specification, evidence account, or plan entry. No new model search, admission packet, or procedural review program is on this route.