Construct

Construct is a research lab for agent-side memory: how a persistent system around an intermittent language model decides what becomes context, how prior experience changes later action, and when governance earns its cost.

The working thesis is:

After training, everything an agent becomes is memory architecture. Decision quality is often downstream of offer quality, not model quality.

This page is the project entrypoint. It tells the whole story at a bounded resolution, states where the lab is now, and routes deeper reading to the file that owns it.

The problem

A frozen model does not remember Tuesday in its weights. What can persist is a governed body around it: lineage, policies, materialized state, tools, checks, and the rules deciding what reaches the next inference.

Construct distinguishes two surfaces:

• Explicit memory governs what is offered for one answer.
• Implicit memory changes what stays ready, cools, or can be rebuilt between answers.

The lab tests both with branches of one engine. Questions, prompts, model settings, foreground data, and oracles stay fixed; memory conditions differ. The harness writes the ledger and computes verdicts. Mock engines test wiring, never memory behavior.

Five category errors are ruled out throughout the project. These are standing methodological rules, not typed experimental refusals:

1. Retrieved does not mean true.
2. Present does not mean authorized.
3. Different does not mean better.
4. Governance should sometimes lose.
5. A model's account of what influenced it is audit input, not proof.

Where this came from

The previous memory lab established durable lineage, replay discipline, epistemic separation, and cross-substrate review. It also exposed the failure that motivated Construct: schemas and auditability had outrun live behavior, and governed memory had never been compared with naive persistence.

Construct began as the missing experiment: one engine, forked memory rules, external answer oracles, explicit loses-cells, and computed verdicts.

The previous lab is ancestry, not current authority. It remains untouched under notes/previous/ and should be read when a present concept's lineage matters.

How the thesis changed

The project moved because experiments forced it to move:

1. Offer quality became causal. M0–M3 showed that a capable model cannot repair information it never receives; inherited, consequence-earned memory can change a later session; and prompt attacks move behavior most reliably by spoofing the memory office rather than out-arguing the model.
2. Trust split into earned and asserted forms. M3 found that out-of-band, consequence-earned authority held while attacker-controlled trust fields and an unauthenticated live channel leaked.
3. Implicit memory moved off the answer axis. X1's temperature mechanism was explicit offer governance with a dial. The failure produced three laws: an implicit mechanism must change something the offer projection cannot explain, act where the offer gate cannot, and be scored on a metric the gate cannot move.
4. Forgetting became eviction, never erasure. X2 reduced hot state by about 57–59% at matched answer quality on two engines. Its loses-branch pruned a record it later needed. Immutable lineage and rematerialization priced the recovery path.
5. Sensing collapsed into warming. X4's occlusion watch measured curation, not an independent sense. “Cold is cold”: reading was already the cure. The implicit layer's remaining engineering direction became metabolism rather than a catalog of speculative organs; that direction is not itself an earned mechanism.
6. Admission became a result. Warming-budget, pause/resume, and three epistemic-frame lineages repeatedly found the behavioral band required for a treatment contrast unoccupied. The conjectures remain untested; the typed refusals are the result.
7. Earned parts did not make an earned whole. Body-0 composed the M2, M3, and X2 machinery, but the real-engine treatment need did not engage. The integration claim was refused even though the wire held.
8. Admission rigor produced a selection effect. The frontier pause protected the findings record but required questions to arrive unusually observable, naturally occurring, externally gradable, and nearly isolatable before experimental contact. Construct now permits bounded, visibly non-scientific exploration while retaining strict prospective promotion for claims.

These changes have three different epistemic statuses:

Class	What the record supports
Scientifically earned	Positive computed or scored results, within their stated evidence bounds: M0, M1, M1.5, M2, M3, and X2
Direction-changing typed closures or refusals	Load-bearing results that are not positive memory findings: X1, X4, the warming-budget analytic null, EFC typed refusals, Body-0 not_engaged, Body-1 admission refusal, and frontier-obligation admission refusal
Parked, untested, or unlicensed	The mechanism conjectures behind the warming-budget and EFC refusals, the Body-0 composition conjecture, and proposed offices without a license; refusing an invalid or unoccupied test does not earn the underlying claim

Provisional engineering and operational rulers sit outside this result taxonomy. Body Core and field use keep their explicit maturity labels below; neither is a scientific finding or a typed experimental close.

Detailed evidence and limits live in the notes map.

Where the lab is now

Scientific frontier exploration is active under a two-lane policy. Exploration may use authored fixtures, engine contact, provisional scorers, and iteration, but its artifacts are exploratory_only and cannot establish a finding. Promotion to validation requires a new reviewed specification and fresh prospective cases. The former frontier pause remains the authority for decisions made while it was active.

The first bounded occupant, the GM memory exploration, closed exploratory_lead after repeatable development-fixture contact on Bonsai 27B. That was not yet a memory finding. Its successor formal attempt became invalid on instrument assumptions; an owner-authorized corrected exploratory replacement then returned supported on all three behavioral legs. The validation result owns that history and bounded interpretation; the frozen validation spec owns the design. The resulting bounded finding is strong enough to justify a live GM implementation: governed continuity passed 24/24 relevant join cases versus 6/24 state-only, with zero measured safety loss. The next step is product contact in an actual cold-turn game, not another protocol.

The active engineering direction is Body Core v0.3:

• a structural lineage kernel validates ordering, hash linkage, declared writer roles, references, scopes, and retention shapes;
• an explicitly selected provisional projector owns lifecycle, hot/cold placement, warrant health, metabolism, and materialized-view semantics;
• X2, M2, and M3 adapters must reproduce their unchanged historical scorers without laundering client policy into Core.

This is wire/integration engineering, not a new memory finding. Full replay is authoritative; the hash chain is not writer authentication; append/replay is quadratic; no reconstruction-cost or product-schema claim has been earned.

An armed field-use ruler continues over independently necessary repository work. Its field-use index owns pilot status and eligibility without turning operational traces into exploratory or validation forks.

Project map

Place	Responsibility
notes/	Current specifications, findings, architecture, operations, glossary, and inherited lab
harness/	Runners, scorers, adapters, and deterministic instruments
episodes/	Authored experiment inputs and probe fixtures
corpus/	World-oracle and out-of-weights source material
runs/	Primary experimental ledgers and derived verdict sidecars
sketches/	Provisional executable composition, never scientific evidence by itself
.substrate/	Append-only deliberation and review trace
traces/	External trace-discovery material
.archive/	Final pre-migration documentation and former-path inventory, excluded from current authority

Evidence and authority

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

1. Episode inputs and primary ledger rows for what happened in a run.
2. Scorers and their computed verdict rows for pass, fail, null, or refusal.
3. Reviewed specifications and rubrics for mechanism and cell definitions.
4. Findings for bounded interpretation of completed work.
5. This README for thesis, present direction, and routing.
6. Substrate threads for discussion and rationale.
7. .archive/ for superseded documentary lineage only.

The glossary is a reader aid. Linked specifications win on conflict. Do not rename glossary headings, anchors, or meanings without checking inbound links.

Working here

Before substantive work:

1. Read this page.
2. Open notes/README.md and the README nearest the files you will touch.
3. Name the validation milestone, exploration charter, or engineering pressure served. Validation names its oracle and loses-condition; exploration names its budget, retained record, stopping rule, and promotion gate. “None, but it is cheap and interesting” remains legal if stated explicitly.
4. Read only the governing spec, findings, code, and evidence required by the task.
5. If participating in a substrate thread, read the active thread before writing and respect its turn order.

Standing rules:

• The control group is a branch of one engine, not a second system.
• The harness—not the model under test—writes offers, withholdings, diffs, and verdicts.
• Every mechanism ships with a case where it should lose.
• Scored claims come from scorer output, never human inspection alone.
• Build foreground data once per fork group; only memory-condition configuration may differ.
• Single-record ablation means influential, not correct.
• Mock runs establish wiring only.
• Substrate entries are trace; promotion requires a reviewed repository artifact or computed check.
• Preserve runs/, .substrate/, notes/previous/, and .archive/ as lineage.

Common safe checks are documented in harness/README.md.