# What changes when context emerges from records?

2026-10-23 · en

RGM / Relational Space investigates deriving contexts from record memberships rather than assigning classes in advance. H1a is the narrow experiment line testing their value for evidence retrieval. Design, calibration records and open gates are collected here.

This guide is an editorial summary, not the original package, an independent reproduction or a final H1a outcome.

Consider two systems with the same records, encoder and retrieval budget. One uses fixed contexts; the other derives them from record distribution. Does the latter find evidence for a query more effectively? That is H1a’s primary question, not a test of every claim in the broader architecture.

## Separate sources from organization.

Persistent epistemic content stays in records. Contexts, topology and representations can be rebuilt; they do not become new sources of knowledge.

## Use is not evidence.

Similarity, centrality, repetition and usage are distinct quantities. Query history must not change canonical knowledge or topology.

## A minimal test, a bounded result.

H1a tests anonymous membership and derived context for retrieval. Role ports, paths, facets, online organization and conceptuality need separate hypotheses.

## From source records to evidence retrieval

1. **Preserve records and provenance** — One message is the record unit; session, order and source provenance remain distinct. Evaluation questions are not inserted into record content.
2. **Construct context memberships** — Fixed schemas or offline-derived structures assign records to contexts. Hidden class information derived from queries is not used.
3. **Retrieve under the same budget** — Use the shared first stage and bounded expansion. A record does not gain extra votes merely by appearing in more active contexts.
4. **Compare arms and controls** — The primary contrast is E_off − F. Null models, neighbourhood, stability and interpretability test different explanations; one score is insufficient.

## Seven comparison arms

| Arm | Role | Distinction |
| --- | --- | --- |
| `B0` | Shared first stage | Baseline without expansion. |
| `N` | Neighbour expansion | Does any gain come merely from adding nearest neighbours? |
| `F` | Fixed-schema context | The fixed-structure side of the primary contrast. |
| `R_F` | Null model of F | Degree-preserving rewiring of F memberships. |
| `E_off` | Offline-derived context | The derived-structure side of the primary hypothesis. |
| `R_E` | Null model of E_off | Degree-preserving rewiring of the derived structure. |
| `E_on` | Online-derived context | A separate secondary test of incremental organization. |

## What is measured?

- **Primary metric** — EvidenceRecall@10; LongMemEval-S is confirmatory, LoCoMo is directional replication.
- **Control gates** — Null and neighbourhood comparisons, stability and interpretability are considered together.
- **Cost and scope** — Report candidate budgets, latency and exclusion rules. Partially resolved gold evidence is not rescued by shrinking the gold set.

## From calibration to the decision gate

This is a read-only review dated 23 October 2026 of local design documents and calibration reports only. No new experiment was run. Raw conversations, decision questions and decision-split retrieval outcomes were not opened.

### A · shared retrieval settings

Recorded checks passed

Encoder and BM25 selection remain in the calibration report. This is not an H1a hypothesis result.

### B · retrieval budget

Recorded checks passed

n, m and C choices are fixed in a separate report; decision metrics do not drive new selection.

### C · context configuration

Recorded checks passed

Offline and online context selection records remain preserved. They do not independently establish superiority.

### v0.4 · null mixing calibration

mixing_plateau_not_reached

The required α could not be selected. Report checks passed, but the calibration condition did not; the confirmatory experiment was not reached.

### v0.5 · method amendment

Draft v0.7 / 14 October

Changes only null-model mixing adequacy. The hypothesis, datasets and primary success criteria are unchanged.

### Reference machine and Lock 2

Separate reproduction gate

The inspected gate report has all_checks_pass=false and retrieval_used=false. It does not show that required transfer and locking gates passed on the new reference machine.

Passing technical report checks does not mean passing a research gate. The v0.4 closure neither supports nor falsifies H1a: the decision partition was not reached.

## Primary source register

These documents were inspected in the local research source. Public repository access could not be verified, so no guessed original-file links are supplied. The downloadable guide contains scope, citations and file hashes, not the original package or raw data.

### Constitution · v0.4

Source, representation, usage and the minimal/broader architecture distinction.

### Reference architecture · v0.4

Records, memberships, contexts, derived views and retrieval responsibilities.

### Research program · v0.4

Hypothesis scopes and separate tests beyond H1a.

### Literature and decision record · v0.4

Related work and limits of research positioning.

### H1a preregistration · v0.4

Primary contrast, arms, data separation, metrics and locking rules.

### H1a method amendment · v0.5 · draft v0.7

Pre-decision mixing-adequacy change and transfer conditions.

### A–C calibration reports · Historical record

Separate selection records for shared retrieval, budgets and context settings.

### Null α calibration report · Historical v0.4 record

mixing_plateau_not_reached; α was not selected and no decision result followed.

### Reference-machine gate · Inspected record

Failed technical gate and a record that retrieval metrics were not used.

### Research decision log · Record through 14 October

Closure, amendment, reference-machine deviation and Lock 2 gates.

## Source files and SHA-256

- `docs/00-anayasa-v0.4.md`
  - SHA-256: `34393a9b9b8da9d250b6c2a1656ff0af13c4533cb6a96c6527efa1a1c629a933`
- `docs/01-referans-mimari-v0.4.md`
  - SHA-256: `d9b1b8847795b0e1b021b34a8878517a21e936f15fd7c113b548bb4d98b3b662`
- `docs/02-arastirma-programi-v0.4.md`
  - SHA-256: `501da391455e0955089369743e6bda3c80698c173caf257ef50ca3a3b9373014`
- `docs/03-literatur-karar-kaydi-v0.4.md`
  - SHA-256: `3adc6d4be1e210ffb9ae63c56f88a39af7e8535da7271e65e599177cd2ec2780`
- `docs/H1a-on-kayit-v0.4.md`
  - SHA-256: `6da25b5c92051667218563da6a1f5310e6cfe21ec3d672392abd1532fcafd118`
- `docs/H1a-on-kayit-v0.5-degisiklik.md`
  - SHA-256: `cc93e8fdcc577c4010e00f951f9b186b3b3ab7db77cccfc7527b59470ba0b68f`
- `DECISIONS.md`
  - SHA-256: `486d6eb4be6eaf8bed6e313af47ed9264087fb158c99e96e7e41478b6add065a`
- `reports/calibration/stage_a_report.json`
  - SHA-256: `16c85bb5b6322f29e8865afc63e7681848b896bd3b373e6c5f7bde2dbac1270b`
- `reports/calibration/stage_b_report.json`
  - SHA-256: `6b692be006ddf1f33ceafdb095ded3358f2678600d91b84dae0c55ebbd2c279c`
- `reports/calibration/stage_c_report.json`
  - SHA-256: `79a4da8cf50b0f7d08bd207163ce0fd96aaf556ea62da0706424c0c3285804ea`
- `reports/calibration/alpha_report.json`
  - SHA-256: `6a59b1e39b0de971294853c5fbd8252f43553b7340817c435fc395746ba574f6`
- `reports/calibration/refmachine_gate_report.json`
  - SHA-256: `be359f31cfcac7a021d72489bc89df56b24f37e5b33f7aa7ab6259186d3805c0`

## Dataset and encoder sources

- [LongMemEval research repository](https://github.com/xiaowu0162/LongMemEval)
- [LongMemEval cleaned dataset source](https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned)
- [LoCoMo research repository](https://github.com/snap-research/locomo)
- [E5-base-v2 model card](https://huggingface.co/intfloat/e5-base-v2)

## Status and open limits

Prepared on 23 October 2026 for RGM / Relational Space H1a, using the v0.4 calibration closure and the v0.5 amendment draft dated 14 October. The inspected records do not establish a final E_off − F superiority result or a completed Lock 2.

Calibration and decision partitions must remain separate. Carrying A–C from the old Linux reference to a new machine is a recorded deviation and reproduction obligation. Documenting an amendment does not establish completion of a signed amendment lock or a new confirmatory run. Retrieval value from derived contexts is not general meaning learning or completion of the broader architecture.

Calibration ≠ final decision. Technical checks ≠ hypothesis support. Derived context ≠ learned general meaning.

## Separate line: ERM relational-memory prototype

ERM v0.12 evaluates patterns and declarations within the same event frame. It is not this H1a retrieval mechanism, and neither line’s results establish the other’s success.
