Internal stress test · Memory Pressure

Stop outdated facts from reaching the live answer path.

An internal seeded-authority memory-pressure benchmark measuring stale and superseded answer errors: cases where plausible old records compete with approved/current memory for the next answer.

Approved/current recovery

600/600

seeded reviewed/current authority state

Stale-context leakage

0/600

seeded authority state; governed path under exact-token scoring

Clean exact answer

600/600

approved current token recovered

Candidate memories exposed

2.00/10

mean answer-path exposure

The everyday version is simple: a team approved one authoritative memory, but older plausible answers are still in the archive. Standard retrieval can surface the old version as if it were still current. FieldHash keeps the archive intact while enforcing which version is allowed to influence the model after authority state has been set.

This is not the AI making something up. It is often harder to catch: the AI answers with a fact that was true once, still looks relevant, and is no longer authorized to govern.

Superseded contract clause

An older clause still matches the question, but should no longer shape the answer.

Deprecated API endpoint

The old integration path remains documented, but should be blocked from live model input.

Obsolete compliance standard

A prior policy can stay auditable without becoming the next recommendation.

Rejected research hypothesis

Discarded analysis remains in the record while current conclusions govern.

Chapter I

The Governance Gate Simulation

Observe the mechanism in action. Toggle between standard and governed memory states to see how Governed Memory enforces reviewed/current state before prompt assembly.

Interactive Governance Gate Simulation

Toggle to observe governed-state enforcement under controlled conflict.

Retrieved Context (81 records total)Showing 6 of 81
AXIOM_014_CURRENT
approved

Project Alpha controlling codeword. Validated by Governed Memory gate.

canonical_status: currentreview_status: reviewed
DECOY_014_03
stale

Prior operational codeword. Superseded 2 weeks ago.

canonical_status: supersededreview_status: unreviewed
DECOY_014_01
stale

Codeword draft. Rejected during static scan.

canonical_status: ignoredreview_status: rejected
DECOY_014_02
stale

Codeword proposal. Never promoted.

canonical_status: ordinaryreview_status: unreviewed
DECOY_014_04
stale

Codeword alternative. Superseded.

canonical_status: supersededreview_status: unreviewed
DECOY_014_05
stale

Prior run tag. Outdated.

canonical_status: supersededreview_status: unreviewed
+75 stale context records repeating in unstructured history...
Base LLM Output Log
Idle
_
Execution Assessment:❌ Wrong Codeword (Stale)
Chapter II

Stale and Superseded Answer Errors

The primary failure mode in this benchmark family isn't that the model fails to remember. It is that the model remembers the wrong version. In regulated and enterprise environments, persistent memory is what lets an agent carry approved decisions forward between sessions. It is also the channel most prone to hidden staleness.

Adding more context, the common remedy for retrieval gaps, did not resolve supersession conflicts in these runs. The arms that exposed all ten candidate memories produced more stale and superseded answer errors than the governed arm that exposed a mean of two.

In production, stale memory is not a harmless retrieval miss. It surfaces as the superseded contract clause, the deprecated API endpoint, the obsolete compliance standard, or the rejected research hypothesis influencing a live answer. Standard memory systems can rank candidates by semantic similarity, recency, or local metadata. Without active answer-path governance, standard memory can allow obsolete data to reach the answer path when no explicit governance gate intervenes.

FieldHash handles this by making authority explicit at the handoff. In this diagnostic, that means evaluating review_status and canonical_status before prompt assembly. In production, those signals can come from review workflows, systems of record, supersession links, or rollback events. The schema fields are plumbing; the product is enforcing them before generation and preserving the evidence.

The buyer-relevant claim is enforcement after authority exists: given a reviewed/current signal, can the answer path keep stale memories from steering the model while preserving them for review? Upstream promotion from configured authority evidence is measured separately in the automatic-promotion diagnostics.

The evidence catalog keeps this page in context with the adjacent retrieval-vs-resolution question: whether a memory system can surface the current record in candidates but still allow stale context to shape the answer. This page remains the seeded-authority pressure test; the supporting diagnostics track the same failure mode.

Chapter III

The Performance Portfolio

A unified look at approved/current recovery, stale-context leakage, clean exact answers, candidate-memory exposure, and cross-provider replication under controlled conflict.

I. Approved/Current Recovery

Frontier LLMs + FieldHash governed state

600 / 600 clean exact

Same frontier LLMs + retrieval-only context

415 / 600 approved/current

Gemini Run

Full evaluation run with row-level provider/model guard.

+60 exact

LLM + FieldHash200/200
Same LLM + retrieval-only memory140/200

Claude Opus 4.7 Run

Full evaluation run; retrieval-only recovered 140/200. Three control bridge errors are counted as baseline misses.

+60 exact

LLM + FieldHash200/200
Same LLM + retrieval-only memory140/200

GPT-5.5 Run

Full evaluation run with row-level provider/model guard.

+65 exact

LLM + FieldHash200/200
Same LLM + retrieval-only memory135/200

Across three distinct n=200 provider runs, the governed-memory path recovered the approved canonical decision with 100% precision when reviewed/current authority metadata was present. Standard retrieval-only context missed 185 cases when that structured governed state was withheld and all memories were exposed as ordinary context.

How to read this fairly: a plain metadata filter (status = current) over the same seeded labels would also select the approved record. On a separate public RFC authority task, a metadata-aware deterministic baseline reached 299/300 on visible-authority selection; that is different from the MemConflict strong labeled prompt result, where an LLM was shown both governing and blocked records and scored 297/300. This study measures enforced answer-path exposure control under pressure, supersession and rollback handling, and reviewer-inspectable evidence of what was allowed, blocked, and why. Selection superiority remains outside the claim.

II. Candidate Exposure and Stale Suppression

Gemini Run

80.00% less · 2.00 / 10 shown

Claude Opus 4.7 Run

79.95% less · 2.00 / 10 shown

GPT-5.5 Run

80.00% less · 2.00 / 10 shown

By applying governance gates before prompt compilation, FieldHash exposed the base models to the approved-current record plus one supporting record on average, while suppressing stale competitors from the answer path. The diagnostic measures influence control. Broad token-savings claims remain outside it.

Claude's 79.95% exposure reduction is reported from the guarded run as logged. The displayed 2.00/10 exposure rate rounds to the same average as Gemini and GPT; the non-round percentage reflects the run's control bridge miss accounting, not additional stale exposure in the governed answer path.

III. Multi-Provider Telemetry

Gemini Synthesis

Exact Recall

200/200

Exposed

2.00 / 10

Retrieval-only baseline recovered 140/200 with 42 decoy mentions.

Claude Opus 4.7 Synthesis

Exact Recall

200/200

Exposed

2.00 / 10

Retrieval-only baseline recovered 140/200; three control bridge errors are counted as misses, not governed-path leaks.

GPT-5.5 Synthesis

Exact Recall

200/200

Exposed

2.00 / 10

Retrieval-only baseline recovered 135/200 with 50 decoy mentions.

Prompt-Only Instruction Diagnostic

To test whether ordinary context simply needed stronger instructions, we gave the baseline models explicit directions to prioritize current/reviewed records and use recency as a tie-breaker. Recall improved to 464/600 but still failed on 136 cases. This supports the enforcement result, but it is not a same-information comparison because the governed path reads structured authority state that the prompt-only baseline does not receive.

Recall

464/600

Exposed

10.00 / 10

Chapter IV

Technical Scope for Reviewers

The rigorous mathematical constraints, baseline parameters, and structural boundaries of the evaluation.

Harness Parameters

Paired Adversarial Testing

Each evaluation item generated a fresh synthetic state containing 1 reviewed/current project decision and 80 plausible historical decoy fragments. The governed path received structured authority state; the baselines received ordinary retrieved context. The benchmark asks whether governed state is enforced under pressure without leaking the target codeword.

Total Cases

600

Decoys per Case

80

Methodology at a Glance

Task

Recover the reviewed/current codeword for a named project.

Seed

One governed reviewed/current fact plus 80 plausible stale or superseded context records per synthetic user.

Scoring

Deterministic exact-substring match against the approved codeword.

Runs

Three complete comparisons (Gemini, Opus, GPT) with row-level provider guards.

Smart Baseline

Added explicit current/reviewed and recency tie-breaker instructions, but not structured governance metadata.

Exposure Bound

The governed path filters active context down to 2 memories shown instead of 10.

Provenance

Scenario generated by an external model, run through the FieldHash harness.

Scope Boundary

Internal supported mechanism benchmark for seeded authority enforcement, not raw-text authority inference or third-party validation.

Schema Definitions & Validation

Retrieval-Only Baseline

A standard RAG architecture where retrieved memories are appended directly as prompt context. The approved fact is present, but the answer path does not receive FieldHash governance metadata, supersession handling, or governed arbitration.

Stale/Superseded Answer Error

The primary error mode where a base model accepts a superseded, outdated context record as current, outputting obsolete parameters that still look plausible because they came from a real older record.

The "Approved" Schema

Explicit metadata flags mapping to review_status: reviewed and canonical_status: current. In this diagnostic, those flags are seeded by the harness to test enforcement after authority state exists; in production they are set via user validation or trusted ingestion gates.

Biotech & Workflow Context

An illustration is biomarker panels (e.g. changing biomarker selection criteria APOE-3 vs APOE-4 across study iterations). Governed memory quarantines historical records while preserving audit trails.

What This Benchmark Supports

Approved decisions carry forward safely under context pressure after authority state exists, and structural governance minimizes model exposure to obsolete data without loss of precision. Private local/on-prem audit diagnostics exist for technical review, but this public page does not use them as headline evidence. The public claim here is governed-state enforcement under controlled memory pressure, not open-ended intelligence.

Verified Scope

Boundaries for Reviewers

The benchmark is narrow by design: seeded authority enforcement under controlled stale-context pressure. It is not independent external validation, legal compliance evidence, token-savings evidence, or a model-superiority claim. Local audit portability, arbitrary prose inference, and learning over time are evaluated separately and should not be inferred from this page.

Explicit Limits

Technical diligence

The benchmark is narrow by design.

The tested Governed Memory behavior is specific: reviewed state survives, stale and superseded context is filtered, and the resulting answer remains inspectable after authority has been established. Universal intelligence remains outside the claim.

Read loop synthesisDownload methods report