HOLO-Invariant

1. Executive Summary

HOLO-Invariant is "AI continuity that preserves corrections instead of hiding them" — MIT, Python 3.10+, 91,867 lines across roughly two hundred modules with 1,603 test functions in 218 test files, and zero runtime dependencies. It is a framework for preserving, reconstructing and checking state across sessions, models, tools and execution environments.

Its README states the contrast as a table, and every row is a failure this atlas reports somewhere: correction replaces history, the latest value hides its lineage, resume trusts supplied context, memory authority is implicit. Against each: correction becomes a verified relation, original and correction and target all stay inspectable, resume checks externally retained evidence, authority stays explicit and bounded.

The benchmark is the reason to read it. Five bounded metrics over one committed fixture:

Four things make that table mean something. The fixture is public and hashed, and "fixes the target before results are observed". The condition schema is closed "so undeclared fields cannot alter the scoring contract" — a candidate cannot add a field that changes how it is scored. The reference result is "regenerated from the same committed fixture and condition on every change" in CI. And the comparison test asserts holo_reference["fixture_hash"] == fixture["fixture_hash"] before reading a single metric.

Then it scores the trivial alternative, and reports which metrics the trivial alternative wins. A plain latest-value store is run against the identical fixture, and the test records that it passes two of the five: its latest_justified_recall is 1.0, and its superseded_resurrection_count is 0 — "[i]t does not resurrect the overwritten value as current." It fails exactly three: uncertainty recall, lineage recall, and stale-continuation blocking. And the deltas are asserted to be zero on the two the baseline already wins, under a comment that states the claim in one sentence:

"The difference is specifically uncertainty + lineage + stale-continuation behavior, not latest-value recall."

This atlas has read a great many benchmarks. Very few name the metrics their baseline already passes. Doing so converts a five-for-five scoreboard into a narrow claim — this system represents uncertainty, lineage, and staleness, and ordinary storage does not — which is both smaller and worth something, because it can be checked and could have come out the other way.

The pass condition is why this earns the negative-evaluation mark rather than just admiration. passes_bounded_continuity_fixture is

latest_recall == 1.0 and not resurrected and uncertainty_recall == 1.0
and lineage_recall == 1.0 and stale_blocked

— a must-not-retrieve assertion (no superseded claim comes back as current) that cannot pass vacuously, because the same condition demands full recall of everything that is current. A store that returned nothing would fail the first clause. The control is in the pass condition itself.

The artifacts refuse to claim more than they show. Every result payload carries truth_claimed: false and accepted: false. The demo's description ends: "[i]t does not modify an existing chain or grant truth, acceptance, write, or execution authority." A correction is bound to one exact baseline transition candidate and sits at READY_FOR_EXACT_TARGET_AUTHORIZATION until authorized against that target — the same bind-to-one-operation property this atlas found in Anda DB's approvals and Edda's verdicts.

What this is not: a memory an agent writes to. There is no store with a read path, no scoping over stored content, and no mutation record — the transitions are the subject. A reader arriving for a memory backend will find a scoring contract and a correction model, which is the valuable thing here and not the thing the category name suggests.

And the metrics are bounded in the literal sense the code names them: five properties over one fixture of a chosen shape. Passing says the system represents uncertainty, lineage and stale continuation on that fixture; generalisation is the reader's inference and not the benchmark's claim — which passes_bounded_continuity_fixture concedes in its own identifier.

2. Mental Model

A correction is a relation, not a replacement.

A fixture is hashed before anyone runs against it.

A baseline is scored on the same target, and told what it won.

A result records that it claimed no truth.

%% caption: the pass condition pairs a must-not-resurrect assertion with full recall of current claims, and the naive baseline is scored on the same hash-pinned fixture
flowchart TB
    FIX[("benchmarks/continuity-v1.fixture.json —<br/>public, hashed, 'fixes the target<br/>before results are observed'")]
    SCH["schemas/continuity-condition.schema.json —<br/>closed, 'so undeclared fields cannot<br/>alter the scoring contract'"]
    FIX --> SCORE{"score a condition"}
    SCH --> SCORE
    SCORE --> M1["latest_justified_recall"]
    SCORE --> M2["superseded_resurrection_count<br/>= len(superseded_resurrected_as_current)"]
    SCORE --> M3["uncertainty_recall"]
    SCORE --> M4["lineage_recall"]
    SCORE --> M5["stale_continuation_blocked"]
    M1 & M2 & M3 & M4 & M5 --> PASS{"passes_bounded_continuity_fixture =<br/>latest_recall == 1.0 AND not resurrected<br/>AND uncertainty_recall == 1.0<br/>AND lineage_recall == 1.0 AND stale_blocked"}
    PASS -.->|"the must-not-resurrect clause cannot<br/>pass vacuously — the same condition<br/>demands full recall of what IS current"| CTRL["the control is in the pass condition"]
    BASE["a naive latest-value store,<br/>scored on the IDENTICAL fixture"] --> SCORE
    BASE --> WINS["wins two metrics:<br/>latest_justified_recall = 1.0<br/>superseded_resurrection_count = 0"]
    BASE --> LOSES["fails three:<br/>uncertainty_recall = 0.0<br/>lineage_recall = 0.0<br/>stale_continuation_blocked = false"]
    WINS -.->|"deltas asserted to be exactly zero:<br/>'the difference is specifically uncertainty +<br/>lineage + stale-continuation behavior,<br/>not latest-value recall'"| CLAIM["a narrow claim instead of<br/>a five-for-five scoreboard"]
    HASHCHK["the comparison asserts both results<br/>share a fixture_hash before reading a metric"] --> SCORE
    OUT["every result payload:<br/>truth_claimed = false,<br/>accepted = false"] --> PASS
    CORR["a correction"] --> CAND["bound to ONE exact baseline<br/>transition candidate —<br/>READY_FOR_EXACT_TARGET_AUTHORIZATION"]
    CAND -->|"authorized against that exact target"| NEWBASE["becomes the baseline;<br/>original, correction and target<br/>all remain inspectable"]

3. Architecture

4. Essential Implementation Paths

holosim/continuity_baseline_benchmark.py:153-190 — the five metrics and the pass condition that keeps one of them honest.

tests/test_latest_value_continuity_baseline.py:75-150 — the baseline, what it wins, and the deltas asserted to zero.

benchmarks/continuity-v1.fixture.json — the target, fixed and hashed.

5. Memory Data Model

A claim with retained evidence, correction edges forming a lineage, and an uncertainty record. The model's commitment is that all three survive a correction: "[o]riginal, correction, and target remain inspectable."

6. Retrieval Mechanics

None of its own. Reconstruction and checking rather than search.

7. Write Mechanics

A correction becomes a transition candidate bound to one exact baseline, held at READY_FOR_EXACT_TARGET_AUTHORIZATION until a typed authorization for that exact target arrives. Binding the authorization to the target is what stops one approval from licensing a different transition.

8. Agent Integration

A CLI and a system-neutral condition document, so another continuity system can be exported into the same shape and scored against the same fixture. That is an unusual offer — a benchmark designed to be run against its author's competitors on equal terms — and its value depends entirely on the fixture being fair, which is why the hash pinning and the baseline scoring matter as much as the metrics.

9. Reliability, Safety, and Trust

The trust story is methodological rather than architectural: the fixture, the closed schema, the CI regeneration, the baseline comparison, and the payload fields that decline to claim truth or acceptance.

10. Tests, Evals, and Benchmarks

1,603 test functions across 218 files, including tests that assert the README's own table matches the regenerated reference result — so the published number cannot drift from the computed one. Nothing was installed or run for this reading; the numbers quoted are the committed ones.

11. For Your Own Build

Score the trivial alternative on your own benchmark, and publish what it wins. A metric the naive baseline already passes is a metric your system should not take credit for, and saying so is what turns a scoreboard into a claim.

Hash the fixture and close the schema. "Fixes the target before results are observed" and "undeclared fields cannot alter the scoring contract" are two sentences that prevent most of the ways a benchmark stops being one.

Put the control in the pass condition. Requiring full recall of current claims alongside zero resurrections means the negative half cannot be satisfied by a system that returns nothing.

And let your artifacts say what they do not establish. truth_claimed: false in every result is a small field that keeps a downstream reader from importing a score as a verdict.

12. Open Questions

Whether the authorization is human. Transitions are bound to an exact target and typed; who supplies the authorization was not traced.

How the fixture was chosen. It is fixed and hashed, which makes it fair; whether its shape favours the model it was built alongside is a question the fixture cannot answer about itself.

What the spine protocol contributes. At 1,641 lines it is the largest module and sits outside the benchmark path read here.

Appendix: File Index

History

2026-09-16 — 2d36396a… — first reading, at a commit dated 15 September 2026. Screened before opening, from a shallow clone: four files scanned, no auto-run surfaces, no build-time execution points, one unpinned surface and one dependency file inside the seven-day cooldown. Nothing was installed, built or run, and no benchmark was executed.