Tessellum

1. Executive Summary

Tessellum is "[t]yped atomic notes in a graph — a Zettelkasten that scales" — MIT, Python, 138,183 lines with 2,905 test functions across 179 test files. Its unit is a tessellum, a small mosaic tile: a note small enough to make one point and tagged with what kind of point it makes. Underneath is a one-way CQRS split — the authored notes are the source of truth, the searchable index is a projection rebuilt from them, "and changes only ever flow one way — from the notes to the index, never back". On top is a dialectic cycle that turns an observation into competing arguments, records where they conflict, names which part of the losing argument broke, and emits a revised rule, with a Dung grounded-semantics solver deciding which arguments survive.

It says in its second sentence that it "is a knowledge-construction system, not an agent-memory store". It is in this atlas because it keeps claims that can turn out to be false and gives them an identity a correction can name — which is this atlas's scope rule — and because of one module.

The demotion gate. Its header states the finding this atlas's own corpus keeps producing, and states it better:

"Promotion turns an episodic derivation into durable knowledge; a promoted claim that stops being true has no way to notice on its own, so a promotion path without a demotion path is a mechanism for entrenching whatever was believed first. That is not a hypothetical failure mode — it is the one the memory literature documents most consistently — which is why this gate ships ahead of consolidation rather than beside it."

Shipping the demotion path before the promotion path it guards is the ordering decision, and almost nothing in this corpus made it.

The gate scores an abstraction for truth rather than usefulness — "systems that abstract memory score an abstraction for utility (was it recalled? did it help?) and none of them scores a produced abstraction for truth" — through a blinded protocol in three parts, each closing a way the test could cheat:

Suppress. The gate "hands over neither its text nor its id. ReDerivationRequest has nowhere to put them, and require_suppression refuses a caller who smuggles the claim into the question or cites the claim's own rendering as its own source." The blinding is a property of the request type, not a convention — the same discipline codemem applies by subtracting a flag from a context.

Freeze the model. An explicit model_id and frozen_at, enforced, because "[a] model that re-tunes on the corpus it is checking would eventually regenerate its own promoted claim from the promotion, and the gate would certify itself."

Decide by arithmetic. A token-overlap ratio against a floor, and every trigger likewise, because "[a] model must never decide the verdict: a demotion nobody can recompute is a demotion nobody can appeal."

Three triggers, and the first is the one nothing else looks for. rederivation_failure fires when the frozen model does not regenerate the claim from its cited sources "even with no attack against it anywhere in the log". Every argumentation-driven system in this corpus notices a claim when something argues against it; this one notices a claim that quietly stopped following from its own evidence. contradiction fires when the computed status flips out of answerable — "[n]early free: status is a pure function of the edge set". independence_below_floor fires when the origin-distinct contexts behind a claim fall under a floor.

The entrance and the exit used to measure different things under one name. The independence trigger counted episodes while the promotion gate counted origin-distinct contexts — "the entrance strict, the exit lax" — and both now consume one measurement function. A field measured one way on the way in and another on the way out is a defect this atlas has found repeatedly; here it is named and closed.

And the missing-data convention inverts between them, which is the subtle part. Applying the entrance's conservative lower bound at the exit "would let a hole in the log lower the count and demote a sound claim". So the exit fires conclusively only when the optimistic bound misses the floor; where the floor is missed only because origins are unreported, the finding is inconclusive and the claim is "withheld — quarantined, never retracted". A conservative bound is safe admitting and dangerous removing, and the direction has to flip. Very few systems notice that, and this one has a named rule for it.

Five outcomes, each carrying whether it is conclusive, under a paragraph that refuses to overstate what the method can show:

"collapsing every finding into 'the claim is false' overstates what this protocol can establish: a frozen model can reproduce a source's error perfectly consistently, and it can fail to reproduce one of several legitimate abstractions of the same sources. Re-derivation tests reproducibility and fidelity, not world truth, causal validity or transfer."

failed_reproduction and grounding_failure are conclusive; stale_evidence and surviving_contradiction depend — the latter "conclusive once the labelling settles against the claim; a live undecided dispute (Dung undec) has settled nothing"; and failed_generalisation is never conclusive, because two legitimate abstractions of the same sources can disagree.

No marks. The statuses here are computed from the argument edge set rather than stored, the scope is one knowledge base, and the atlas's marks attach to a store's read path. That is a statement about the mark definitions, not about the quality of what is here.

Two limits the project states about itself are worth carrying out. By default the dialectic cycle "treats two arguments as conflicting when their claims are worded differently", with evidence-based incompatibility and attack direction an opt-in mode — so the default conflict detector is textual, which is exactly the failure chitta-field's claim-centric detector is built to avoid, and this project ships the better mode behind a flag while saying which is the default. And each cycle is labelled on its own, "so a later argument does not yet overturn an earlier cycle's verdict" — a stated boundary on how far a verdict travels.

2. Mental Model

A note makes one point and says what kind of point it is.

A promotion must have an exit before it is allowed an entrance.

A re-derivation asks a blinded, pinned model to rebuild the claim from the claim's own sources — and counts tokens, not opinions.

Inconclusive means quarantine, not retraction.

%% caption: the gate blinds the claim by construction, pins the model so it cannot certify itself, and decides arithmetically — and an inconclusive finding quarantines rather than retracts
flowchart TB
    OBS["an observation"] --> CYCLE["dialectic cycle: competing arguments,<br/>a counter, a revised rule;<br/>Dung grounded semantics labels survivors"]
    CYCLE --> DERIV["an episodic derivation"]
    DERIV --> PROMO{"promotion gate — counts<br/>ORIGIN-DISTINCT contexts"}
    PROMO --> CLAIM[("a promoted, durable claim")]
    CLAIM --> GATE["the re-derivation gate:<br/>REQUIRED AND SCHEDULED,<br/>and it shipped BEFORE promotion"]
    GATE --> S1["1. SUPPRESS — ReDerivationRequest<br/>'has nowhere to put' the claim's text or id;<br/>a caller smuggling it into the question,<br/>or citing the claim's own rendering, is refused"]
    S1 --> S2["2. FREEZE — explicit model_id + frozen_at,<br/>because a model re-tuned on the corpus<br/>'would eventually regenerate its own promoted<br/>claim from the promotion, and the gate<br/>would certify itself'"]
    S2 --> S3["3. ARITHMETIC — token-overlap against a floor;<br/>'a demotion nobody can recompute<br/>is a demotion nobody can appeal'"]
    S3 --> TRIG{"three independent triggers"}
    TRIG -->|"rederivation_failure"| T1["the claim did not come back out of its<br/>own sources — EVEN WITH NO ATTACK<br/>anywhere in the log"]
    TRIG -->|"contradiction"| T2["computed status flipped out of answerable<br/>— status is a pure function of the edge set"]
    TRIG -->|"independence_below_floor"| T3["origin-distinct contexts fell under the floor<br/>— the SAME measurement the entrance used,<br/>after the two measured different quantities<br/>under one name"]
    T1 & T2 & T3 --> CONV{"missing-data convention INVERTS at the exit"}
    CONV -->|"optimistic bound misses the floor"| CONC["conclusive — demote"]
    CONV -->|"floor missed only because origins<br/>are unreported"| QUAR["inconclusive — WITHHELD,<br/>quarantined, never retracted<br/>(the entrance's conservative bound here<br/>would demote a sound claim)"]
    CONC & QUAR --> FIVE["five outcomes, each carrying conclusiveness:<br/>failed_reproduction · grounding_failure (yes) ·<br/>stale_evidence · surviving_contradiction (depends) ·<br/>failed_generalisation (never)"]
    FIVE -.->|"'Re-derivation tests reproducibility and fidelity,<br/>not world truth, causal validity or transfer'"| LIMIT["what the protocol cannot establish"]

3. Architecture

4. Essential Implementation Paths

src/tessellum/dks/demotion.py:1-100 — read the module docstring end to end; it is the report.

src/tessellum/dks/consolidation.py — measure_origin_distinct, the one measurement both gates now consume.

5. Memory Data Model

Typed atomic notes with a Building-Block type system, linked into Folgezettel trails that "record how one idea led to the next". Claims promoted out of the dialectic cycle carry their cited sources, which is what makes re-derivation possible at all: a claim with no usable evidence produces grounding_failure rather than a pass.

6. Retrieval Mechanics

Hybrid keyword and vector search over the projection. The query protocol abstains with status_not_answerable rather than returning a claim whose argument labelling does not support it — an abstention, not a filter, and computed rather than stored.

7. Write Mechanics

Author or digest; the index follows. The CQRS direction is the invariant: "the notes you author are the source of truth, the searchable index is a projection rebuilt from them, and changes only ever flow one way".

8. Agent Integration

A Python package and CLI, for humans and agents both. It does not present itself as a memory an agent writes to during a session, and the report should not either.

9. Reliability, Safety, and Trust

The gate is the trust story, and its most valuable property is epistemic modesty expressed as structure: five outcomes with conclusiveness attached, a stated list of what re-derivation does not establish, and an inconclusive branch that quarantines. A system that can distinguish "this is false" from "I could not tell" is doing something most of this corpus does not attempt.

10. Tests, Evals, and Benchmarks

2,905 test functions across 179 files. Nothing was installed or run for this reading.

11. For Your Own Build

Build the demotion path before the promotion path. A promoted claim cannot notice that it stopped being true, so an entrance without an exit entrenches whatever was believed first — and the ordering here, exit first, is the concrete form of taking that seriously.

Blind the test by construction. If the request type has nowhere to put the claim, no reviewer has to check that nobody leaked it.

Pin the model that checks your model. One that re-tunes on the corpus it audits will eventually reproduce the claim from the promotion and certify itself.

Decide with arithmetic. "A demotion nobody can recompute is a demotion nobody can appeal" is the argument for a token-overlap floor over a judge.

Invert the missing-data convention between your entrance and your exit. A conservative bound is safe when admitting and dangerous when removing, and using the same one in both directions demotes sound claims on a gap in the log.

And attach conclusiveness to every finding. "I could not tell" is a real answer and collapsing it into "false" is how a correction mechanism starts causing the errors it was built to catch.

12. Open Questions

Whether a quarantined claim ever returns. The inconclusive branch withholds rather than retracts; the path back was not traced.

How the opt-in evidence-based conflict mode performs against the textual default. The project ships both and names which is the default; no comparison was found.

What the labelling boundary costs in practice. Each cycle is labelled on its own, so a later argument does not reach back into an earlier verdict.

Appendix: File Index

History

2026-09-16 — 6e167169… — 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 two dependency files inside the seven-day cooldown. Nothing was installed, built or run.