Products · SynaptYon SARL

Document Intelligence, from the first page to the last answer.

SynaptYon SARL, which I co-founded, builds the whole chain, from ingestion to downstream applications, for business, financial, historical and healthcare documents, in four products: read, extract, ask and remember. None is generally available yet. The memory layer comes from Amnesia by Design, Memory by Necessity, and below it is shown in research detail, with demos that run in your browser on fictional documents.

§ 01 The chain

Four products, one chain.

Read every page, extract the data you need, ask questions with sources, and keep what was learned for the next document. Each product links to its page on the SynaptYon SARL site.

  1. Grammata

    Read. Every page, read: printed or handwritten, new or centuries old.

  2. Kleis

    Extract. The data you need, from any document.

  3. Sibyl

    Ask. Ask your documents. Get answers you can check.

  4. Mnemos

    Remember. Document AI that gets better with every document.

    Research view: Mnemos

§ 02 Built for

One chain, six document families.

Each family draws on part of the chain. These are the documents the products are designed for, not a list of customers.

§ 03 The research view

The memory layer, from the paper.

Mnemos is where the chain keeps what it reads. Its six parts implement the persistent-state architecture specified in Amnesia by Design, Memory by Necessity: read a document, gate what it says, consolidate it, reconcile it with what came before, keep the trail back to its source, forget it on request, and measure whether any of it helped. Each layer names the part of the paper it implements, the product that owns it, and the work it builds on.

The stack · 7 layers, 6 products Interactive
  1. Acquire Read the whole document

    Extract the multimodal observation: text, layout coordinates, table structure and visual elements, for questions not yet asked.

    Paper
    Observe · Dt = (V, T, B, G, H, M)
    Product
    Mnemos In development
    Research
  2. Validate Gate before writing

    Corroborate each observation against the provenance ledger and reject it when confidence is below threshold, leaving state unchanged.

    Paper
    Pre-write gate · Theorem III.4
    Product
    Argus In development
    Research
  3. Consolidate Write and abstract

    Append admitted evidence with its timestamp and provenance, then abstract recurring families into reusable schemas.

    Paper
    Et, Kt, Γt · max J(α) − λR(α)
    Product
    Mnemos In development
    Research
  4. Reconcile Know what is in force

    Resolve conflicts with one of five outcomes and close superseded claims at their effective time, keeping history for audit.

    Paper
    Temporal supersession · 𝒱t
    Product
    Chronos In development
    Research
  5. Trace Cite the box

    Keep the chain from every conclusion back to document, region and version, through consolidation, supersession and repair.

    Paper
    Provenance invariant · Pt
    Product
    Ariadne In development
    Research
  6. Forget Erase, and prove it

    Remove a target across every store, repair the schemas that depended on it, and certify what can no longer be recovered.

    Paper
    Five-level unlearning · Algorithm 2
    Product
    Lethe In development
    Research
  7. Measure Show that memory helps

    Compare warm, retrieval-only and cold arms on the same future documents and report benefit, cost, risk and governability together.

    Paper
    Longitudinal harness · five metrics
    Product
    Anamnesis In development
    Research
Each layer lists the concept it implements in the papers, the product it belongs to with that product’s current status, and the published work it builds on. Pick a layer; a product chip marks the layers that product covers.

§ 03.1 Acquire · Consolidate

Document memory

In development

Part of Mnemos · codename Mnemos

Documents, remembered.

St = (Et, Kt, Γt, 𝒱t, Pt, At)

Mnemos keeps what your Document AI reads. Every validated observation is written into six linked stores (evidence, schemas, relations, validity, provenance and policy), so the next contract, invoice or report is read against everything that came before it.

The problem

Today's systems can parse a 200-page filing and ground an answer to a single table cell, then reset. The system that reads the ten-thousandth invoice is, computationally, the one that read the first.

For: Teams with recurring document streams: contracts and amendments, supplier invoices, clinical reports, research literature.

How it works

  1. Read everything once

    Each document becomes a tuple of pixels, tokens, boxes, structure graphs, parse trees and metadata. The observation is document-conditioned, not query-conditioned, so it serves questions nobody has asked yet.

  2. Write with provenance

    Evidence that passes the gate is appended to the episodic store with a timestamp and a link back to its source region.

  3. Consolidate what recurs

    Repeated document families become schemas, chosen by expected future utility minus storage cost. Common queries use the schema; tail queries still reach the episodic detail.

  4. Answer from state

    Queries are answered from the updated state, so reading the first document changes how the second is understood.

State lifecycle · Eq. 4 · Fig. 3 · Algorithm 1 Interactive Illustrative

St = (Et, Kt, Γt, 𝒱t, Pt, At)

Et Episodic evidence buffer
Admitted observations, appended with a timestamp and their provenance.
Kt Consolidated semantic store
Schemas and facts abstracted from episodic traces.
Γt Associative relation graph
Relations between entities, documents and schemas.
𝒱t Temporal validity index
When each claim holds; closed when newer evidence supersedes it.
Pt Provenance ledger
Links every claim back to the document, region and version it came from.
At Authorization and policy state
Who may read or change what; updated only by policy events.

Invariant. Provenance is an invariant across every operation.

Algorithm 1 · State transition, one session

  1. Observe Fig. 3 · Acquire

    Read the incoming document Dt as (V, T, B, G, H, M).

    Reads: — Writes: —

  2. Corroborate

    Check the observation against prior provenance.

    Reads: Kt Pt Writes: —

  3. Validate Fig. 3 · Validate

    Pre-write gate: admit the evidence or leave the state unchanged.

    Reads: — Writes: —

    If the gate rejects, St+1 = St and the algorithm skips to Answer.

  4. Write Fig. 3 · Write

    Append the admitted evidence to the episodic store and extend provenance.

    Reads: — Writes: Et Pt

  5. Consolidate Fig. 3 · Consolidate

    Abstract episodic traces into schemas and relations.

    Reads: — Writes: Kt Γt

  6. Reconcile Fig. 3 · Update & reconcile

    Check against temporal validity and apply supersession.

    Reads: — Writes: Kt 𝒱t Pt

  7. Keep policy

    Authorization stays as is unless a policy event occurs.

    Reads: At Writes: —

  8. Forget (on request) Fig. 3 · Forget

    If a deletion request arrives, run unlearning across every store.

    Reads: — Writes: Et Kt Γt 𝒱t Pt

  9. Answer

    Answer the query from the updated state, citing provenance.

    Reads: Et Kt Γt 𝒱t Pt At Writes: —

The six components are Eq. 4 of the paper; the steps, their order and the components each one reads or writes follow Algorithm 1 and the lifecycle of Fig. 3. Register contents are our illustration on fictional documents: a dashed outline means the step reads a component, a tint means it writes it. Fictional documents. The demo runs entirely in your browser; it illustrates the ideas and is not production software.

Implements

  • §III-A Document tuple and state tuple, Eq. (3)–(4)
  • §III-B Per-session state transition, Algorithm 1
  • §III-C Architecture and degenerate regimes, Fig. 3
  • §III-E Substrate composition, Table III

Sections of arXiv:2609.32041, a specification rather than a validated implementation.

Harness metrics

Counterfactual metrics from the paper’s harness that apply to document memory. The paper specifies them; it reports no measurements.

§ 03.2 Validate

Quality gate

In development

Part of Mnemos · codename Argus

Nothing enters memory unverified.

Pcont(k) = 1 − (1 − pe · pc)k

A memory that stores its own mistakes gets worse with use. Mnemos sits between reading and remembering: every observation is corroborated against the provenance ledger and scored for confidence, and anything below threshold is rejected with the state left exactly as it was.

The problem

Writing a model's output straight into memory turns one misread tax ID or transposed table cell into a durable error that every later session inherits. Unvalidated writes drive contamination towards certainty.

For: Anyone who lets a model write to a store that other sessions read.

How it works

  1. Corroborate

    Retrieve supporting or contradicting evidence for the new observation from the prior provenance ledger.

  2. Score confidence

    Compute epistemic confidence as the calibrated complement of predictive entropy over the extracted fields.

  3. Admit or reject

    Below the confidence threshold nothing is written; the session still answers, but the state is unchanged.

  4. Measure the gate

    Estimate the false-accept rate by injecting controlled corruption into a parallel stream, then read off the safe horizon.

Theorem III.4 · Contamination bound Interactive Illustrative values

Contamination after k updates

Pcont(k) = 1 − (1 − pe · pc)k

Expected Memory Harm

E[MH(k)] = wm · Pcont(k) · DKL(Ptrue ‖ Pfalse)

Safe horizon · Corollary III.7

k ≤ ln(1 − δ) / ln(1 − pe · pc)

Variables · assumes step independence

pe
observation error rate
pc
gate false-accept rate
k
number of state updates (sessions)
δ
tolerated contamination probability

Below, q = pe · pc is the chance that a single update admits an erroneous observation, so Pcont(k) = 1 − (1 − q)k.

Worked example · illustrative values pe = 0.05, pc = 0.1, δ = 0.05

Per-update contamination
0.005
q = pe · pc = 0.05 · 0.1
Contamination after 100 updates
39.42%
Pcont(100) = 1 − (1 − 0.005)100
Safe horizon for δ = 0.05
k* = 10.23
ln(1 − 0.05) / ln(1 − 0.005): the first 10 updates stay within δ.
Values of Pcont(k)
Worked example: q = 0.005. Illustrative values, not measurements.
Updates kPcont(k)
10.50%
104.89%
10039.42%
1,00099.33%
10,000≈ 100%

Corollary III.7

  • A perfect gate (pc = 0) gives zero contamination.
  • For any pc > 0, contamination approaches 1 as k grows: the gate controls the rate, not the asymptote.

Conjecture III.8

schema repair suffices for zero leakage under three stated conditions, which the authors note do not generally hold.

Move the sliders or pick a gate. The chart plots Pcont(k) with faint curves for a gate ten times weaker and ten times stronger; the dashed line is δ and the marker is the horizon k*. Equations from Theorem III.4 and Corollary III.7 of the paper; every number here is computed from illustrative inputs, not measured.

Implements

  • §III-B Observe, corroborate, validate, Eq. (5)–(8)
  • §III-D Theorem III.4 and Corollary III.7
  • §IV-D Controlled corruption for Memory Harm
  • §V-A Memory contamination, Fig. 5a

Sections of arXiv:2609.32041, a specification rather than a validated implementation.

Harness metrics

Counterfactual metrics from the paper’s harness that apply to quality gate. The paper specifies them; it reports no measurements.

§ 03.3 Reconcile

Versions & amendments

In development

Part of Mnemos · codename Chronos

Knows which clause is in force today.

D1 valid [ta, tc) · D2 valid [tc, te)

Documents revise one another. Mnemos records, for every stored claim, the interval during which it is authoritative, and resolves each new conflict into one of five outcomes instead of overwriting: support, supersede, scope-coexist, retain conflict, or escalate for review.

The problem

An amendment supersedes a clause only after its effective date. A preliminary statement coexists with its audited version because the two carry different authority. Overwrite-on-conflict loses both distinctions, and the audit trail with them.

For: Contract, compliance and audit work where documents amend, restate and contradict each other.

How it works

  1. Keep four clocks

    Publication, event, effective and ingestion time are distinct axes; an amendment ingested today may take effect next quarter.

  2. Classify the conflict

    Each new claim is reconciled against the store into one of five outcomes, and the outcome itself is recorded.

  3. Close, don't delete

    Under supersede, the earlier claim's interval is closed at the effective time and kept for audit.

  4. Answer for a date

    Ask what holds today, or what held on the day a decision was taken, and get the clause that was valid then.

Temporal supersession · the five outcomes of 𝒮sup Interactive Illustrative
The five outcomes, as a table
OutcomeWhat happensExample from the demo stream
Support New evidence agrees with the stored claim and corroborates it. Named in the paper; description is our reading. Liability cap AMD-1, “unchanged”Corroborates §7.1 of the MSA.
Supersede The earlier claim's validity interval is closed at the effective time; the new evidence is authoritative from then on. Defined in the paper. Payment terms AMD-1, “45 days”Closes “30 days” on 2026-10-31; “45 days” is valid from 2026-11-01.
Scope-coexist Claims with different populations or authority both stand, e.g. a preliminary statement next to its audited version. Defined in the paper. Liability cap SOW-2, “€400,000 / year”Different scope from the MSA cap: both stand.
Retain conflict The contradiction cannot be resolved yet, so both claims are kept and marked as conflicting. Named in the paper; description is our reading. Express pallet rate INV-2026-0117, “€68.00”Contract rate since 2026-11-01 is €71.50; the line is €42.00 below it. Both claims are kept and flagged.
Escalate for review The conflict is handed to a person before the state changes. Named in the paper; description is our reading. Payment terms INV-2026-0117, “30 days”Contradicts the 45 days in force since 2026-11-01; routed to a person before anything is paid.
How to read it. Read the four fictional documents in order: each field meets the claims already stored and gets one outcome; the ledger is the resulting validity index. Outcome names come from the paper, which defines supersede and scope-coexist and only names support, retain conflict, escalate for review; for those, the descriptions are our reading. “New claim” is not one of the five: nothing was stored yet. Fictional documents. The demo runs entirely in your browser; it illustrates the ideas and is not production software.

Implements

  • §III-B Temporal supersession operator 𝒮sup
  • §III-C Temporal layer 𝒱t
  • §V-B Bitemporal supersession, Fig. 5b

Sections of arXiv:2609.32041, a specification rather than a validated implementation.

Harness metrics

Counterfactual metrics from the paper’s harness that apply to versions & amendments. The paper specifies them; it reports no measurements.

§ 03.4 Trace

Source tracing

In development

Part of Mnemos · codename Ariadne

Every answer, traced to the bounding box.

σ₂ ↝ σ₁ ↝ {e₁, e₂, e₃} ↝ {Di, ℓi, vi}

Mnemos keeps the thread from any conclusion, however consolidated, back to the document, region and version it came from. Provenance is not an extra step; it is an invariant that every operation preserves, through consolidation, supersession and repair.

The problem

Assistant "memory" features retain the fact but discard the evidence. They cannot trace a remembered liability cap back to the box on page four, so they cannot be used in an audit.

For: Audit, legal and regulated teams who must show where every number came from.

How it works

  1. Tag at write

    Every admitted observation carries its document, location and version into the ledger.

  2. Survive abstraction

    When evidence is consolidated into a schema, the schema keeps links to every observation it was built from.

  3. Survive change

    Supersession and forgetting repair the chain instead of cutting it.

  4. Cite on answer

    Answers point to the span, box or table cell they rely on, or abstain when nothing supports them.

Ariadne · Document inspector Interactive Illustrative
Every extracted field, as a table

Four fictional documents read in order. Each row is a field: the value read, where it was read from, and how it met what memory already held.

Provenance gives the document, the section or region type, and the page.
DocumentFieldValueProvenanceOutcome against memory
MSA-2026-014Contract · 2 Mar 2026 Parties Halden Freight SAS ↔ Orvelle Analytics SA MSA-2026-014 §1 · p. 1 New claim
Term 24 months from 2026-03-02 MSA-2026-014 §2 · p. 1 New claim
Payment terms 30 days MSA-2026-014 §4.2 · p. 1 New claim
Standard pallet rate €42.00 MSA-2026-014 §5 · p. 1 New claim
Express pallet rate €68.00 MSA-2026-014 §5 · p. 1 New claim
Storage rate €3.10 / pallet-week MSA-2026-014 §5 · p. 1 New claim
Liability cap €250,000 / year (all services) MSA-2026-014 §7.1 · p. 1 New claim
Governing law French law, courts of Rennes MSA-2026-014 §9 · p. 1 New claim
SOW-2Statement of work · 1 Jun 2026 Liability cap €400,000 / year (cold-chain services) SOW-2 §3 · p. 1 Scope-coexistDifferent scope from the MSA cap: both stand.
Cold-chain pallet rate €95.00 SOW-2 §4 · p. 1 New claim
AMD-1Amendment · 15 Oct 2026 Payment terms 45 days AMD-1 §1 · p. 1 SupersedeCloses “30 days” on 2026-10-31; “45 days” is valid from 2026-11-01.
Express pallet rate €71.50 AMD-1 §2 · p. 1 SupersedeCloses “€68.00” on 2026-10-31.
Liability cap unchanged (all services) AMD-1 §3 · p. 1 SupportCorroborates §7.1 of the MSA.
Data residency EU only AMD-1 §4 · p. 1 New claim
INV-2026-0117Invoice · 5 Nov 2026 Express pallet rate €68.00 INV-2026-0117 Table · p. 1 Retain conflictContract rate since 2026-11-01 is €71.50; the line is €42.00 below it. Both claims are kept and flagged.
Payment terms 30 days INV-2026-0117 Paragraph · p. 1 Escalate for reviewContradicts the 45 days in force since 2026-11-01; routed to a person before anything is paid.
Total €2,076.00 INV-2026-0117 Table · p. 1 New claim
How to read it. Dashed boxes are layout regions, drawn in the page's 600 × 848 coordinate box; highlighted ones yield fields. Each field keeps its provenance (document, section, page, box) and an outcome against what memory already held: one of the paper's five reconciliation outcomes, or a new claim. Fictional documents. The demo runs entirely in your browser; it illustrates the ideas and is not production software.

Implements

  • §III-A Provenance ledger Pt
  • §III-B Provenance invariant across operations
  • §V-B Recursive provenance
  • §IV-B Harness property 6: provenance queries

Sections of arXiv:2609.32041, a specification rather than a validated implementation.

§ 03.5 Forget

Verified deletion

In development

Part of Mnemos · codename Lethe

Forget on request, and prove it.

FF = 1 − Leak(Xdel | S′t) / Leak(Xdel | St)

Deleting the file is only Level 1. Mnemos runs the five-level hierarchy: physical removal, representation invalidation, consolidated schema repair, parametric unlearning and certified non-recoverability. It reports Forgetting Fidelity and Coverage Retention, so a deletion is measured rather than assumed.

The problem

Current Document AI stops at Levels 1 and 2. A schema consolidated from a deleted document still carries information about it, so an erasure request under GDPR Article 17 is not actually honoured.

For: Organisations with erasure, retention and data-residency obligations on what their systems remember.

How it works

  1. Remove the object

    Purge the raw file, page images and OCR cache.

  2. Invalidate representations

    Delete the embeddings, graph nodes and index entries derived from it.

  3. Repair schemas

    Re-consolidate every schema that depended on the target and repair its provenance links.

  4. Certify

    Bound the residual leakage with membership and attribute-inference probes, and issue a certificate.

Verifiable forgetting · Fig. 6 · Algorithm 2 Interactive Illustrative
  1. Level 1

    Physical object removal

    Delete the raw file.

    Implemented in current Document AI Evaluated by: Forgetting Fidelity · Coverage Retention

  2. Level 2

    Representation invalidation

    Remove embeddings and graph nodes derived from it.

    Implemented in current Document AI Evaluated by: Forgetting Fidelity · Coverage Retention

  3. Level 3

    Consolidated schema repair

    Re-consolidate schemas that depended on the deleted evidence and repair links.

    Unbuilt Evaluated by: Forgetting Fidelity · Coverage Retention

  4. Level 4

    Parametric unlearning

    Remove facts absorbed into model weights: exact retraining, influence-based approximation, certified residual checks.

    Unbuilt Evaluated by: Additional certification

  5. Level 5

    Certified non-recoverability

    Audit residual leakage cryptographically or statistically, then issue a certificate.

    Unbuilt Evaluated by: Additional certification

Passing one level does not imply success at the next. Current Document AI implements Levels 1–2; skipping Level 3 leaves non-zero leakage in the semantic store.

Levels, their status and how each is evaluated are Fig. 6 of the paper. The erasure run is our illustration on fictional documents; leakage is shown qualitatively (present or removed), nothing is measured. Fictional documents. The demo runs entirely in your browser; it illustrates the ideas and is not production software.

Implements

  • §III-D Forgetting Fidelity and Coverage Retention, Eq. (13)–(14)
  • §III-D Conjecture III.8, schema-repair sufficiency
  • §V-C Five-level hierarchy, Fig. 6 and Algorithm 2

Sections of arXiv:2609.32041, a specification rather than a validated implementation.

Harness metrics

Counterfactual metrics from the paper’s harness that apply to verified deletion. The paper specifies them; it reports no measurements.

§ 03.6 Measure

Evaluation on your documents

In development

Part of Mnemos · codename Anamnesis

Prove that your system remembers.

EG(k) = Q(Dk+1 | Sk) − Q(Dk+1 | Rk)

Leaderboards reward single-shot reading, so nobody builds memory. Mnemos evaluates persistence directly: the same model and the same future documents under three arms (warm state, retrieval-only and cold), over streams with session breaks, amendments, recurring families, deletion requests and unanswerable questions.

The problem

None of the ten benchmarks audited in the paper tests cross-session state evolution. Because benchmarks do not reward persistence, researchers do not build it, and the reverse.

For: Research and engineering teams who need to show that a memory system helps, and what it costs.

How it works

  1. Build the stream

    Eight mandatory properties, from sequential presentation and cross-document dependency to forgetting requirements and abstention.

  2. Run three arms

    Warm state, an append-only retrieval index over the same stream, and a cold start, under identical weights.

  3. Report five metrics

    Experience Gain, Cost Efficiency, Memory Harm, Forgetting Fidelity and Coverage Retention, always together.

  4. Model the dependence

    A mixed-effects fit with a log-exposure term, family-level random effects and AR(1) residuals, or a paired bootstrap for small samples.

Counterfactual harness · after Fig. 4 Interactive Hypothetical inputs

Hypothetical inputs. The paper specifies this harness; it reports no measurements.

Stream

  1. D1
  2. D2
  3. Dk

One document per session, before the test.

Skclean

Clean twin for MH

The warm state the same documents would give with no corrupted observations.

  1. Sk

    Warm state

    Carries consolidated state from documents D1 … Dk.

  2. Rk

    Retrieval-only

    An append-only index of the same documents, with no consolidation.

  3. S0 = ∅

    Cold baseline

    Starts from empty state on every task.

Every arm is tested on

Dk+1

The same next document; quality Q and cost C are compared.

Deletion audit for FF and CR

St

State before the deletion request

S′t

State after forgetting Xdel

Success criterion

Positive Experience Gain and Cost Efficiency, bounded Memory Harm, and high Forgetting Fidelity and Coverage Retention, under all eight mandatory properties.

The five metrics
MetricDefinitionMeasuresTarget
EG Experience Gain{"id"=>"eg", "abbr"=>"EG", "name"=>"Experience Gain", "equation"=>"EG(k) = Q(D_k+1 | S_k) − Q(D_k+1 | R_k)", "measures"=>"Benefit of consolidation beyond retrieval access.", "good"=>"positive"}Benefit of consolidation beyond retrieval access.Positive
CE Cost Efficiency{"id"=>"ce", "abbr"=>"CE", "name"=>"Cost Efficiency", "equation"=>"CE(k) = 1 − C(D_k+1 | S_k) / C(D_k+1 | S_0)", "measures"=>"Inference cost saved by warm state, in tokens or FLOPs.", "good"=>"positive"}Inference cost saved by warm state, in tokens or FLOPs.Positive
MH Memory Harm{"id"=>"mh", "abbr"=>"MH", "name"=>"Memory Harm", "equation"=>"MH(k) = Q(D_k+1 | S_k^clean) − Q(D_k+1 | S_k)", "measures"=>"Degradation induced by state; positive values mean memory hurt.", "good"=>"bounded"}Degradation induced by state; positive values mean memory hurt.Bounded
FF Forgetting Fidelity{"id"=>"ff", "abbr"=>"FF", "name"=>"Forgetting Fidelity", "equation"=>"FF = 1 − Leak(X_del | S′_t) / Leak(X_del | S_t)", "measures"=>"How much of a deleted target still leaks from the post-deletion state (probe-based mutual-information bound).", "good"=>"high"}How much of a deleted target still leaks from the post-deletion state (probe-based mutual-information bound).High
CR Coverage Retention{"id"=>"cr", "abbr"=>"CR", "name"=>"Coverage Retention", "equation"=>"CR = exp( Q(Y_ret | S′_t) − Q(Y_ret | S_t) )", "measures"=>"Whether knowledge that should survive a deletion still does.", "good"=>"high"}Whether knowledge that should survive a deletion still does.High

Memory Harm protocol. Run a clean stream and a corrupted one over the same documents; the corrupted stream replaces a fraction ρc of observations with plausible but wrong values, such as a misread tax ID or a transposed table cell.

Cumulative cost. CEcum(k) = 1 − Σ warm costs / Σ cold costs over sessions 1 … k+1. A memory system pays for itself when CEcum(k) > 0 for the target stream length, even if per-task CE is zero in some sessions.

Analysis. A mixed-effects model with a log-exposure term f(k) = log(1 + k), a warm–cold indicator, family-level random intercepts and slopes, and AR(1) residuals; a paired bootstrap over streams for small samples.

Pick a metric to see which arms it compares, then move the hypothetical inputs or try a scenario. Adapted from Fig. 4 of the paper, with the retrieval-only arm described in its text. Arms, metrics and the criterion are the paper's; the paper reports no results for them.
Required design · Eight mandatory properties Interactive

Untick what a benchmark lacks: the success criterion only holds under all eight.

  1. Sequential presentation Documents arrive across sessions, and state persists between them.
  2. Cross-document dependency Some questions need evidence seen only in earlier sessions.
  3. Evolving content Controlled revisions, amendments and supersessions.
  4. Consolidation opportunities Repeated instances of a document family.
  5. Forgetting requirements Deletion requests, checked for an unrecoverable target and intact unrelated knowledge.
  6. Provenance queries Conclusions must be traced back to their source regions.
  7. Session boundaries Context is cleared, so only the persistent store survives.
  8. Abstention Unanswerable queries the system must decline rather than answer from stale or contaminated state.
Property names from the paper; the one-line glosses are ours.

Implements

  • §IV-A Statefulness audit, Table IV
  • §IV-B Harness and its eight properties, Fig. 4
  • §IV-C Counterfactual metrics, Eq. (15)–(16)
  • §IV-D Statistical design, Eq. (17)–(18)

Sections of arXiv:2609.32041, a specification rather than a validated implementation.

Harness metrics

Counterfactual metrics from the paper’s harness that apply to evaluation on your documents. The paper specifies them; it reports no measurements.

§ 04 Research prototypes

Two papers you can try in the browser.

These demos are not products. Each illustrates the ideas of one published paper on fictional documents, entirely in your browser, and its ideas feed one of the products above.

Multilingual retrieval-augmented generation with strict grounding and explicit abstention, on French and English.

Used in: Ariadne

Research prototype · MTRACE · Cite, or abstain

Interactive Illustrative

With JavaScript on, this demo answers questions like these from the 12 passages below, in English or French and as of a chosen date, quoting and citing its evidence, or abstaining when no passage in force supports an answer.

  • What are the payment terms?
  • Quel est le délai de paiement ?
  • What is the liability cap for cold-chain services?
  • Where is client data stored?
  • What does an express pallet cost?
  • Who is Orvelle’s CEO?
The collection 12 passages · 5 documents
Section Passage Valid Status
MSA-2026-014 Master Services Agreement ·
§2 Term. Twenty-four (24) months from 2 March 2026. –
§4.2 Payment terms. Invoices are payable within thirty (30) days of receipt. – superseded by AMD-1 §1
§5 Rates: standard pallet €42.00; express pallet €68.00; storage €3.10 per pallet-week. – superseded by AMD-1 §2
§7.1 Liability. The Supplier’s aggregate liability is capped at €250,000 per contract year. –
§9 Governing law. This Agreement is governed by French law; courts of Rennes. –
SOW-2 Statement of Work No. 2 — Cold-chain services ·
§1 Scope. Temperature-controlled storage and transport at 2–8 °C for the Client’s EU sites. –
§3 For cold-chain services only, the Supplier’s liability is capped at €400,000 per contract year. –
AMD-1 Amendment No. 1 to MSA-2026-014 · · in force from
§1 Clause 4.2 is replaced: invoices are payable within forty-five (45) days of receipt. –
§2 In the rate table of Clause 5, the express pallet rate becomes €71.50. –
§4 New Clause 11. Client data is processed and stored in the European Union only. –
INV-2026-0117 Invoice INV-2026-0117 ·
Lines Invoice INV-2026-0117: 12 express pallets at €68.00 and 30 standard pallets at €42.00; total €2,076.00. –
DPA-2026-014 Data Processing Addendum ·
§3 Sub-processors. The Supplier may appoint sub-processors only with the Client’s prior written consent. –

Illustrative · BM25 + query expansion + reciprocal rank fusion over 12 fictional passages, in your browser. A research prototype, not a product.

How to read it. Ask in English or French, or pick a question, then switch the date to see superseded clauses become history. An answer is given only from a passage in force that covers the question’s key terms, quoted with its span highlighted and cited by document, section and page; otherwise the demo abstains and says what it searched. Semantic query expansion, reciprocal rank fusion and strict grounding with abstention come from MTRACE (arXiv 2025), validity intervals from the 2026 survey; a simple extractive rule stands in for the generator. Fictional documents. The demo runs entirely in your browser; it illustrates the ideas and is not production software.

Research prototype · arXiv 2026

Fresh at Decision, Stale at Action: Routing New Information to LLM Decisions Already Made

A. Merimi, S. Bakkali, A. Sakat

A timestamp theory of LLM decision validity and a decision ledger that routes new information to decisions already made.

Used in: Chronos

Research prototype · Decision ledger · Fresh at Decision Interactive Illustrative

Ledger entry

Open · revisable until tr =

Decision

Ship the December campaign by express pallet with Halden Freight

Chosen
Halden Freight, express
Alternatives
  • Halden Freight, standard
  • Second carrier, express

EvidenceMSA §5: express pallet €68.00 (verified 2026-10-01).

Dossier

  • Halden Freight SAS
  • Halden’s December express route runs through the port of Le Havre
  • Orvelle Analytics SA

Watch conditions

  • Halden Freight express rate or capacity changes
  • Disruption on the Le Havre route
  • Orvelle December volume forecast changes

Timestamps

The six timestamps of this decision
SymbolMeaningDate
tw Evidence written
tv Evidence verified
td Decision made
tr Revision deadline
ta Action
te Outcome observed

ta − tv = Δfresh + Δact = 4 + 57 = 61 days with Δfresh = td − tv and Δact = ta − td

News stream · 5 items

Routing signals: name mentions a name from the decision · dossier matches its dossier · watch matches a watch condition.

  1. n1

    Amendment No. 1: Halden Freight’s express pallet rate becomes €71.50 from 1 November.

    • name : yes
    • dossier : yes
    • watch : yes

    Cost rises by €3.50 per pallet; the decision can stand.

  2. n2 Decoy

    Halden Marine announces a fleet maintenance stop in December.

    • name : yes
    • dossier : no
    • watch : no

    A different company with a similar name.

  3. n3 Decisive

    Dock workers at the port of Le Havre call a strike from 25 November to 3 December.

    • name : no
    • dossier : yes
    • watch : yes

    The December express route is blocked: switch to the second carrier before 20 November.

  4. n4

    A strike at the port of Le Havre is planned for 15–20 January 2027.

    • name : no
    • dossier : yes
    • watch : yes
    • starts after the activity

    Starts after the activity ends: dropped by the time filter.

  5. n5

    Orvelle raises its December volume forecast by 10%.

    • name : yes
    • dossier : yes
    • watch : yes
    • after tr

    Arrives after the revision deadline: too late to change this decision.

Why a decision goes stale, and what routing buys

Proposition 1 · Freshness ceiling

Pr[f holds at ta] = e−λfΔfresh · e−λfΔact ≤ e−λfΔact

If the fact f behind the evidence changes at a constant rate λf, the chance it still holds at action time is capped by the gap between decision and action, however fresh the evidence was when the decision was made.

Accuracy factorisation

Hypothetical inputs

A = (1 − π)K + π[ρS + (1 − ρ)L]

Accuracy splits into routing and judging: π is the share of decisions that should change, ρ the routing recall, S the rate of switching when the news is delivered, L the rate of switching when it is not, and K the rate of keeping a decision that should stand.

How to read. Pick a router: it decides which news reaches the open decision before its revision deadline tr, and the outcome is judged when the decision is carried out at ta. The demo assumes the agent switches whenever decisive news reaches it in time. Timestamps, routing rules, Proposition 1 and the factorisation follow Fresh at Decision, Stale at Action (A. Merimi, S. Bakkali, A. Sakat, arXiv 2026), which also studies routing all items, BM25 and dense retrieval over the decision text, the dossier or the watch conditions, dependency queries and oracle routers. Fictional documents. The demo runs entirely in your browser; it illustrates the ideas and is not production software.

§ 05 The research behind it

Earlier work, and what it taught us.

Beyond the survey itself, these publications shaped the products: what each one is, the lesson we took from it, and where that lesson is used.

  • WACV 2025

    GlobalDoc

    A cross-modal transformer for real-world document image retrieval and few-shot classification.

    What it taught us: A document's identity lives in the joint vision-language representation, not in either modality alone.

    Used in: Mnemos

  • Pattern Recognition 2023

    VLCDoC

    Contrastive vision-language pre-training for cross-modal document classification.

    What it taught us: Aligning modalities, rather than merging them, keeps what each one carries.

    Used in: Mnemos

  • ICDAR 2024

    Anytime early exits

    The first multimodal early-exit design for document understanding: over 20% lower latency at full baseline accuracy.

    What it taught us: Compute can be spent where a document needs it, which is what a memory-warmed parser should do.

    Used in: Mnemos

  • arXiv 2025

    MTRACE

    Multilingual retrieval-augmented generation with strict grounding and explicit abstention, on French and English.

    What it taught us: An answer that cannot cite its evidence should not be given.

    Used in: Ariadne

  • arXiv 2026

    Fresh at Decision

    A timestamp theory of LLM decision validity and a decision ledger that routes new information to decisions already made.

    What it taught us: Being right when you decide is not being right when you act; validity needs a clock.

    Used in: Chronos

  • ICDAR 2025

    WildKhmerST

    29,601 annotated Khmer text lines from 10,000 real-world images, with a detection and recognition benchmark.

    What it taught us: Most document corpora are English and Latin-script; reading has to work beyond them.

    Used in: Mnemos

§ 07 Pilots

Shape the products with your documents.

The fastest way to know whether the chain helps your team is a pilot on your own documents. For the memory layer, we would measure it with the counterfactual metrics from the survey: does warm state beat retrieval alone, at what cost, and with what risk.

SynaptYon SARL

Tell us about your documents

What they are, the decisions they drive, and how often they change. Write to sbakkali.sirajai@gmail.com.