VisaPredict AIVisaPredictAI
← VisaPredict AIPlan and progress

Verifiable MLOps plan status: progress, epics, user stories, execution route and recent evidence.

The plan, without hand-waving

A living dashboard for the MLOps v2 plan. Progress is calculated from user stories; paused work is neither hidden nor used to inflate completion.

Active plan

26 / 61 US

25 done · 1 under observation

Current phaseD · MLOpsD7 observation 0/2
NextD9 → D8architecture and documentation
Productionfresh2026-09-158ec972c234
UpdatedSep 04, 2026data main · 80b3bfb

From platform to evidence

We close the platform first, then reduce debt with measured payoff, synchronize all six surfaces, and only then train cohorts.

0+A+BFoundation
DMLOps platform
CClean code
FFull synchronization
ECohorts and models
GAcademic closeout

Now: D7 accumulates two real rebuilds in parallel; the next product change is D9. D5 waits for the causal F2 campaign.

Progress by epic

Every bar uses the same weighting: done 100%, observing 75%, active 50%, and all other states 0%.

0Done

Prepare main without losing work

Isolated baseline, backports, minimum CI and cleanup with proven recovery.

100%7/7 done
ADone

Bulletin ingestion

Safe manual ingestion and a cron resilient to the official source being blocked.

86%6/7 done
BDone

Green CI and minimum security

Main protection, cron authentication, supply chain and web security.

100%7/7 done
DIn progress

MLOps platform that earns its keep

Release integrity, observability and reproducible operations before training more models.

64%5/9 done
CPlanned

Clean code with a measured payoff

Reduce duplication and complexity only where value is demonstrated.

0%0/10 done
FPlanned

Full synchronization

One canonical truth across data, paper, web, RAG, model card and production.

0%0/8 done
EPlanned

Cohorts, DeepAR and router

Answer the stability hypothesis with preregistered evaluation and positive or negative results.

0%0/7 done
GPlanned

Academic closeout

Turn the platform and its results into a reproducible thesis, paper and defense.

0%0/6 done

Plan by epic and story

Open an epic to see the expected outcome and available evidence for every story.

0Prepare main without losing work7/7 · 100%Done
  1. 0.1
    Integration worktree

    Keep main isolated from the preserved R9 work.

    Done
  2. 0.2
    Re-audit the ER diagram

    Port the diagram and generator with visual equivalence.

    Done
  3. 0.3
    LaTeX backports

    Integrate and compile the documents without losing branches.

    Done
  4. 0.4
    Minimum viable CI

    Keep a strict ci-gate and linear history.

    Done
  5. 0.5
    Preserve PR #4

    Keep R9 open and in draft pending an independent decision.

    Done
  6. 0.6
    Verifiable cleanup

    Clean only accredited targets and retain recovery receipts.

    Done
  7. 0.7
    State documentation

    Record only completed facts and verified SHAs.

    Done
ABulletin ingestion6/7 · 86%Done
  1. A1
    Single network port

    Inject the fetcher, apply one retry and type WAF blocking.

    Done
  2. A2
    Validated manual ingestion

    Validate, promote atomically and upload to S3 in create-only mode.

    Done
  3. A3
    Resilient cron

    Separate source access from gates, CI and publishing.

    Done
  4. A4
    Honest signal

    Version source state and maintain one canonical issue.

    Done
  5. A5
    Offline tests

    Cover fetch, freeze and ingestion without depending on the Internet.

    Done
  6. A6
    BrowserFetcher

    Optional time-boxed spike; the manual ritual remains the reliable path.

    Deferred
  7. A7
    August–September cut

    Publish 300 snapshots, 298 months and the 2026-09 cut under rule zero.

    release/2026-09-158ec972c234
    Done
BGreen CI and minimum security7/7 · 100%Done
  1. B0
    Protected publishing

    Use a least-privilege GitHub App to publish through gates.

    Done
  2. B1
    Minimum supply chain

    Retain locks, contracts, triage and SHA-pinned actions.

    Done
  3. B2
    Lock contracts

    Verify toolchain and locks inside live jobs.

    Done
  4. B3
    Strict ci-gate

    Require policy, consistency, lint and modeling success.

    Done
  5. B4
    Web security

    Reduce 17 production advisories to zero and document triage.

    Done
  6. B5
    Source watchdog

    Distinguish blocked source from broken pipeline without losing alerts.

    Done
  7. B6
    Baseline measurement

    Measure the real CI critical path on a published SHA.

    Done
DMLOps platform that earns its keep5/9 · 64%In progress
  1. D1
    DVC friction

    Block pushes when the lock does not represent the DAG.

    f67edf9
    Done
  2. D2
    Visible advisories

    Reject empty artifacts and turn warnings and advisories into contracts.

    0092af0
    Done
  3. D3
    Ingestion state

    Maintain a closed, atomic and fail-closed state feed.

    004e478
    Done
  4. D4
    One release_id

    Break the card–manifest cycle with normalized identity.

    f9c22a2
    Done
  5. D5
    Derived provisional caveat

    Remove it only after the causal F2 campaign; never by hand.

    Deferred
  6. D6
    Disk hygiene

    Inventory, preserve and clean with conservative GC and receipts.

    Done
  7. D7
    Monthly tracking

    Observe two real rebuilds before deciding whether track_run stays.

    80b3bfb · 0/2
    Under observation
  8. D8
    Documentation consolidation

    Create ENGINEERING.md with a source matrix and backlinks, without deleting yet.

    Planned
  9. D9
    MLOps architecture

    Publish a one-page system overview and SVG using canonical figures.

    Planned
CClean code with a measured payoff0/10 · 0%Planned
  1. C1
    Single taxonomy

    Centralize categories and metadata with a golden master.

    Planned
  2. C2
    Shared extraction

    Remove scraper duplication with pure functions.

    Planned
  3. C3
    Re-runnable mega-audit

    Replace mutable globals with a testable AuditReport.

    Planned
  4. C4
    Specific errors

    Remove silent failures and log country and month for every skip.

    Planned
  5. C5
    Figure kit

    Extract shared theme, language and saving from three generators.

    Planned
  6. C6
    Modular database build

    Separate migrations, loading and governance while preserving the fingerprint.

    Planned
  7. C7
    Dead code

    Remove consumerless paths with anti-resurrection guards.

    Planned
  8. C7b
    Tree-dirty semantics

    Use one verifiable definition of a dirty tree.

    Planned
  9. C8
    Honest tooling

    Measure coverage and complexity across the real product.

    Planned
  10. C9
    LOC by role

    Prevent tooling from again exceeding half the product.

    Planned
FFull synchronization0/8 · 0%Planned
  1. F1
    Derived model count

    Remove the web literal and read key_facts fail-closed.

    Planned
  2. F2
    Source banner

    Show source blocking without promising automatic updates.

    Planned
  3. F3
    Generalized guardian

    Guard macros, tables, proposal and documentation.

    Planned
  4. F4
    Canonical RAG

    Answer figures from JSON and require a coherent production pin.

    Planned
  5. F5
    Derived paper

    Move literals and caveats into versioned facts.

    Planned
  6. F6
    Gallery readiness

    Add the cohort hook without duplicating data.

    Planned
  7. F7
    Stale documents

    Correct stale counts and claims.

    Planned
  8. F8
    Propagation runbook

    Mechanize the data → web → production order.

    Planned
ECohorts, DeepAR and router0/7 · 0%Planned
  1. E0
    Single scale

    Centralize the pre-training naïve scale.

    Planned
  2. E1
    Causal cohorts

    Build stability without using hold-out information.

    Planned
  3. E2
    Exploratory scan

    Measure already-scored results by cohort without retraining.

    Planned
  4. E3
    Global models

    Train a registered ladder and document failure too.

    Planned
  5. E4
    Stability router

    Compete against each cohort's naïve baseline under the canonical gate.

    Planned
  6. E5
    Scientific propagation

    Carry results into thesis, web, RAG and model card under rule zero.

    Planned
  7. E6
    Paid-for cleanup

    Unify universes and correct stale catalog entries and docstrings.

    Planned
GAcademic closeout0/6 · 0%Planned
  1. G1
    Overleaf

    Update and visually review both documents.

    Planned
  2. G2
    Threats to validity

    Document Cloudflare and semi-automatic ingestion.

    Planned
  3. G3
    Cohort result

    Publish the subsection even if the result is negative.

    Planned
  4. G4
    MICAI paper

    Propagate the cut and cohorts with derived caveats.

    Planned
  5. G5
    Defense deck

    Generate presentation figures from key_facts.

    Planned
  6. G6
    Next review

    Bring cohorts, multi-horizon results and open decisions to Chente.

    Planned

Latest progress

This log changes only when verifiable evidence exists: a commit, CI, release or receipt.

  1. D7 reached main

    Under observation

    Monthly tracking merged with 5/5 CI; observation over two real rebuilds begins (0/2).

  2. D4 closed the identity cycle

    Done

    The single release_id contract is on main and will activate with the next real bulletin.

  3. D2 complete

    Done

    Empty artifacts, advisories and warnings are now covered by visible contracts and green CI.

  4. D1 and B4 complete

    Done

    The DVC hook blocks stale locks and the web audit records zero vulnerabilities.

  5. Cleanup 100/100

    Done

    Repositories, platform, evidence and recovery were verified without losing work.

  6. 2026-09 cut published

    Done

    Production became fresh with 300 bulletins and an immutable release.