Skip to content

About

Project SLSSM-01 public research dossier - observable spectral dynamics in a recurrent sequence model

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Trinity Research

Live site: trinity-research.com

Source for the Trinity Research site. Trinity Research is Freddie Hunt and Joshua Hunt, PhD: language model research and optimization, business process automation, predictive modeling, and data ingestion systems.

Pages

Page What it is
index.html Capabilities, selected work, the signals strip, and the team.
slssm-01.html The Project SLSSM-01 research dossier — evidence, synthetic visualizations, and the disclosure boundary.
research-note.html The long-form research note: metric definitions, skeptic FAQ, limitations, citation blocks.
signals.html Generated. Business automation and applied AI, updated on weekdays.
hnted/ HNTED hunting dashboard: sun/moon timing, both daily lunar windows, and regional whitetail rut context.
hnted/field-notes/ Four short, source-linked hunting research articles.
hnted/waterfowl/ Free waterfowl guide: upstream forecasts, dated snow/cold/river observations, and sourced bird reports.

HNTED

HNTED is a static, browser-based field-planning project by Trinity Research. The deer dashboard's weather integration remains deferred. The waterfowl guide connects saved NWS/ECCC forecasts, NOAA snow analysis and daily station history, USGS river stages, and dated agency reports. Camera import, accounts, and automatic article generation are not enabled. Sun/moon calculations and geographic rut matching run on the visitor's device. Saved spots and preferences stay in that browser.

Town search uses a bundled GeoNames US town index under CC BY 4.0, with source and license attribution in the app. It has no geocoding API calls, API key, ZIP-code lookup, or property-address search. GPS and manual coordinates are available for locations outside that index.

Edit data/hnted-articles.json and rebuild reviewed articles with node scripts/build-hnted-notes.mjs --public. Publication settings live in data/hnted-publication.json. Add new article URLs to data/hnted-pages.json; node scripts/render-signals.mjs preserves these entries in the root sitemap. Hunting articles are separate from the business-automation Signals feed. The HNTED app requires no server route; only its static files are published. Its astronomy library license is retained in hnted/vendor/SUNCALC-LICENSE.txt.

Waterfowl data can be refreshed with node scripts/refresh-hnted-waterfowl.mjs. The workflow in .github/workflows/daily-waterfowl.yml is scheduled at 4:17 AM America/Chicago, with hourly backup checks at minute 41 UTC and a manual run available in GitHub Actions. Healthy backup checks skip provider collection. Its source reference is retained in docs/daily-waterfowl-workflow.yml. It refreshes weather and snow/cold/river snapshots, checks all 18 configured agency report feeds, and imports six numeric survey series after strict date and table validation: Loess Bluffs, Iowa, both INHS river inventories, Missouri, and Tennessee Black Bayou. Other sources receive newer-report notices or unavailable-check warnings. Hourly backup checks trigger another source sweep after six hours. Forecast collection has its own three-hour target and refreshes the other public sources in that same bounded run; GitHub's schedule is approximate. The job validates its data, commits only generated public records, and explicitly requests a Pages build, waits for it, and verifies current live data. Normal pushes retry transient errors without force pushes or overwriting branch changes. Pages polling checks all recent matching builds, so a failed sibling does not override a successful or in-progress build of the same commit. Persistent build failure permits at most two additional rebuild requests within the bounded wait; actual failures and stale live data still fail the workflow. Narrative summaries and unsupported numeric formats still require review; the guide links newer reports without relabeling old observations. Daily checks do not create daily surveys. Source-check and survey dates stay separate, with older data retained when a refresh fails and partial coverage labeled when permitted. No external API keys or local Windows routine are needed for this data job.

The Weather fronts layer decodes NOAA WPC's public coded surface analysis and short-range forecast bulletins into hnted/waterfowl/data/fronts.json. It shows front/boundary lines and H/L pressure centers at explicit valid times, using the same map projection and zoom. The selected spot's calendar day filters the available positions; no missing times or local arrival estimates are invented. The cloud routine collects fronts alongside weather. Failed sources retain their saved dates; analyses older than 12 hours are not drawn. The pure decoder is scripts/wpc-fronts.mjs; the collector is scripts/update-hnted-waterfowl-fronts.mjs. The coding guide is NOAA's bulletin documentation. An isolated malformed analysis boundary can be omitted with an explicit partial-geometry notice. Missing coordinates are never guessed; broader errors retain the previous source frames. Forecast parsing remains strict.

Run node scripts/check-hnted-waterfowl-health.mjs --require-fresh --live to check saved timestamps, actual forecast coverage, river observation times, agency-source monitoring, failed jobs, and the published data. Numeric PDF imports require Poppler's pdftotext; the workflow installs it from Ubuntu's package repository when needed. No PDF-reader executable is committed. GitHub schedules can be delayed or dropped; the hourly backup uses the same scheduler. On November 5 the missed daily update was recovered manually, and a transient GitHub push rejection cleared on retry. A successful manual run does not prove scheduled triggering. Local development snapshots update separately. The page checks its published completion marker every 15 minutes while visible and when returning after at least five minutes. Check for updates forces a check. New files update in place, preserving the selected day, city, species, filters, and map/list view. Failed or inconsistent downloads retain the prior view and show a warning. These requests fetch shared static files and do not send visitor coordinates. Daily cold-history cards distinguish NOAA archive observation dates from source-check times; those records can lag even when today's forecasts and collection job are current.

slssm-01.html, signals.html, feed.xml, and sitemap.xml are the only pages with generated content; everything else is hand-written.

Project SLSSM-01

The dossier covers an independent research effort studying recurrent language models with observable spectral dynamics. Its primary sequence mechanism uses state evolution rather than quadratic self-attention, which is what makes its internal memory behavior measurable in the first place.

Every number in the dossier is author-reported, rounded, from one run, and externally unreplicated. The site says this before any reader has to ask. Model weights and checkpoints, training code and recipes, exact architecture equations, raw activations, and raw mode tables are deliberately excluded.

The measurement method has a standalone, reproducible demonstration on public Pythia-14m training checkpoints: a pre-specified protocol with dated post-run amendments, pinned model commits, shuffled-pairing and persistence controls, and CPU-only runtime of a few seconds.

github.com/trinity-research/pythia-edmd-demo

The demo validates the measurement method on a public model. It does not validate this model's claims; the dossier is explicit about that distinction.

The signals pipeline

scripts/fetch-signals.mjs   RSS feeds  ->  data/inbox.json     deterministic
/daily-signals skill        inbox      ->  data/signals.json   Claude writes prose only
                                       ->  data/skipped.json   rejects, so they stay rejected
scripts/render-signals.mjs  signals    ->  signals.html, feed.xml, sitemap.xml,
                                           and the strip in index.html

The split is the point: Claude writes takes into JSON and edits nothing else, so a weak generation costs a weak entry rather than a broken page. The render script validates every entry and fails rather than publishing a malformed one.

Every inbox item must end up in either signals.json or skipped.json. Anything in neither is fetched again tomorrow and re-read every day until it ages out.

Scheduling

This runs locally, not in CI. A working GitHub Actions version is kept at docs/daily-signals-workflow.yml as a reference copy — deliberately not under .github/workflows/ because it needs the Claude GitHub App (not installed). This separate Signals workflow has not been enabled by the HNTED data publication. Activation steps are in the file's header.

.\scripts\run-signals.ps1 -DryRun      # everything except commit and push
.\scripts\run-signals.ps1              # the real thing
.\scripts\register-signals-task.ps1 -At 03:30   # weekdays, as you, while logged on
.\scripts\register-signals-task.ps1 -Remove

Logs land in %LOCALAPPDATA%\TrinityResearch\signals-logs — outside the repo, so they are never committed or served publicly. They are pruned after 30 days.

The runner refuses to start on a dirty working tree, so nothing of yours gets swept into an automated commit. It also reverts inbox.json and exits cleanly if Claude judged every item off-topic.

Two things the schedule depends on, both tied to your user session:

  • The gh token lives in the Windows keyring, protected against your profile. The task is therefore registered as interactive — as you, while logged on. A task set to "run whether user is logged on or not" starts without your profile loaded and usually cannot read it, so the push fails. To run while logged off, switch the remote to an SSH key first.
  • A powered-off machine does not run. -WakeToRun wakes a sleeping one; StartWhenAvailable catches a missed run at next logon.

Check on it with Get-ScheduledTaskInfo -TaskName TrinityResearch-Signals (LastTaskResult of 0 is success).

Tuning the feed

Edit scripts/sources.json — no code change needed:

  • priority — 1 is the core niche and always gets slots first. Default 3.
  • maxPerSource — stops one prolific blog taking the whole day.
  • maxPerPriority — ceiling per band. This is what keeps a quiet Power Platform week from turning the page into a general AI news feed: on those days it publishes three on-topic items instead of six off-topic ones.

After editing, run node scripts/check-sources.mjs to confirm every feed still resolves and parses. Feeds rot; this tells you which one.

Running the stages by hand

node scripts/fetch-signals.mjs     # writes data/inbox.json
claude "/daily-signals"            # writes data/signals.json + data/skipped.json
node scripts/render-signals.mjs    # regenerates the pages

Repository notes

Pure static site, no build step and no dependencies: clone and open index.html to preview locally. The scripts use Node built-ins only, so there is no package.json, no lockfile, and nothing to install. Deployed by GitHub Pages from main at the repository root; CNAME points at trinity-research.com.

social-card.png is a raster export of social-card.svg at 1200×630. If you edit the SVG, re-export the PNG — the OG tags declare those exact dimensions.

Maintained by Freddie Hunt and Joshua Hunt, PhD. Contact: contact@trinity-research.com

About

Project SLSSM-01 public research dossier - observable spectral dynamics in a recurrent sequence model

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages