Brief the crew

The pre-job tailgate talk is being built. Until it lands, describe the job in Search the past to see what went wrong on jobs like it and what BSEE recommends.

Report it

Near-miss capture by voice is being built. Until it lands, type what happened in Search the past: it checks what the report is missing and how to fix it.

Search the past

Paste a cause statement or describe a planned job. Tail Gate suggests the cause, finds the closest past offshore incidents and the BSEE alert that applies. A person decides.

One or two sentences on what failed and why. Leave out people's and companies' names; the page redacts the ones it recognises, but you are the first check.

Press Ctrl or Cmd with Enter to read. The local model takes a few seconds.

Or try a real report from the dataset:

Your corrections

Nothing read yet

Paste a cause statement or pick a real report on the left. You will see, in this order:

  1. the suggested cause, how sure the reading is, and what the report is missing;
  2. the past offshore incidents that read the same way, shared words highlighted;
  3. the BSEE safety alert that applies, with its recommendation;
  4. a place to record the action a reviewer chose.

New reports, sorted by what they need

An agent watches BSEE's public investigation index, files each newly published report here, and reads every cause statement. You start with the lane that needs you most.

    Pick a report

    Each report shows the investigator's own cause next to Tail Gate's reading, how often the mechanism has happened before, and what the report is missing.

    What keeps going wrong

    Mechanisms that recur across separate incidents.

      Cause statements by year

      Modern era only, where investigators use the six categories. Counts are cause statements, so one incident can add several.

      How Tail Gate works, and how far to trust it

      How the numbers were checked

      Organisers' protocol: 374 cause rows from 159 incidents in the modern era, the shipped folds, out-of-fold predictions, macro-F1 with a 95% interval from 1,000 incident bootstraps.

      Published baseline 0.349 (0.293 to 0.399). Tail Gate's similarity model 0.440 (0.383 to 0.485), a paired gain of +0.091 (+0.024 to +0.162). The local-model agent 0.471 (0.413 to 0.529). An independent evaluator re-scored both from saved predictions and found no leak. Not shown: any difference between our two models.

      Human labels: two teammates labelled 38 held-out rows by a rule written first, and marked 5 as too incomplete to label. On the other 33, the baseline agreed with them on 20, the similarity model on 16 and the agent on 18. With so few rows the intervals run from about 0.33 to 0.75, too wide to rank the three. The two label columns were filled together, so agreement between raters was not measured.

      Filling the gaps: for 4 of those 5 incomplete rows, the missing words were in the same report's other rows. The page rebuilt all 4 using only that report's words, each sentence citing its source rows. The fifth stays marked missing. A word check proves nothing was invented, not that the rebuild is right; a person accepts each one.

      What did not work: the agent's learn-from-mistakes loop (0.471 with it, 0.503 without). A planted label word was found at 0.80, short of our 0.9 bar. Categories with 26 to 28 rows are mostly noise.

      How it works

      Redaction
      Anything that looks like a person's or operator's name is removed before text reaches a model. A ledger check refuses a prompt that still carries one.
      Past incidents
      The five most similar documented incidents from the same era, and BSEE's alerts for that kind of event. Alerts that describe the challenge's own incidents are kept out.
      Cause reading
      A similarity model: nothing trained, instant. The local language model is an optional second opinion.
      Checking the answer
      Only one of the six cause names is accepted. Anything else falls back to the vote of the past incidents, and the fallback is shown.
      Your corrections
      They travel with you inside each request, signed; a tampered state is refused. Reports you choose to share are kept, names redacted, in a chained ledger anyone can inspect.

      Your data

      Without an account
      Searches, briefs and readings are worked out and returned; nothing you type is kept. Actions you plan stay in this browser.
      With an account
      Your organisation's reports, actions and talks are stored for your organisation only. No other organisation, and no Tail Gate reader, sees them. Every change is written to an audit trail where each entry seals the one before, so an edit or deletion shows.
      Sign-in
      An emailed link, no password. Tail Gate stores only a fingerprint of each link and session, never the link itself. Links work once, for 15 minutes.
      Analytics
      Counts of what people do (a search run, a talk given, an action closed), with people and organisations replaced by one-way codes. No names, email addresses or report text. Bots and our own checks are counted separately and left out of the numbers.
      The shared pool
      Only reports you choose to share, with names removed on arrival. A shared report teaches the reader for everyone only after a reviewer accepts it.

      Who it is for

      Field crews and supervisors
      The matches and the alert before the lift, not after.
      HSE and operations
      Pattern review, investigations and corrective actions, and sharing reports with the pool.
      Executives and risk leaders
      Recurring hazards across the fleet: the file upload reads a quarter's reports at once.

      Cause suggestions are model outputs for a person to review, never verified ground truth and never an operational clearance. Source: Bureau of Safety and Environmental Enforcement, US Department of the Interior. Cause element names: Energy Institute. Tail Gate began as InciRep, built on 2026-11-07 at the HAI Hackathon by Team 1, MIT licence.