Read the evidence behind the preview.

ICONIP Paper Method: Public Overview

Document ID: iconip-method-v1 Version: 1.0.0 Evidence class: Accepted-paper overview Language: English

Purpose and scope

The accepted ICONIP paper presents a site-specific way to organise corrosion-related evidence for rock-bolt support screening. It treats distinct material and support responses as separate prediction tasks, then makes their contributions visible in a two-branch engineering screen. This structure is intended to preserve attribution: readers can see whether the static-capacity view or the deformation-related view is more restrictive in a displayed case.

Environmental and geological context is represented through a predefined site-specific input. This input is prepared for the study workflow; it is not inferred by the Copilot and is not an unrestricted control that an operator can adjust until a desired flag disappears. The paper's results and their evaluation belong to the study cohort and stated protocol. They should not be read as evidence that the same behaviour has been independently confirmed at other mines or operating conditions.

From evidence to a screening result

The three responses concern section loss, ultimate-strength reduction, and elongation loss. The static view brings the section and strength responses together; the ductility view keeps the elongation response visible separately. A support can therefore retain a relatively high static reserve while showing a more restricted elongation reserve. Elongation is a material response indicator, not a direct measurement of the installed support system's energy absorption.

Which ML / AI models does LAVA use?

The accepted paper names the learners and their jobs. XGBoost estimates the percentage-mass-loss section proxy and the control-normalised total-elongation loss. Gaussian-process regression estimates coupon ultimate-strength loss relative to its matched supplier/type control. The serviceability index then takes the more restrictive of the static-loss sum and elongation loss: I_svc = max(ΔA + Δσ_U, ΔE_tot). These are the paper's selected component learners; they are not three interchangeable models for one target.

The public benchmark video also shows direct-index comparison models: linear regression, Gaussian process, and XGBoost. Its artificial neural network (ANN) row is a supplemental fixed-configuration comparison, not a LAVA component learner. The composed LAVA row uses the three component predictions described above. Direct-comparison scores therefore answer a different question from the component scores.

The disclosed chemistry input combines chloride with an electrical-conductivity-by-copper interaction; a bounded, pre-assigned Rock Reactivity Index adjusts the resulting score before it is fixed as a downstream model input. pH may appear as context, but it is not an explicit term in the disclosed final score. The public explanation names these inputs and their roles without listing the controlled calibration constants.

At a high level, the method takes recorded input characteristics, evaluates separate response components, and combines them under the paper's screening rules. Keeping the components separate makes the governing pathway inspectable. The resulting values support comparison and review; they are not probabilities of failure and do not by themselves prescribe an intervention.

The paper reports an internal evaluation using the study data. Model selection and evaluation are part of the same research programme, so the reported internal result is not an independent external validation. The public showcase preserves this distinction: stored paper-model outputs may be used to demonstrate the interface, but the Copilot itself is not a component of the paper's model evaluation.

What this overview intentionally omits

The learner names, target roles, and high-level recombination above are reported in the accepted paper and public benchmark video. This overview does not reproduce detailed training settings, feature encodings, fitted artefacts, controlled calibration constants, or an implementation recipe. Consult the accepted paper for its full scientific account where access and use are permitted.

Limitations

The method is scoped to its study data, assumptions, inputs, and screening purpose. No claim of cross-site transfer, prospective performance, calibrated uncertainty, or operational approval follows from the accepted paper alone. Showcase trajectories, synthetic records, maps, and assistant responses are separate extensions and must keep their own evidence labels.

Public retrieval terms: ICONIP method; XGBoost section-loss proxy; XGBoost elongation loss; Gaussian-process ultimate-strength loss; direct-index benchmark; supplemental ANN comparison; chloride; copper; electrical conductivity; Rock Reactivity Index; pH context; serviceability index; response branches; internal evaluation; external validation; controlled calibration constants; training settings undisclosed.

LAVA Project and Copilot Overview

Document ID: lava-overview-v1 Version: 1.0.0 Evidence class: Public showcase overview Language: English

What LAVA is

LAVA is a site-specific workflow for screening corrosion-related changes in rock-bolt support. It separates several engineering response pathways so that a reviewer can see which part of the assessment is driving attention. The purpose is to organise evidence and make a review easier to follow. LAVA is not a universal corrosion predictor, an autonomous safety system, or a replacement for an engineer's site assessment.

The accepted research method and the showcase interface are related but distinct. The research method evaluates selected response components and applies documented screening rules. The website adds interactive views and a Copilot assistant to explain selected records, compare branches, and locate supporting material. Interface features do not become validated research results merely because they appear beside paper results.

What is LAVA Copilot and how does it work?

An invited visitor chooses a fictional showcase region and asks a question. The assistant can inspect approved scenario records, read stored model-result values, search the public document collection, and open the full original source for a citation. It may ask for clarification when the question lacks evidence or refers to information that the showcase does not contain. It returns a review-oriented explanation and possible next checks; it cannot modify stored results, approve an asset, or issue a maintenance order.

The displayed screening values are supplied independently of the assistant's prose. A positive flag remains visible even when an explanation discusses it. Answers identify whether they came from a configured live model or from an evidence-only path; a fallback is not presented as a live model answer. Sources are limited to an allowlist, and a citation should lead back to the readable source rather than an unsupported claim.

Limits

The public collection contains showcase material and synthetic examples, not private site records or an operational mine database. Missing measurements, installation dates, geological assessments, or company rules remain missing. The demonstration can explain its own inputs and boundaries, but it cannot establish field performance, asset safety, failure probability, remaining life, or production readiness.

Public retrieval terms: LAVA overview; Copilot workflow; public corpus; source citation; human review; synthetic showcase; model results; evidence boundaries; no autonomous safety approval.

Fictional Demonstration Review Policy

Document ID: review-policy Version: 1.0.0 Evidence class: Fictional review-workflow fixture Status: Demonstration guidance only; not an approved company policy or site instruction. Language: English

Review routing

When a displayed branch triggers under the showcase screening rules, keep the flag visible and route the example to human review. A language-model explanation cannot remove, downgrade, or overrule the flag. A passing branch does not cancel a trigger in another branch. The thresholds are demonstration screens and do not define an operating acceptance limit.

The showcase begins with a fictional region. Supports shown in one region share one installation batch as a presentation assumption. The display does not identify an individual real support or provide its calendar installation date. Do not turn this assumption into a factual asset record.

Ask for evidence; do not fill gaps

If a question depends on whether conditions changed, request dated water-quality records for the same support or site. Ask for the relevant engineer-reviewed geological assessment and RRI assignment record, together with original material and test records where they are relevant. A single synthetic endpoint cannot establish a historical change. If the required record is missing, say so plainly and ask a focused follow-up rather than inventing a value, date, trend, or site policy.

The Copilot can identify evidence gaps, retrieve approved public sources, and draft a concise review summary with possible next checks. It does not write to an asset system, issue an inspection or replacement order, modify stored predictions, or approve safety. The responsible engineer and the organisation's authorised process retain those decisions.

Review outcome

Record which branch triggered, which supplied evidence was checked, what remains unavailable, and what human review is proposed. Distinguish stored model outputs from observed records and assistant-generated prose. Keep the source citation attached to any statement drawn from the public document collection.

Limitations

This is a fictional demonstration fixture, not a company procedure, regulatory interpretation, engineering instruction, or assurance case. It cannot establish asset safety, remaining life, a failure date, replacement timing, or policy compliance. Real decisions require the applicable site evidence, standards, and authorised engineering judgement.

Public retrieval terms: fictional review policy; human review workflow; flags cannot be overridden; evidence-gap questions; dated water records; geological assessment; RRI assignment; not company policy; no safety approval.

Screening Rules and Branch Attribution

Document ID: screening-rules-v1 Version: 1.0.0 Evidence class: Paper screening overview Language: English

Two views of support response

The showcase keeps two screening branches visible because they describe different concerns. The static branch combines a section-integrity response with an ultimate-strength response. Its displayed reserve is compared with a 66% showcase screening threshold. The ductility branch represents retained deformation capacity and is compared with a provisional 70% sensitivity threshold. The latter is shown to make a different response pathway visible; it is not an approved operational acceptance criterion.

Each branch is shown with its own value and flag. A case may pass the static screen while triggering the ductility sensitivity screen, or it may show a different combination. The display preserves this disagreement rather than averaging it away. A triggered branch is a reason for human review within this demonstration protocol, not a probability that failure will occur. A non-triggered branch is not a statement that the support is safe.

How to read the displayed result

Read the value, threshold, and flag together, then identify which response pathway is more restrictive. The selected case and age determine which stored result is shown. The assistant may explain the distinction, retrieve the public rule card, or identify missing evidence. It cannot change the numerical result or clear a flag. Explanatory text, a smoothed curve, a response band, or a user-selected region never overrides the raw screening record.

The thresholds are showcase screening rules, not universal material limits. They do not include every site condition, support design decision, inspection observation, company standard, or regulatory requirement. A reviewer needs the applicable site records and authorised engineering judgement before drawing an operational conclusion.

Limitations and authority

These rules organise a demonstration of branch attribution. They do not estimate failure probability, set inspection frequency, determine replacement timing, or grant safety approval. A triggered flag routes a synthetic example toward review; only the responsible human process can assess a real asset. A flag should not be suppressed because another branch passes or because a model-generated explanation sounds reassuring.

Public retrieval terms: screening rules; static branch; ductility branch; 66% threshold; 70% sensitivity threshold; branch disagreement; screening flags; flags cannot be cleared; not a failure probability; no safety approval.

Synthetic Scenario Data Card

Document ID: synthetic-data-card-v1 Version: 1.0.0 Evidence class: Public synthetic demonstration Language: English

What the showcase records represent

The current showcase presents eight selected CTGAN synthetic input records. They were generated for a public demonstration; they are not measurements from eight real rock bolts, supports, mines, or inspection histories. They are used to exercise the same visible workflow a visitor can follow: select a scenario, inspect stored responses, see the screening branches, and ask which evidence would be needed for a real review.

The displayed component responses and screening values are evaluations from frozen paper-model versions applied to the synthetic inputs. They are not hand-authored forecasts and are not live model predictions generated from a visitor's arbitrary data. The assistant can retrieve stored results for supported showcase cases and ages. It does not retrain the paper models or expose their fitted artefacts.

Fictional regions and missing records

The visible records contain environmental descriptors such as chloride, copper and electrical conductivity, together with exposure age and supplied site-context descriptors. The water tool retrieves the synthetic snapshot. The result tool retrieves the selected case and age from stored evaluations and checks that its component values and screening flags agree. Neither tool turns the snapshot into a measured time series or computes a new site calibration.

The map and region names are fictional presentation devices. As a display assumption, all supports represented within one region share one installation batch. This makes a region useful as a simple selection context; it does not supply individual support identities, measured installation dates, or a real mine layout. A selected synthetic row stands for the region in the interface, not a verified population of supports.

Operational evidence that is not present stays absent. In particular, the showcase does not supply a dated history for a real support, a mine plan, confirmed site chemistry, an engineer's geological assessment, or company-approved maintenance rules. The Copilot should request appropriate records instead of inferring or inventing them.

Appropriate use and limits

Use these records to demonstrate the interface, stored-result lookup, source citation, branch explanation, and evidence-gap workflow. They are not a statistical substitute for the original field cohort, independent validation, field prevalence estimates, or proof that a method will generalise.

Is the synthetic heat map a prediction?

No. The synthetic heat map uses inverse-distance weighting (IDW) to shade the eight synthetic endpoint severities across a fictional map. It only fills visual space between those examples: it does not estimate risk at an unobserved location, use mine geometry, or show model predictions for new locations. The displayed endpoints come from stored model evaluations; the colours between them are visual interpolation.

Do not interpret a synthetic value as an observed inspection result, a causal chemical mechanism, a failure date, or an intervention recommendation. The displayed synthetic rows also do not establish a privacy guarantee for every possible derived release.

Public retrieval terms: synthetic data card; CTGAN synthetic inputs; frozen paper models; stored outputs; eight scenarios; fictional regions; shared installation batch; not field measurements; not external validation; not a real mine map.

Trajectory and Input-Response FAQ

Document ID: trajectory-faq Version: 1.0.0 Evidence class: Public visualization explanation Language: English

What does the shaded response range show?

The demonstration illustrates a simple sequence: vary one selected environmental input within a stated showcase assumption, repeat the evaluation, and observe how the two screening branches respond. Each altered input is passed through the same frozen response models used for the stored showcase outputs. The resulting range describes how the displayed outputs vary under that chosen input scenario. It is a conditional input-response illustration, not an estimate of every source of uncertainty.

The range is not a calibrated 95% prediction interval or confidence interval. It does not include model prediction error, uncertainty in model parameters, unvaried environmental factors, or variation between real supports. A wider band means the output is more responsive to the selected input assumption in that illustrated case; it is not by itself a measure of danger, reliability, or probability of failure.

The two sides need not be symmetric around the central result: the same input change can affect the outputs differently in either direction. A narrow or wide band is not determined directly by the model's R-squared score. Repeating the simulation more often can make the sampled response distribution more stable, but does not add the missing model prediction error.

Is every point on the line a new model result?

No. The showcase stores model evaluations at selected query points. A smoothed line connects those points to make the display easier to read. Values between the stored points are visual interpolation, not additional measurements and not a fresh model evaluation for every displayed day. The underlying queried values remain available, and the card values and screening flags follow the original stored results.

The response band is also a visualization of selected query points. Its appearance between those points does not add evidence or create a new daily uncertainty estimate. Smoothing does not modify the model results, move the thresholds, or clear a flag. Read the raw value and its branch flag when interpreting the screen.

Could the same idea be used with real records?

An input-sensitivity workflow can be applied to real data only after the input variation has been justified from relevant records, sampling practice, or measurement knowledge. The demonstration assumption must not be copied automatically into a real site assessment. Real use would also require appropriate model validation and a clear account of the uncertainties that remain outside the varied input.

Limitations

The current band explains a synthetic showcase, not future degradation, remaining life, a failure date, or a calibrated probability. The figure is a communication extension alongside the paper protocol. Visual smoothness and band width do not establish model accuracy or field validation.

Public retrieval terms: trajectory FAQ; Monte Carlo input response; input perturbation; repeated evaluation; response range; not a calibrated prediction interval; not a confidence interval; smoothed display; stored query points; flags unchanged.

SYN-03 / Region B1

Source: CTGAN_0001, CSV record 1. Version: paper-selected-ctgan-b-v3 Raw bundle SHA256: adbf1805af54f3a41ba18898456e205bdf31bbbe511619c436f0ed4fcfc6403b

Fictional region: all supports share one installation batch/date. One selected synthetic row represents the region; individual supports and calendar installation dates are not supplied.

Environmental inputs

{
  "CAI": 0.4435449772309673,
  "RRI": 0.5,
  "pH": 7.997173403591517,
  "Chloride": 1156.7243046807885,
  "Copper": 0.203735723168239,
  "EC": 7028.229018494521,
  "Sulfate": 2931.8308804475555,
  "Temp C": 27.92534041203984,
  "age_days": 437,
  "steel_type": "gal",
  "supplier_alias": "Supplier A",
  "scenario_zone": "SCENARIO_01"
}

Original endpoint predictions

{
  "section_loss_pct": 17.2645834816322,
  "ultimate_strength_loss_pct": -1.0962509802059683,
  "total_elongation_loss_pct": 41.302825927734375,
  "static_reserve_pct": 83.83166749857376,
  "ductility_reserve_pct": 58.697174072265625
}

Provenance and gaps

CTGAN synthetic input evaluated by frozen paper models. Stored model results, not hand-authored losses or live inference. Daily queries cover 90–437 days. No dated water history, inspected support inventory, geological assessment, approved company policy, or calibrated prediction interval is supplied. RRI is an input, not a newly verified geological judgement.