The Intelligent Gate — Can Procurement Red Flags Forecast PPP Renegotiation?

A deployable structural-screening instrument for the Ministry of Finance's fiscal-approval gate, demonstrated — and stress-tested to a null — on Chilean and Brazilian concession panels

Author: David Duarte Arancibia, Austral Intelligence Date: June 2026 (draft) Audience: Ministries of Finance and their fiscal-risk teams; supreme audit institutions and anticorruption bodies; sector regulators; MDB and DFI integrity and infrastructure-advisory units; the open-contracting community Companion software: Austral platform — https://platform.austral-intelligence.com Series position: Paper 5 of the Austral series. It is the renegotiation follow-up flagged in Paper 3, MoF as PPP Orchestrator (§8.3 caveat, §14.3 roadmap), and the empirical front-end of the Renegotiation Triangle idea. Same DNA as the rest of the series: a planner's-perspective decision framework, a reference implementation on the Austral platform, cross-country empirics, a tamper-evident audit chain, reproducible in ten minutes.

Abstract

Under the orchestrator partition set out in Paper 3 of this series, the Ministry of Finance (MoF) does not run the tender — procurement sits with the contracting authority. But the MoF does hold the fiscal-approval gate at award, and renegotiation is the draw it later pays on the fiscal envelope it guaranteed at that gate. This paper asks a narrow, practical question: do a small set of structural procurement red flags — single-bidding, thin competition, award below the sector's competitive norm — computable from the tender record at award, before signature, carry information about the downstream hazard of renegotiation? If they do, the MoF's gate can stop being blind to the competitive integrity of the tenders it underwrites.

We built the instrument and we tested it honestly. The instrument is a structural-indicator layer — now shipped in the Austral platform — that computes participation, timing, and (where bid-level data exist) price-dispersion flags from a normalized concession-tender record, benchmarked within sector. We assembled two project-level concession panels: Chile (123 concessions with complete bidder counts, drawn from the Dirección General de Concesiones workbook, with renegotiations observed in the Modificaciones register) and Brazil (250 concessions with bidder counts from the World Bank's Private Participation in Infrastructure database, of which 92 are transport). On the participation core the two countries look strikingly similar: single-bidding runs at 23% of Chilean concessions and 13% of Brazilian concessions (20% in Brazilian transport) — neither a rarity nor an epidemic.

The headline result is a robust null. In Chile — the only one of the two panels with a programmatic renegotiation outcome, and the country with the longer track record — a naive reading of the raw data appears to show that single-bidder concessions renegotiate less (46% versus 63% on our primary outcome). That apparent protective effect is an artifact of vintage and exposure: single-bidder concessions are systematically younger (mean award year 2014 versus 2009) and so have been at risk for fewer years. Once we model time-to-first-renegotiation properly — a Cox proportional-hazards model and a discrete-time hazard model, each adjusted for sector and award vintage — the effect dissolves. The single-bidder hazard ratio is 0.91 (95% CI 0.47–1.73, p = 0.77) on the primary outcome; three estimators across three outcome definitions all agree, and no structural flag — nor the procurement procedure, nor an unsolicited-proposal origin — forecasts elevated renegotiation, whether in occurrence, timing, recurrence, or severity. (Two associations reach nominal significance and neither survives: a recurrent-count effect that runs the wrong way for a red flag — single-bidder concessions renegotiate slightly fewer times, a vintage artifact — and a two-stage-tendering effect that runs the right way but dissolves under a project-size control and fails a multiplicity correction.) The signal in the data is not competition at all: it is vintage and, secondarily, scale — the award-year hazard ratio is on the order of 1.3 per year and larger concessions renegotiate sooner, while the single-bidder flag on its own discriminates renegotiation at a concordance of 0.52, a coin flip. The finding is consistent with the Engel–Fischer–Galetovic and Guasch account of Chilean concession renegotiation as endemic and institutionally driven — roughly two-thirds of concessions renegotiate, and the draw they pay on the fiscal envelope runs to a median of ~14% and a 90th-percentile of ~42% of the original project budget — the magnitudes a gate should provision against, off the base rate rather than off any flag. That endemic base rate leaves little tender-level variance for any competition flag to explain.

We treat that null as the paper's contribution, not as a disappointment, and we make three claims on it. First, a cautionary one: the survival correction is itself the result. Naive red-flag rates — exactly the rates that integrity dashboards and screening tools report — point the wrong way here, and the structural flags that the procurement-integrity literature has validated as corruption signals do not transfer into renegotiation forecasts in an endemic-renegotiation regime. A gate that screened on them naively would mis-rank its portfolio. Second, a constructive one: the instrument is real, shipped, and reproducible across two countries with one schema and one engine — the value of a screening layer does not depend on its predicting renegotiation, and §8 reframes what the gate should do with a flag that is a process-integrity signal rather than a renegotiation predictor. Third, a policy one: the test was only possible because Chile publishes a structured concession register. Almost everywhere else it is not possible at all. Concessions — the highest-value, highest-risk public contracts a state signs — are carved out of open contracting in most of the jurisdictions we examined; even strong-governance regimes publish little or no bidder-level data. The binding constraint on screening the fiscal gate is not the model. It is that you cannot screen what you cannot see. We close with a call for structured concession-tender disclosure and a roadmap that positions the instrument as the route from a passive gatekeeper to an intelligent gate.

Keywords: Ministry of Finance, public-private partnership, concession, renegotiation, procurement red flags, single-bidding, survival analysis, fiscal commitments and contingent liabilities (FCCL), open contracting, fiscal-approval gate.

Feedback & discussion. This paper is published openly so that practitioners — Ministries of Finance and their teams, supreme audit institutions, anticorruption bodies, sector regulators, MDB and DFI advisors, the open-contracting community, academics — can challenge it, refine it, and put it to work on real programmes. A null result especially invites adversarial reading: we welcome counter-examples, additional country panels, and re-specifications of the hazard model on any section.

• Comment in-page at https://austral-intelligence.com/research (open comments on every section)
• Email the author: [email protected]
• Live demo or walkthrough on request

Table of contents

1. The decision problem — the blind gate
2. Why competition red flags — and why structural, not textual
3. The instrument — a structural-indicator layer for the tender record
4. The panels — Chile and Brazil, hand-built
5. The outcome — renegotiation as a draw on the fiscal envelope
6. The test, and the null
7. Why the null is the answer, not a power failure
8. What the instrument is still for
9. Architecture and reproduction in ten minutes
10. Limitations
11. The policy lead — concessions are dark

Appendices A–E follow the page break.

1. The decision problem — the blind gate

Why this matters: the moment a Ministry of Finance commits the public's contingent exposure to a PPP is the moment of maximum leverage and minimum information. Once the contract is signed, the state's options narrow to renegotiation and the costs of renegotiation fall on the fiscal envelope the MoF guaranteed. If a cheap, auditable signal at the gate could tell the MoF when to clear a deal, when to demand enhanced due diligence, and when to refer or retender, the gate would stop underwriting deals it cannot see inside. This paper asks whether the most-validated family of procurement signals delivers that — and reports, honestly, that in the regime where we can test it, it does not, and why that answer is worth a paper.

1.1 The gate, located inside the orchestrator partition

Paper 3 of this series, From Gatekeeper to Orchestrator, argued that the analytical chain of a PPP programme should be partitioned by information cost: the contracting authority (CA) owns the project's technical half — engineering, demand, environmental risk, operations — and the MoF owns the financial half and the programme-level methodology that binds the technical half to the country's fiscal capacity¹. Even under that stronger, orchestrator reading of the MoF's role, the partition deliberately keeps procurement with the contracting authority. The MoF does not run the tender, score the bids, or award the contract. What the MoF owns is the methodology and the verdict — including, at Stage E of the partition, the decision to commit the country's fiscal envelope behind the contract.

That commitment is an affirmative gate. The MoF must say yes (or no) to a contingent exposure before signature, and several of the orchestrator regimes Paper 3 identified make the gate explicitly affirmative: South Africa's National Treasury issues Treasury Approvals at the procurement-documentation stage (TA:IIB) and at final award (TA:III); Ghana requires the Minister of Finance to personally issue a Seal of Quality before a PPP may proceed¹. The gate is where the public's money is put behind the deal. It is the natural home of a screening instrument.

1.2 The gate is blind to process quality

What the gate sees today is the output of the tender: a winning bidder, a price or an availability payment, a financing structure, a headline fiscal cost. What it does not see — what no line of the fiscal memo records — is the competitive integrity of the process that produced those numbers. Was the tender genuinely contested, or did a single qualified bidder show up? Was the submission window long enough for a real field to assemble, or compressed in a way that favored an incumbent? Did the winning price sit inside the band a competitive field would have produced, or far outside it? These are properties of the procurement process, and they are exactly the properties the MoF's gate cannot observe from the fiscal memo — even though the procurement-integrity literature treats them as among the most informative signals a public contract emits².

The asymmetry is structural, not accidental. The CA holds the tender file; the MoF holds the fiscal envelope. Under the gatekeeper doctrine the MoF accepted a one-page summary because observing the process cost more than it was worth. The premise of this paper — the same premise that animates the whole series — is that the observation cost has collapsed: the participation and timing properties of a tender can be computed from a normalized tender record at near-zero marginal cost, audit-chained, and presented to the gate alongside the fiscal numbers. The question is whether doing so buys the gate anything.

1.3 The decision the instrument is meant to support

We frame the gate's decision as a three-way action, not a binary pass/fail:

• Clear — the tender's structural profile is unremarkable; the fiscal commitment proceeds on its merits.
• Enhanced due diligence (EDD) — the structural profile carries one or more flags; the MoF requires additional review, a documented justification from the CA, or a tighter renegotiation-control clause before committing.
• Refer or retender — the profile is severe enough that the MoF refers the tender to the audit or anticorruption body, or declines to underwrite until the procurement is re-run.

The thresholds between these actions are a cost-asymmetry problem: the expected fiscal cost of underwriting a deal that later renegotiates badly, against the cost of an unnecessary retender or a delayed close. We return to the decision rule in §8 — and we return to it changed, because the empirical core of this paper bears directly on what a flag at the gate actually means.

1.4 What we set out to find, and what we found

We set out to build the predictive link the procurement-integrity literature implies: structural red flags at award → elevated hazard of renegotiation downstream. We built the instrument, assembled two concession panels, and fit the hazard model. The link is not there — not in the country where we can measure it, and not for any flag, outcome definition, or estimator we tried. The bulk of this paper is the honest reporting of that null: why the raw data misleads (§6.1), why the corrected answer is genuinely null rather than merely underpowered (§7), and what a screening instrument is for once you accept that — in an endemic-renegotiation regime — a competition flag is a process-integrity signal and not a renegotiation forecast (§8). The null is not a failure of the instrument; it is a finding about the regime, and it is the kind of finding a fiscal gate needs before it trusts a number on a dashboard.

A note on analysis provenance. This paper began as a confirmatory study. Its pre-specified hypothesis was that the single-bidder flag predicts an elevated hazard of renegotiation; its primary outcome was the contract-driven renegotiation measure (excl_newworks); and the three-outcome × three-estimator grid of §6 was fixed as a robustness lattice, not assembled after inspecting results. The hypothesis failed. "Embracing the null" is an honest post-hoc reframing of a confirmatory design that returned a negative answer — not a null dressed up as a discovery. The scale, procedure, and recurrent-count analyses (§6.4–§6.5) are exploratory extensions prompted by the null; we label them as such, and we apply a multiplicity correction to the competition-axis hazard tests (§6.5). The recurrent-count and severity probes are a separate, uncorrected family — but their one nominal result runs protective (the wrong way for the hypothesis), so it cannot be a false positive for it.

2. Why competition red flags — and why structural, not textual

Why this matters: if the gate is going to screen on a signal, the choice of signal is the whole game. We argue for a specific, narrow family — structural competition indicators — and against the two obvious alternatives: rhetorical/text-based corruption flags, which the most-corrupt tenders can write away, and bespoke per-project forensic review, which does not scale to a portfolio gate. The choice is not cosmetic: it determines what can be computed cheaply, audited, and benchmarked across countries, and it is what makes the null in §6 a clean result rather than a measurement failure.

2.1 The validated core is single-bidding and thin competition

The empirical procurement-integrity literature has converged on a small set of structural indicators — properties of the bidding process itself — as the most robust observable correlates of procurement risk. The single most validated is single-bidding: a tender that attracts only one qualified bid. Fazekas and co-authors built the Corruption Risk Index (CRI) around it and a handful of allied structural flags, validated against independent corruption measures across large European contracting datasets²; the Open Contracting Partnership and a broad red-flags literature have since adopted single-bidding and thin competition as front-line screening signals³. The appeal is that these flags are behavioral and structural: they measure what happened in the contest — how many bidders, how long the window, how the winning price sits against its peers — not what the documents say about it.

We take that core and adapt it to concessions, benchmarked within sector. A four-bidder field is healthy for an urban toll road and thin for a competitively tendered airport; a single bid is less surprising for a large, capital-intensive railway concession (where the qualified field is small everywhere) than for a standard road. The instrument therefore computes single-bidding and a bidder count, a below-sector-norm flag relative to the sector's competitive median, a submission-window/short-window timing flag, and — where bid-level data exist — a price-dispersion flag. §3 gives the definitions and §4 the panels.

2.2 Why not the text scanner: the disclosure-selection critique

A natural alternative is to screen the text of the tender — to scan the procurement documents and the award justification for the rhetorical fingerprints of a rigged process (tailored specifications, unusual eligibility carve-outs, boilerplate single-source justifications). The Austral platform already ships such a scanner. We demote it deliberately, to a qualitative overlay, and make the structural layer the predictive core, for one decisive reason: disclosure selection. Text flags fire on disclosed prose, and disclosed prose is written by the same party that ran the process. The most sophisticated capture produces the cleanest documents — a tender engineered to favor an incumbent will, if anything, read more carefully than an honest one. A text scanner is therefore most likely to miss exactly the cases that matter, and its false-negative rate is correlated with the very thing it is trying to detect. Structural indicators do not have this defect: a single-bidder tender is single-bidder no matter how the file reads, and the count of bidders cannot be written away. For a gate that must defend its verdict to an audit institution, a signal that cannot be drafted around is worth more than one that can.

The demotion is warranted on this paper's own data, not merely asserted. Running the shipped text scanner over the tender text the Chilean register actually publishes returns zero integrity findings across all 123 concessions — while the structural layer fires a single-bidder flag on 28 of them. The reason is the argument: the scanner keys on the rhetorical fingerprints of a rigged process (sole-source justifications, bid-rotation language, conflict-of-interest and beneficial-ownership prose), and a structured concession register exposes none of that — only short, sanitised descriptive fields, never the full tender documents where such language, if it existed, would live. The structural signal survives this disclosure selection because a bidder count cannot be redrafted; the textual signal does not, because the prose a scanner needs is exactly what a sanitised or unpublished record withholds. (We grant a language caveat — the scanner's rules are English, the register Spanish — but it does not rescue the text approach: the deeper point is that no full tender prose is published in machine-readable form at all, in either language. That is the §11 data-darkness argument arriving one section early.)

2.3 Price dispersion is the strongest flag — and the one we mostly cannot see

The most powerful structural flag, where it is observable, is price dispersion: the spread of competing bid amounts, and the gap between the winner and the runner-up. A field that bids in a tight, implausibly coordinated band, or a winner that wins by an anomalous margin, is a stronger signal than head-count alone. But price dispersion requires the amounts every bidder offered, not just their identities — and that is precisely the field that almost no concession regime publishes in structured form. In our two panels it is observable in Brazil (federal auction atas record competing bids) and not in Chile (the winning bid amount is locked in scanned-PDF award records, and losing bids are not systematically published). Price dispersion is therefore a Brazil-only extension and a worked demonstration of the instrument's reach (§3.3, §10), not a flag we can carry into the cross-country test. This gap is not incidental to the paper; it is the first instance of the data-darkness argument that §11 makes the policy headline.

2.4 The flag respects the partition

One property of the structural family matters for the orchestrator argument specifically: it lets the MoF screen the gate without pulling procurement onto the MoF's side of the partition. The MoF does not need the tender file, the bid documents, or the CA's evaluation reasoning to compute a single-bidder flag or a within-sector bidder-count benchmark — it needs only the structured participation record, which is a by-product the CA already produces. The instrument consumes a signal at the gate; the CA continues to execute the procurement. That is the orchestrator partition working as designed: the MoF reads the result of the process at the resolution it needs to set a fiscal verdict, and leaves the running of the process where the technical information is.

3. The instrument — a structural-indicator layer for the tender record

Why this matters: the instrument is the part of this paper that does not depend on the empirical result. Whatever the hazard model finds, a gate that can compute the competitive structure of a tender at near-zero marginal cost, benchmarked within sector, audit-chained, and reproducible across countries, is a capability the MoF did not previously have. We built it as a platform module and shipped it; this section describes what it computes and why each flag is defined the way it is.

3.1 What the layer computes

The structural-indicator layer takes a normalized tender record — a small, country-agnostic object carrying the tender's identifier, sector, country, the number of bidders (or the individual bids, from which the count is derived), the publication and submission-deadline dates, and the awarded amount — and returns a StructuralIndicators object with the following fields:

Each fired flag carries a severity — high, medium, or low — and the composite structural_score starts at 100 and deducts a severity weight per flag (high = 20, medium = 10, low = 5), floored at zero. The score is a convenience for ranking and display; the paper's analysis uses the underlying flags and the score alike, and finds none of them predictive (§6). The full catalogue — code, severity, dimension, and definition for every flag — is served by the live GET …/procurement/structural/indicators endpoint and reproduced in Appendix B.

3.2 Within-sector benchmarking

The single design decision that matters most is that the participation flags are benchmarked within sector, not against a global norm. A large, capital-intensive railway or airport concession draws a small qualified field everywhere on earth; a standard urban road draws a larger one. A global "fewer than three bidders" rule would flag the railway as suspicious and clear the road, exactly backwards. The layer therefore computes a sector benchmark (the median bidder count) across the submitted panel and fires below_sector_norm relative to that sector's median, and thin_competition relative to the sector's typical field. The endpoint is panel-aware: submit a panel and the benchmark reflects that panel; submit a single record and the benchmark is degenerate, so below_sector_norm stays null while the participation and timing flags still compute. This is what lets the same instrument run on a Chilean road panel and a Brazilian electricity panel and mean the same thing in each.

3.3 The price-dispersion extension — a Brazilian worked example

Where the amounts every bidder offered are available, the layer computes two price flags: low_price_dispersion (an implausibly tight coefficient of variation across competing bids, a signature of coordinated bidding) and winner_field_outlier (an anomalous gap between the winner and the rest of the field). These are the strongest structural flags in the integrity literature — and, as §2.3 explained, the ones we can almost never see, because concession regimes publish winner identities far more often than they publish the full bid table. In our two panels the price flags are computable in Brazil and not in Chile, so we demonstrate them on a hand-coded set of Brazilian federal concession auctions whose competing-bid amounts are public. This is an instrument demonstration, not a test: it shows the engine computes the price family correctly on real auctions. It is emphatically not evidence that price dispersion forecasts renegotiation — Brazil has no programmatic renegotiation outcome to test against (§4.2), so these flags play no part in the cross-country result of §6.

Brazilian concession auctions use two different competitive variables, and the convention matters for reading the gap. Airport rounds are won on the largest grant payment (outorga, in reais); toll-road rounds on the largest tariff discount (deságio, in percentage points). In both, the winner is the high offer — so a winner-field outlier here is the mirror image of the lowball the flag's prose describes: the winner over-promised relative to the pack, the same winner's-curse and later-recovery risk with the sign reversed. The engine computes |winner − median(rest)|, so it is indifferent to the sign; we read the magnitude. The coefficient of variation is scale-free and both quantities are computed within a single auction, so mixing reais and percentage points across rows is sound.

Running the shipped engine (default thresholds: low_dispersion_cv = 0.05, large_gap_pct = 0.15) over the set:

Each row exercises a different part of the layer. Confins (2013) drew only two bids — R$1.82bn and R$1.80bn, twenty million reais apart on a ~R$1.8bn grant — a coefficient of variation of 0.78%, well under the 5% threshold, so low_price_dispersion fires alongside thin_competition (two bidders) and below_sector_norm (two is under the airport panel's median of four). Tight bids are not proof of anything; the flag is a prompt for review, never a finding (§10), and two serious bidders converging on a similar valuation is one innocent reading among several. Nova Dutra (2021) is the winner-field case: CCR took the maximum 15.31% tariff discount the edital allowed — and paid a R$1.77bn grant on top — while EcoRodovias offered 10.60%, a 44.4% relative gap that trips winner_field_outlier. The winner did not lowball; it over-promised, maxing the discount cap on a corridor previously handed back for re-tender — the winner's-curse pattern the flag is meant to surface. Paraná Lote 2 (2023) drew a single proponent bidding a token 0.08% discount: single_bidder, and with one bid there is no dispersion to compute. Natal (2011) is the clean control — four consortia, eighty-eight live rounds, healthy competition, no flag.

The within-sector benchmark behaves exactly as designed across this tiny panel: with three airports it is usable (median four bidders, so Confins's two trips below_sector_norm), while with only two roads it falls under the min_sector_n = 3 guard and stays degenerate — which is why neither road carries a below-norm flag even though Paraná had a single bidder. The instrument refuses to benchmark a sector it has not seen enough of.

Galeão (2013) is the row that proves the paper's point by its blanks. Five consortia submitted sealed bids ranging from R$6.56bn to R$19.018bn; the winner, an Odebrecht–Changi consortium, bid R$19.018bn — a 293.9% premium over the R$4.82bn floor. Within a decade the concession had failed and been handed back; it was re-tendered in 2026, with Aena winning at R$2.9bn. It is, in hindsight, the textbook winner's-curse outcome the price flags exist to anticipate. And it is precisely the auction we cannot score on dispersion: only the floor and the winning bid were ever published, the three middle sealed bids are dark, so the engine sees five bidders (healthy participation, no flag) and no computable bid field. The one auction history proved was a disaster is the one whose bid table was never disclosed. That is the data-darkness argument of §11 compressed into a single row — and the reason the price family, powerful as it is where visible, cannot carry the cross-country test.

Scaling the harvest. The five auctions above were the original worked set; we have since extended it to 14 sourced federal auctions (seven airport blocks and seven road concessions, 2013–2025), every bid traced to a public source and run through the same shipped engine. The price family computes on the ten with fully-disclosed competing bids, and the pattern is consistent and striking: the winner-field-outlier flag fires on eight of the ten, with winners routinely out-bidding the rest of the field by multiples — the 2019 and 2021 airport blocks saw the winner offer two to thirty times the runner-up's grant — while two auctions show the opposite signature, a suspiciously tight cluster that trips low_price_dispersion. Four further auctions drew a single bidder.

This is the Galeão pattern generalised: aggressive winner over-promising is endemic in Brazilian concession auctions, and the engine surfaces it wherever the bid table is public. It remains an instrument demonstration, not a test — Brazil still has no programmatic renegotiation outcome to regress these flags against (§4.3, Appendix E.2) — but it shows the price family, the strongest in the integrity literature and the one Chile cannot see, computing correctly across a real and varied set. Every figure traces to a public source; auctions disclosing only the winner and floor, or resolved through viva-voz rounds that leave competing bids non-comparable, were left uncoded rather than guessed.

3.4 It ships

The layer is a registered module in the Austral platform's procurement package, with its own request/response contract (POST …/procurement/structural-scan), a unit-test suite, and the tamper-evident audit chain the rest of the series uses. It is not a paper prototype: the same endpoint that produced the indicators for the panels in this paper is the one a Ministry of Finance would call at its gate. §9 walks the call end-to-end and Appendix A gives a copy-pasteable example.

4. The panels — Chile and Brazil, hand-built

Why this matters: the credibility of a null rests on the credibility of the data behind it. We are explicit that this is a built dataset — hand-assembled from a published workbook and a structured international database, not drawn from a live open-contracting feed, because (as §11 argues at length) no such feed exists for concessions. The honesty cuts both ways: it bounds the claims we can make, and it is itself the evidence for the policy argument.

4.1 Chile — the primary panel

Chile is the primary panel because it is the rare jurisdiction that publishes both the tender's participation record and its renegotiation outcome in structured form. The Dirección General de Concesiones (DGC) of the Ministry of Public Works maintains a concessions workbook on the national open-data portal⁹, keyed by a stable project code (e.g. 028_RVLC1), with separate sheets for contracts, bidders (Oferentes, carrying each bidder's name and an Adjudicado award flag, from which the bidder count is directly countable), participants and their nationality, and — critically — modifications (Modificaciones), the register of contractual renegotiations. The workbook lists 140 concessions; we retain the 123 with a complete bidder count. The bidder count is the number of distinct Oferentes — a consortium counts as the single bidder it is, which is the unit the workbook itself records; a separate participant sheet, which we do not use for the count, lists each consortium's members. They span interurban and urban roads, airports, hospitals, prisons, public buildings, and water solutions, with contract starts from the 1990s onward.

One structural feature of the register matters for every rate and hazard below, so we surface it here rather than bury it in the limitations: the DGC workbook records only concessions in an active or completed state — 46 in operation, 42 finished, 18 in construction, 17 in mixed construction-and-operation — and carries no cancelled, caducated, or relicitación hand-back state at all. The panel is therefore conditioned on survival into the register: a concession that failed early and was struck from the active programme does not appear. We carry this forward in §10 and read the renegotiation base rate as a property of the surviving population.

The participation picture: 28 of 123 (23%) are single-bidder. The within-sector competitive medians are unremarkable and sector-typical — interurban roads (n = 45) run a median of 3 bidders, airports (n = 29) a median of 4, hospitals (n = 13) a median of 4, prisons (n = 6) a median of between 1 and 2. The structural flags fire as follows across the panel: single-bidder 28, below-sector-norm 21, thin-competition 18, short submission window 5.

4.2 Brazil — the complement

Brazil enters through a different door. We had expected to hand-code Brazilian participation from federal auction records, but the World Bank's Private Participation in Infrastructure (PPI) database carries a populated bidder-count field (numberb) for Brazilian concessions — which means the participation core is available programmatically, without parsing a single PDF. We retain 250 Brazilian concessions with bidder counts, of which 92 are transport (57 toll roads, 17 airports, 11 seaports, 7 railroads), alongside 107 electricity and 43 water-utility concessions, with financial-closure years spanning 1998–2024.

The participation picture is strikingly close to Chile's: 33 of 250 (13%) are single-bidder overall, and 18 of 92 (20%) in transport — against Chile's 23%. Single-bidding in Latin American concessions is, on this evidence, neither a rarity nor an epidemic; it is a steady background rate in the low-to-mid tens of percent, somewhat higher in the capital-intensive transport sectors. One sector stands out: Brazilian railroads run a median of one bidder (n = 7) — a structurally uncompetitive sector everywhere, and exactly the case the within-sector benchmark exists to keep in proportion.

4.3 What we can and cannot join

The asymmetry between the two panels is the hinge of the whole paper. Chile gives us a programmatic outcome: the Modificaciones register is the renegotiation dependent variable (DV), joinable to the tender record on the stable project code. Brazil does not. PPI's status field is near-degenerate for our purpose — only 3 of 250 concessions are recorded as cancelled, and there is no renegotiation field at all — and the federal renegotiation record (ANTT aditivos for roads and railways, the analogous ANAC instruments for airports, and relicitação hand-backs under Law 13,448/2017) is PDF-only, with no stable key back to the auction¹⁰. We therefore have a clean participation comparison across two countries, but a time-to-event renegotiation test in Chile only. We state this plainly here because it determines the scope of every empirical claim that follows: the hazard model is a Chilean result, and §10 and §11 carry that limitation forward rather than papering over it.

5. The outcome — renegotiation as a draw on the fiscal envelope

Why this matters: the choice of outcome variable is what makes this an MoF instrument rather than an integrity curiosity. Renegotiation is not a governance abstraction; it is the mechanism by which a signed PPP reaches back into the public purse after the gate has closed. Choosing it as the DV — and framing it as a draw on the fiscal envelope the MoF guaranteed — is what connects the procurement signal to the fiscal commitment.

5.1 Why renegotiation, not cancellation

The integrity literature often studies cancellation or outright failure. For a fiscal gate, and at our sample sizes, that is the wrong outcome: cancellation is rare (3 of 250 in Brazil) and carries almost no statistical power. Renegotiation is frequent and early. Engel, Fischer, and Galetovic document that the large majority of Latin American transport concessions are renegotiated within a few years of award; Guasch's canonical World Bank study put the incidence across Latin American transport and water concessions at roughly half to three-quarters, with renegotiation arriving on average within about 2.2 years of award⁵⁴. In our Chilean panel — a regime where renegotiation tends to come later than that regional average — 78 of 123 concessions (63%) renegotiate at least once, with a median time to first renegotiation of about 6.6 years (a cause-specific incidence; §6.2 bounds the long-horizon figure under competing risks). An outcome that common is one a hazard model can actually resolve.

5.2 The FCCL bridge

Framed in the language of Paper 3, a renegotiation is a draw on the Stage-E fiscal envelope — the contingent commitment the MoF underwrote at the gate¹. In Chile that envelope is not an abstraction: it is built from named instruments — the ingreso mínimo garantizado (minimum revenue guarantee, IMG) and the revenue-distribution mechanism (mecanismo de distribución de ingresos, MDI) that share demand risk between the state and the concessionaire, and, in the contracts designed to curb renegotiation, the least-present-value-of-revenue (LPVR) variable-term auction that Engel, Fischer, and Galetovic devised and Chile pioneered⁶. A scope addition, a term extension, a compensation for higher costs: each is a post-signature transfer that lands on those accounts. This is the bridge that makes a procurement red flag a fiscal instrument: if a structural property of the tender forecast the hazard of that draw, the MoF could price the contingent commitment more accurately at the gate. The empirical core of this paper tests precisely that "if" — and answers it in the negative for the flags we have.

5.3 Three definitions of the outcome

Renegotiation is not monolithic, so we test three nested definitions of the DV, from broadest to strictest:

The excl_newworks variant is our primary outcome: it strips out renegotiations dominated by state-initiated programme expansions, retaining the contract-driven renegotiations a red flag might plausibly forecast. We are careful not to over-claim the exogeneity of what we remove — in the Chilean literature, additional-works (Nuevas obras) are a prime opportunistic channel, the very recovery route a lowballing single bidder might exploit, not a clean state-only category⁸. That is precisely why we also report the broadest any variant, which retains them — and which is also null (§6.3). The excl_newworks cut is the cleaner read of contract-driven renegotiation, not a claim that new works are competition-orthogonal. The financial variant isolates the purest fiscal draws but, as §7 discusses, is underpowered in the single-bidder arm (only 5 single-bidder events) and cannot by itself carry the inference. Severe outcomes (cancellation, relicitação hand-back, arbitration) are kept only as a coarse robustness check.

6. The test, and the null

Why this matters: this is the analytical core. It contains the result the rest of the paper is built around — and it contains a worked illustration of exactly the mistake an unguarded screening dashboard would make. The section is deliberately structured to show the trap first, then the correction, because the gap between them is the cautionary finding.

6.1 The bivariate trap

Read naively, the Chilean data appears to say that single-bidding is protective. The raw renegotiation rates run the wrong way for a red flag:

A screening tool that reported these rates — and many do report exactly this kind of cross-tab — would conclude that single-bidder tenders are the safer ones, and a gate calibrated on them would systematically misallocate its scrutiny. The conclusion is an artifact. Single-bidder concessions in Chile are markedly younger: their mean award year is 2014 versus 2009 for multi-bidder concessions (medians 2018 versus 2009). They have simply had fewer years at risk, and a raw rate that does not condition on exposure mistakes youth for safety. The correction for this is not a robustness check; it is the analysis (Figure 1).

Figure 1 — The bivariate trap. Panel A: the raw renegotiation rates of §6.1 make single-bidder concessions look seventeen points safer. Panel B: single-bidder concessions are systematically younger (mean award 2014 versus 2009), so they have been at risk for fewer years. Conditioning on time at risk dissolves the gap (adjusted hazard ratio 0.91, p = 0.77). Reading youth as safety is precisely the mistake an unguarded screening dashboard makes.

6.2 The survival correction

We model time to first renegotiation with three estimators, each from an established library for reproducibility: a Kaplan–Meier estimator with a log-rank test stratified by single-bidder status; a Cox proportional-hazards model adjusting for award vintage and sector; and a discrete-time hazard model (complementary log-log link on annual person-period data). Concessions are right-censored at an administrative snapshot (the renegotiation-register extraction date, 2026-01-09), or at contract end if earlier. The Kaplan–Meier curves for single- versus multi-bidder concessions are essentially coincident — on the primary outcome the median time to first renegotiation is 14.3 years for single-bidder and 13.1 years for multi-bidder concessions, with a log-rank p of 0.27 (Figure 2).

Figure 2 — Time to first renegotiation, primary outcome (excluding state-initiated new works), single- versus multi-bidder Chilean concessions. The two curves are essentially coincident (log-rank p = 0.27); the apparent single-bidder advantage in the raw rates of §6.1 does not survive conditioning on time at risk. This near-coincidence, not a separation, is the result.

Two checks confirm the correction is sound rather than a modelling artifact. First, the proportional-hazards assumption holds for the single-bidder term (Schoenfeld test p = 0.13); the transport sector violates it (p = 0.01), so we stratify the baseline hazard by sector, and the single-bidder hazard ratio is unchanged at 0.88 [0.43–1.77]. Second, renegotiation is not the only exit from the risk set: 26 of the 123 concessions reach the end of their contract term without ever renegotiating — a competing terminal event that a single-decrement Kaplan–Meier curve mistreats as censoring. Re-estimating with an Aalen–Johansen competing-risks estimator (Figure 3) puts the 20-year cumulative incidence of first renegotiation at about 58%, against the 69% the Kaplan–Meier curve implies — the single-decrement curve overstates long-horizon incidence by roughly eleven points — while leaving the single-versus-multi comparison null. We therefore read renegotiation incidence throughout as a cause-specific quantity.

Figure 3 — Cumulative incidence of first renegotiation. The Kaplan–Meier curve (which treats a concession reaching contract-end without renegotiating as censoring) overstates the competing-risks-correct Aalen–Johansen incidence by about eleven points at twenty years. The headline null is unaffected; the figure bounds the long-horizon incidence claims of §5.1 and §8.3.

6.3 The result

Once exposure, sector, and vintage are accounted for, the apparent protective effect disappears and nothing replaces it. The single-bidder hazard ratio sits on top of 1.0 on the broadest outcome and drifts only modestly below it on the others, never approaching significance, and the discrete-time estimator agrees with the Cox estimator across the board:

The result is not specific to single-bidding. Every structural covariate we tested is null on every outcome and under both the covariate-adjusted and sector-stratified specifications — the bidder count (HRs 0.92–1.18, all p ≥ 0.13), the below-sector-norm flag (HRs 0.66–1.32, all p ≥ 0.27), and the composite structural score (HRs ≈ 0.99–1.03, all p ≥ 0.21). There is no flag, no outcome definition, and no estimator under which a competition red flag forecasts elevated Chilean renegotiation (Figure 4).

Figure 4 — Every structural flag is null. Adjusted hazard ratios (vintage + sector controls) for the four structural covariates across the three outcome definitions; the primary outcome is bolded, and the sector-stratified specification (not shown) agrees. Every interval straddles a hazard ratio of 1.0. The absence of separation — a wall of confidence intervals on the no-effect line — is the result.

6.4 What carries signal: vintage, and secondarily scale

The model is not failing to find signal — it is finding a different signal, and it is not competition. Award vintage dominates: in the primary model the start-year hazard ratio is 1.31 per year (95% CI 1.23–1.40, p ≈ 10⁻¹⁶) — each later year of award multiplies the renegotiation hazard by about a third. A second, weaker signal is scale: the official-budget term carries a hazard ratio of 1.49 per standard deviation of log budget (95% CI 1.05–2.14, p = 0.03) — larger concessions renegotiate sooner, independent of vintage and sector. Both are properties the gate can read off the fiscal memo; neither is competition. The decomposition of model fit:

A model with nothing but vintage and sector reaches a concordance of 0.85; adding scale lifts it to 0.86; adding every structural flag on top moves it by a further 0.001. The discriminatory power in the data is vintage and, secondarily, scale; the competition flags contribute essentially nothing. That is the null in one line: at the Chilean gate, when a concession was signed and how large it is carry information about its renegotiation hazard; how competitively it was tendered does not.

That the model separates curves at all answers the obvious objection to any null — that the data simply has no signal to find. It has: stratified by award vintage, the survival curves fan out cleanly (Figure 5). The competition flags lack the discrimination vintage has; they do not lack a model capable of showing it.

Figure 5 — A positive control. Stratified by award-year tercile, the survival curves separate markedly (vintage hazard ratio ≈ 1.31 per year, concordance 0.85) — the discrimination the competition flags conspicuously lack. The null of §6.3 is a property of the flags, not a limitation of the design.

The vintage effect is not a mechanical artifact of right-censoring at a fixed snapshot. Restricting to the fully-exposed cohort awarded on or before 2020 (n = 98, 65 events), the vintage hazard ratio is essentially unchanged at 1.34 [1.25–1.44] and the single-bidder hazard ratio remains null at 0.86 [0.42–1.77]. We read vintage as a reduced-form cohort proxy — later concessions were let under a progressively more renegotiation-permissive framework, around and after the 2010 reform (Law 20.410) that created Chile's Panel Técnico de Concesiones — and we do not claim to have decomposed which institutional change drives it.

6.5 The null is not an artifact of the flag set, the outcome, or the procedure

A sceptical reader will press the null on three further margins. It holds on all three.

Recurrence and severity. The hazard model tests when a concession first renegotiates. A fiscal gate also cares how often and how much; neither rescues the flag. A negative-binomial recurrent-count model (renegotiation count, log-exposure offset, vintage + scale + sector) returns a single-bidder incidence-rate ratio of 0.46 (p = 0.03) — single-bidder concessions renegotiate fewer times, not more. That is the §6.1 trap again: it is the wrong sign for a red flag, and it is an exposure-shape artifact — renegotiations accrue late in contract life, and single-bidder concessions, being younger, sit earlier on that back-loaded curve, so a constant-rate offset under-credits their shorter histories. The conditional severity of the draw — log renegotiated UF among the 72 concessions with a recorded amount, on the same covariates — is cleanly null for single-bidding (coefficient −0.19, p = 0.72), while scale predicts severity strongly (coefficient 1.43, p < 0.001), echoing §6.4. Across occurrence, timing, recurrence, and severity, no flag identifies the concessions that renegotiate more often or pay out larger — the only thing a screening flag would be for.

Procedure and origination. Perhaps the flag is the wrong kind of competition signal: the realised bidder count is a noisy proxy for the pressure a bidder actually faced — a sealed-bid entrant who believed rivals were coming bids competitively whether or not any materialised. A cleaner signal is the procedure, which the panel records. Two procedure variables have usable variance in the Chilean data, and both are null. Origination as an unsolicited private initiative (iniciativa privada; 25 of 123) — a structurally less-contested route, in which the original proponent enjoys a scoring bonus — carries a hazard ratio of 0.82 [0.43–1.55], p = 0.53: null, and again the wrong sign. Two-stage tendering with prequalification (62 of 123) appears at first to bite — 1.73 [1.03–2.91], p = 0.04 — but it is a scale confound: prequalification is used for the larger projects, so once budget enters the model the effect falls to 1.66 [0.97–2.86], p = 0.07, and to p = 0.09 with the full control set. Two-stage tenders also draw more bidders, not fewer (mean 3.5 versus 2.9), so the residual is not a thin-field effect.

Multiplicity. Testing this many covariates invites the forking-paths objection, so we meet it directly. The one nominally significant competition-axis result (two-stage tendering, p = 0.04) does not survive correction over the six structural and procedure covariates tested on the primary outcome: its Benjamini–Hochberg-adjusted p-value is 0.23 and its Bonferroni-adjusted value is also 0.23. No competition-axis covariate is significant after correction. The procedure and count probes are exploratory by design, and we report them as such — not dressed as confirmatory tests (see the pre-specification note in §1.4).

7. Why the null is the answer, not a power failure

Why this matters: a null invites one obvious dismissal — "you just didn't have the data to find the effect." We take that objection seriously and answer it directly, because the difference between an underpowered study and a genuine null is the difference between "we don't know" and "we know it isn't there." On the broad outcomes, we are in the second case.

7.1 The broad outcomes are well-powered

Seventy-two to seventy-seven events on 123 concessions is not a thin design for a single binary covariate, and "well-powered" is a quantitative claim we make quantitatively. With 72 events on the primary outcome and single-bidder concessions making up 23% of the panel, the design has 80% power to detect a single-bidder hazard ratio of about 2.2 (or, symmetrically, a protective 0.46) at the 5% level. A doubling of the renegotiation hazard is the order of magnitude a red flag would need to earn a place at the gate, and it is excluded: the estimate is 0.91 and its interval, 0.47–1.73, sits on no effect. We are careful not to over-claim the other direction — the upper bound does admit a hazard ratio as high as 1.73, so the data does not rule out a modest (≈1.7×) elevation. But an effect the study is only borderline-powered to see, whose point estimate sits just below one, and which would have to be weighed against a 63% base rate, is not a signal a fiscal gate would screen on. The one outcome where the power objection genuinely bites is the financial variant, with 36 events and just 5 in the single-bidder arm; we flag it as underpowered and do not lean on it. The inference rests on the broad outcomes.

7.2 The effect is identified, not absorbed by vintage

A second objection: perhaps single-bidding is dangerous, but vintage "ate" its coefficient because the two are collinear. They are not. The correlation between single-bidding and award year is only 0.23 — single-bidder concessions are somewhat younger, but single-bidding is nowhere near a proxy for vintage. The model can and does separate the two, and when it does, vintage stays strong while the flag collapses to ≈1. The null is a property of the data, not an identification failure.

Nor is it a measurement artifact of how bidders are counted. A noisy bidder count would bias the hazard ratio toward 1.0, which could manufacture a spurious null — so we re-fit on the unambiguous open single-stage subsample (n = 61), where the count is cleanest (a consortium is one Oferente and there is no prequalification round to muddy who submitted). The single-bidder hazard ratio there is 1.07 [0.39–2.91] — if anything moving away from a protective reading, and certainly not the elevation a red flag would need. The cleanest half of the panel does not show the effect re-emerging.

7.3 It is what the regime predicts

The cleanest reason to believe the null is that it is exactly what the literature on Chilean and Latin American concessions would lead one to expect. Renegotiation in this regime is endemic and institutionally driven — roughly two-thirds of Chilean concessions renegotiate, consistent with the half-to-three-quarters that Guasch and the EFG work report region-wide⁴⁵⁷. Bitran and co-authors, studying Chile and Colombia specifically, locate the drivers of renegotiation in the contractual and institutional environment — opportunistic bidding against a renegotiation-permissive framework, government-initiated scope changes, and the structure of the concession contract — not in the size of the bidding field at award⁸. When two-thirds of contracts renegotiate regardless of how they were tendered, there is very little tender-level variance left for a competition flag to explain. The null is not a surprise against this backdrop; the confirmatory result would have been.

7.4 What would have to be true for the flag to predict

For single-bidding to forecast renegotiation, the mechanism of renegotiation would have to run through the competitive deficit at award — for instance, a lone bidder lowballing to win and then renegotiating the price up. In Chile the dominant mechanism is different: renegotiation is the routine administrative channel for scope changes, ministry-initiated works, and cost-overrun compensation, and it operates more or less independently of how many bidders showed up. The null is therefore informative about the mechanism: it tells the MoF that, in this regime, controlling renegotiation is about contract design and the renegotiation framework itself, not about policing the size of the bidding field.

7.5 The null has a country-level echo

The project-level null sits inside a country-level literature that points the same way. Guasch's canonical study of more than a thousand Latin American concessions located the drivers of renegotiation in the regulatory and contractual environment, not in the competitiveness of the award — and his data run the same direction as ours on competition itself: concessions awarded by competitive bidding renegotiated more often than those awarded bilaterally (46% versus 8%), while the largest reducers were a regulatory body in place at award (17% versus 61% without it) and the regulatory regime (rate-of-return 13% versus price-cap 42%), not the size of the bidding field⁵⁷. Whatever the confounds in that cross-section, it is not a setting in which award competition looks protective.

Against the most apt comparison — road concessions, the closest match to our transport-heavy panel — Chile is endemic but comparatively disciplined. Bitran and co-authors record 18 of 21 Chilean road concessions (86%) renegotiated, against 84% in Colombia and 73% in Peru; but Chilean concessions were renegotiated about 3.3 times each, against 20.5 in Colombia, with a first renegotiation at 2.7 years versus 1.0⁸. Our panel's 63% (all sectors, on the broadest any outcome; 59% on the contract-driven primary outcome) and ~6.6-year median sit comfortably inside that frame: Chile renegotiates endemically, but less furiously than its neighbours.

Two cautions keep this honest. Guasch's own headline figure for Chile is far lower (~8% overall) — but it is an early, energy-weighted sample running only to 2000, not comparable to the post-2000 transport-and-social MOP panel here; we read it for the cross-country ranking and the drivers, not as a level for Chile. And the country-level evidence is observational and confounded throughout. What it supplies is not a second test but a consistent backdrop: wherever the determinants of renegotiation have been studied across countries, they are institutional and contractual, and award competition is not among them — exactly what a fiscal gate should expect before it screens on a competition flag (§8).

8. What the instrument is still for

Why this matters: it would be easy to read §6–§7 as "the instrument is useless" and stop. That reading is wrong, and getting it right is what turns a null into a usable conclusion. The flag does not forecast renegotiation — but renegotiation forecasting was only ever one of the things a competition signal at the gate is good for, and arguably not the most defensible one.

8.1 A flag is a process-integrity signal, not a renegotiation forecast

The structural flags measure something real and verifiable: whether the tender was competitively contested. That is a legitimate concern of the fiscal gate in its own right — it bears on value for money, on corruption exposure, and on the audit-defensibility of the commitment — independently of whether it moves the renegotiation hazard. The error this paper warns against is not computing the flag; it is interpreting the flag as a renegotiation predictor. In an endemic-renegotiation regime, that interpretation is unsupported. The flag should be read as what it is.

It is worth being concrete about how the predictor reading fails, because the failure is not merely "no signal" — naively, the flag points the wrong way. A gate that screened its portfolio on the structural flags, treating flagged tenders as the high-renegotiation risks, would target the group that renegotiates least: on the primary outcome, the 54 flagged concessions renegotiate at 48% against 67% for the 69 unflagged ones, and single-bidder concessions at 46% against 62% for multi-bidder ones (Figure 6). The flag's concordance for renegotiation is 0.52 — a coin flip — against 0.85 for award vintage. A naive screening gate would thus spend its enhanced-scrutiny and retender budget on the safer half of its book while the riskier half passed through: the dashboard error of §6.1 turned into a resource-allocation error. This is the sharpest case for reading the flag as a process-integrity signal (§8.2) and provisioning off the base rate (§8.3), rather than as a risk-ranker.

Figure 6 — What a gate screening on the flag would do. Renegotiation rate (primary outcome) for the groups a structural-flag screen would sort on. The flagged groups — any structural flag, and single-bidder — renegotiate below the base rate and below the unflagged group, so a gate that scrutinised or retendered them would target the lower-risk half of its portfolio. The flag's concordance for renegotiation is 0.52 (a coin flip) versus 0.85 for award vintage.

8.2 The three-way action survives, re-calibrated

The clear / enhanced-due-diligence / refer-or-retender structure of §1.3 survives the null, but its trigger changes. The action is driven by process integrity, not by a predicted renegotiation hazard. Where a flag fires, the gate's enhanced-due-diligence step is not "expect this one to renegotiate" — the data does not support singling it out — but "confirm the competitive process was sound, document the justification, and ensure the renegotiation-control clauses are in place." The flag prompts a review of how the deal was let; it does not rank the deal's renegotiation risk against its peers.

8.3 The real fiscal lesson is the base rate and the size of the draw, not the flag

The most important numbers in this paper for a Ministry of Finance are not hazard ratios. They are 63% — the share of concessions that renegotiate on the broadest, any-modification measure (the §8.1 screening comparison uses the narrower contract-driven primary outcome, whose base rate is 58%; but provisioning should cover every draw on the envelope, including state-initiated works, so the broader rate is the right one here) — and the size of the draw when they do. Renegotiation here is the base case, not the exception, and it is uncorrelated with the competitive signal the gate might be tempted to screen on; so provisioning must be set off the endemic base rate, not off a flag that does not move it.

We can put a number on the draw, because the Modificaciones register records the estimated amount (in UF) of each modification. Among the 72 concessions with a recorded renegotiation amount, the cumulative renegotiated value runs to a median of 13.5% of the official project budget, a mean of 26.3% — the gap between the two is the long right tail — and a 90th percentile of 41.9%, nearly half the original budget added back through renegotiation. Across the whole portfolio, renegotiation adds about 15.7% to committed value. These are the magnitudes a fiscal gate should carry into its contingent-liability provisioning.

This is where the instrument meets the FCCL machinery. In the language of fiscal-risk tools such as the IMF–World Bank PFRAM, the renegotiation draw is a contingent outflow on the Stage-E envelope the MoF guaranteed at the gate. The honest provisioning rule the null implies is unconditional, not flag-conditioned: provision every concession at roughly the base-rate-weighted expected draw — on the order of the 63% incidence against a right-skewed severity whose median is ~14% and whose tail reaches ~42% of budget — rather than discounting the contracts a competition flag (wrongly) marks as safe. The instrument's genuine contribution to provisioning is to make that base rate and that severity distribution visible, and the contract-level renegotiation-control clause auditable — not to sort contracts into high- and low-risk bins it cannot distinguish. (Six further concessions renegotiated without a recorded amount, so the severity figures are conditional on disclosure and, if anything, understate the tail.)

8.4 Thresholds remain a cost-asymmetry problem

Setting the boundary between clear, EDD, and refer is still a cost-asymmetry calculation — the cost of an integrity miss against the cost of an unnecessary retender — but it is now an integrity calculation, not a predicted-hazard one. That is a narrower and more honest claim than the one we set out to make, and it is the claim the evidence supports.

8.5 A Monday-morning gate configuration

To make the instrument concrete for an adopting Treasury, here is a default configuration — explicitly integrity-triggered, not hazard-triggered (the data does not support ranking renegotiation risk by these flags), and meant to be tuned to a ministry's own cost-asymmetry.

The enhanced-due-diligence step asks was the deal let cleanly? — a value-for-money and audit-defensibility question — not will this one renegotiate?, which the evidence says the flag cannot answer.

Because the fiscal exposure is the renegotiation draw, and that draw is endemic (§8.3), the substantive protection lives in the contract, not in the gate's risk-sorting. The instrument's companion deliverable is therefore a renegotiation-control-clause checklist the gate can require, grounded in the Engel–Fischer–Galetovic and Bitran diagnoses of what actually drives Latin American renegotiation:

• a cumulative-renegotiation-value cap (for example, 20–25% of the original contract value) that binds the right tail the §8.3 severity distribution exposes;
• competitive re-pricing of any material scope addition above a threshold — an independent panel or benchmark, not bilateral negotiation with the incumbent;
• a state-versus-concessionaire firewall separating government-initiated scope changes (which should be budgeted as new public investment) from concessionaire compensation claims;
• mandatory machine-readable disclosure of every modification — the very data this paper had to hand-assemble, which a disclosing regime would produce as a by-product.

9. Architecture and reproduction in ten minutes

Why this matters: the series' discipline is that every framework is a running reference implementation, not a slide. The instrument here is no exception, and the null does not change the engineering claim: a Ministry of Finance can stand this capability up against its own data with one API call and a tamper-evident record of the result.

9.1 The pipeline

The flow is: a normalized tender record (or a panel of them) is submitted to the structural-indicator layer, which computes participation, timing, and — where bid amounts exist — price-dispersion flags benchmarked within sector; the text scanner runs alongside as a qualitative overlay (§2.2); the result is written to the platform's tamper-evident SHA-256 audit chain and can be returned with an optional signed PDF. The scan is deterministic — re-running the same panel reproduces the indicators exactly (the series' X-Replay-Match header is emitted by the platform's replay-scored modules; the structural scan, a pure function of its input panel, is reproduced by re-running it). Nothing in the pipeline is bespoke to a country: the same schema and engine ran the Chilean road panel and the Brazilian electricity panel in this paper.

9.2 One call

A Ministry of Finance gate integration is a single POST …/procurement/structural-scan with a panel of tender records; the response carries each record's indicators, fired flags with severities, and composite score, with the sector benchmark computed across the submitted panel. Appendix A gives the copy-pasteable call and a representative response.

9.3 The FCCL wire — honestly scoped

The score can be wired into the platform's fiscal-rules module so that a gate verdict is recorded against the Stage-E envelope. Given §8, we are explicit about what that wire should and should not carry: it surfaces the integrity flag and the endemic base-rate provisioning, and records the renegotiation-control clause as an auditable commitment. It does not emit a per-contract predicted renegotiation hazard, because we have no evidence such a prediction would be valid in this regime. Wiring the instrument to claim more than the data supports would be the dashboard error of §6.1 rebuilt in production.

9.4 A worked case, end to end

Take a real concession from the panel: the Segunda concesión Ruta 66, Camino de La Fruta (project 094_66RF2), an interurban toll road awarded in 2019. It drew one bidder. Submitted to the structural-scan endpoint inside the roads panel, it returns a single fired flag — single_bidder, high severity — and a structural score of 80. The gate sequence the rest of this paper licenses:

1. The flag fires → enhanced due diligence, not a renegotiation forecast. The gate does not record "expect this one to renegotiate" — §6 says it cannot. It records an integrity prompt: a single-bidder award on a major road warrants a documented justification from the contracting authority that the tender was genuinely contested, and a check that the renegotiation-control clauses (§8.5) are in the contract.
2. The contract carries the protection. A cumulative-renegotiation-value cap and competitive re-pricing of scope additions are what bound the fiscal exposure — because, as it happens, this concession went on to be modified five times, for a cumulative ~2.76 million UF, about 21% of its official budget. That draw is not evidence the flag "worked": 63% of the portfolio renegotiates, and §6 shows the flag does not identify which. It is evidence that the clause, not the flag, is where the fiscal control sits.
3. The provision is portfolio-level. The MoF provisions this concession not at a flag-conditioned hazard but at the endemic base rate against the severity distribution of §8.3 — the unconditional rule the null implies.

Every step runs against the shipped endpoint, is written to the SHA-256 audit chain, and is reproducible from the published DGC workbook in minutes. The instrument is real and the verdict is auditable; what changed, after the null, is the meaning of the flag the verdict rests on.

10. Limitations

Why this matters: the limitations are unusually load-bearing for this paper, because a null result is only as strong as the honesty around it. Two of them — the demotion of the orchestration contrast and the single-country hazard inference — bound the claims materially.

• Small-N, two-country, hand-built. The hazard inference rests on 123 Chilean concessions; the participation comparison adds 250 Brazilian ones. This is a built dataset, subject to survivorship and disclosure bias, not a live open-contracting feed.
• The panel is conditioned on survival. The DGC register carries only active and completed concessions — no cancelled or relicitación hand-back state (§4.1) — so concessions that failed and left the programme are absent from both the numerator and the denominator. That biases the renegotiation rate in no determinate direction (a failure removes an event and an exposure alike), but it means the 63% base rate and the vintage hazard describe the surviving population, not every concession ever signed. Because concessions enter the risk set at contract inception, there is no left-truncation to correct. A subtler version of this bias would bite the covariate, not just the rate: if the early hand-backs that never entered the register were disproportionately single-bidder lowballs that collapsed before renegotiating, survivorship would remove exactly the single-bidder-to-failure cases a positive hazard would need — and could manufacture the protective sign of §6.1. We cannot rule this out. The register does not identify the excluded concessions, and bounding the channel would require an external relicitación / caducidad hand-back list (MOP or the Panel Técnico) linked to award-stage bidder counts, which we do not have. It is the limitation most able to overturn the null, and the one a disclosure regime (§11) is most needed to close.
• The orchestration contrast is not tested. The planning outline for this paper proposed a Chile (CO-LEAD) versus Brazil (GATE) contrast — does stronger MoF orchestration attenuate the red-flag→renegotiation hazard? We demote that hypothesis to untested. Brazil has no programmatic renegotiation outcome, so there is no second hazard to compare; and N = 2 on the role dimension is not a test of a country-level proposition under any circumstances. We make no claim that orchestration moderates an effect we could not measure in the first place.
• One country carries the hazard inference. The null is a Chilean result. Its generalization to other endemic-renegotiation regimes is an argument (§7.3), not a second panel. A regime with low, non-endemic renegotiation is exactly where a competition flag might still carry information, and we have not tested one.
• The null is principally a demand-risk finding. Split by contract family, the null holds in both, but the demand-risk concessions (transport — roads and airports, n = 88) carry the inference (single-bidder HR 0.58 [0.23–1.47]), while the availability-payment family (social — hospitals and prisons, n = 31) is too thin to resolve and even leans the other way (HR 2.05 [0.67–6.34], p = 0.21). We do not generalise the null to availability-payment PPPs, where the state pays for availability rather than bearing demand risk, the renegotiation mechanism differs, and a competition flag might — intriguingly but inconclusively — yet carry information. This is a second reason (with §11.6's low-renegotiation regime) that the decisive next test lies outside Chile's demand-risk core.
• The instrument screens, it does not adjudicate. A fired flag is a prompt for review, never a finding of wrongdoing. Nothing in this paper licenses treating a single-bidder flag as evidence of impropriety in any individual tender.
• The price-dispersion demonstration is not a test. The Brazilian worked example (§3.3) shows the price family computes correctly on real auctions; it is a hand-coded handful, chosen because their bid tables are public, and it is decidedly not evidence that price dispersion predicts renegotiation. Brazil has no programmatic renegotiation outcome, so the price flags were never entered into the §6 hazard model and carry no inferential weight here.

11. The policy lead — concessions are dark

Why this matters: this is the finding with the longest reach. The hazard result is a clean answer to a narrow question in one country. The policy result is that, in almost every other country, the question cannot be asked at all — and that is a more consequential fact than the null itself.

11.1 We could only run the test because Chile publishes

Everything in §6 was possible because Chile's DGC publishes a structured concession register that links the participation record to the renegotiation outcome on a stable key. That register is rare. It is the precondition for the test, and most jurisdictions do not meet it.

11.2 Everywhere we looked, the gate is mostly blind

Concessions — the highest-value, highest-risk contracts a state signs — are carved out of open contracting in most of the jurisdictions we examined. The open-contracting movement has made real progress on routine public procurement, but concessions and PPPs sit largely outside it, and the bidder-level data the instrument needs is the first thing to disappear. In the seven jurisdictions of Appendix D, structured bidder-level concession data exists in only two (Chile, and for participation only, Brazil), and a programmatic renegotiation outcome in only one. Even regimes with strong PPP governance publish little or none of it: a project pipeline or a contract summary is not a participation record (Figure 7). We do not claim this for every country on earth — there are partial exceptions, and the European Union's tender notices and a handful of national portals carry more than the rest — but across the jurisdictions a researcher actually reaches the pattern is stark, and the slogan holds where we could test it: you cannot screen what you cannot see.

Figure 7 — Concessions are dark almost everywhere we looked. Structured, machine-readable bidder-level tender data and a programmatic renegotiation outcome, across seven jurisdictions. Chile is the only row that supports the full hazard test; the recurring entry is "No."

11.3 The binding constraint is disclosure, not modeling

It is worth stating the uncomfortable implication directly. The Chilean null is almost the good case: at least the data existed, the test could be run, and a clear answer came back. The worse and more common situation is the country where the gate is structurally blind and will stay blind — where neither a confirmatory result nor a null is reachable, because the participation record is locked in scanned PDFs or never published at all. The frontier for this work is not a better hazard model. It is disclosure. Consider the ceiling even at the top: a live query of the European Union's procurement database — the most complete in the world — returns 3,410 machine-readable concession awards with bidder counts, but only 196 concession-modification notices for the entire Union since 2023, of which a sixth cannot even be joined back to an award (Appendix E.1). If the gold standard discloses 196 renegotiation events, the screening frontier everywhere else is not the model; it is getting the record published at all.

11.4 The call

We call for structured concession-tender disclosure — at minimum the participation record (bidders, award flag, submission window) and, where possible, the bid table that the price-dispersion flags require — published in a machine-readable form and brought inside the open-contracting umbrella. The natural channels are the open-contracting and transparency communities that already maintain the standards and the distribution; concessions are the conspicuous gap in their coverage.

11.5 From gatekeeper to intelligent gate

The instrument is the route from a passive gatekeeper to an intelligent gate — and the null does not undo that. A gate that can see the competitive structure of the tenders it underwrites, price the endemic renegotiation base rate honestly, and defend its verdict with a tamper-evident record is intelligent relative to one that reads a one-page summary, regardless of whether any single flag forecasts renegotiation. Intelligence at the gate is about seeing and auditing, not about a single predictive coefficient.

11.6 Roadmap

• Add a low-renegotiation regime — and state the prediction that would falsify the null. The decisive next test is a jurisdiction where renegotiation is not endemic. Our reading of the mechanism (§7.3–§7.4) makes a falsifiable prediction: a competition flag should carry information about renegotiation where the base rate is low enough to leave tender-level variance for it to explain, and should stay null where renegotiation is endemic. A regime such as the United Kingdom or Korea — low renegotiation, but, as Appendix D shows, currently publishing no bidder-level data — is where the null should break if it is going to. Finding the flag predictive there would not overturn this paper; it would complete it. The most structured candidate is the EU concession regime via TED, whose schema turns out to support the test end to end; Appendix E.1 sets out the fields, the join key, the falsifiable prediction, and the disclosure obstacle that is itself the §11 argument in a second jurisdiction.
• Add the country-level renegotiation layer. Wire in the IADB / Guasch country-level renegotiation evidence — the follow-up Paper 3 flagged in its §14.3 roadmap — so the project-level instrument sits inside a country-level frame.
• The contestability score — built, and inert by design. The procedure axis (§6.5) motivates a graded contestability index: a procedure ceiling (open tender at the top, an unsolicited proposal with a right-to-match at the floor; Figure 8) modulated by the realised field within sector, and by any structural incumbent advantage. We built it and scored the Chilean panel (Figure 9), and it does exactly what it should: it spreads concessions across a 17–100 range and cleanly separates the procedures — open single-stage tenders average 74, two-stage 72, unsolicited private initiatives just 39 — and it carries no renegotiation signal whatsoever (hazard ratio 1.05 [0.82–1.35], p = 0.70; concordance 0.51, a coin flip, indistinguishable from the bare single-bidder flag at 0.52). That is the point, not a failure: it is a process-integrity score, and reporting it as an outcome forecast would rebuild the §6.1 error in a fancier index. The open item is to test it where it might predict — the floor case (the right-to-match Swiss challenge) needs a regime that uses it, the Philippines, Korea, or parts of India, which Chile's bonus-only USP does not.
• Scale the price-dispersion harvest. The §3.3 worked example hand-codes a handful of Brazilian auctions; the next step is a systematic extraction of competing-bid tables from the B3 / ANAC / ANTT atas into the panel — turning the demonstration into a Brazilian price-dispersion dataset large enough to test, if a renegotiation outcome can ever be joined to it. Appendix E.2 sets out that join concretely: hand-coding the ANTT termo aditivo record for the ~57 federal toll-road concessions would deliver both a second endemic-regime replication and the first hazard test of price dispersion.

Figure 8 — A contestability spectrum. The procedure sets a ceiling on the competitive pressure a tender can carry, from open competitive tender down through an unsolicited proposal with a right-to-match to direct award. It is a process-integrity construct — not, on this paper's evidence, a renegotiation predictor.

Figure 9 — The contestability index, built and applied to the 123 Chilean concessions. Panel A: the index spreads concessions across a 17–100 range and separates the procedure families (open single-stage highest, unsolicited private initiatives lowest), so it captures real procedure-and-field variation. Panel B: like the single-bidder flag, it is a coin flip for renegotiation (concordance 0.51 versus 0.85 for award vintage). It is a process-integrity measure, exactly as designed — not a renegotiation forecast.

Appendix A — Reproducible API call

The structural-indicator layer is reached with a single authenticated POST to the procurement module. The request is a panel of normalized tender records; sector benchmarks are computed across the submitted panel, so the below_sector_norm flag is panel-relative.

Request:

POST /workspaces/{workspace_id}/procurement/structural-scan
Authorization: Bearer <token>
Content-Type: application/json

{
  "records": [
    {"tender_id": "028_RVLC1", "sector": "Vial interurbana", "country": "Chile",
     "n_bidders": 1, "publication_date": "2018-03-01",
     "submission_deadline": "2018-04-15", "awarded_amount": 240000000, "currency": "USD"},
    {"tender_id": "031_RVLC2", "sector": "Vial interurbana", "n_bidders": 4,
     "publication_date": "2017-05-10", "submission_deadline": "2017-08-20"},
    {"tender_id": "045_RVLC3", "sector": "Vial interurbana", "n_bidders": 3,
     "publication_date": "2016-09-01", "submission_deadline": "2016-12-10"},
    {"tender_id": "052_RVLC4", "sector": "Vial interurbana", "n_bidders": 2,
     "publication_date": "2019-02-01", "submission_deadline": "2019-02-20"}
  ],
  "params": {}
}

Response (StructuralScanResponse, abridged — the two unflagged records score 100, and flag objects also carry label/why/value):

{
  "n_records": 4,
  "params": {"short_window_days": 30, "low_competition_max": 2,
             "low_dispersion_cv": 0.05, "large_gap_pct": 0.15, "min_sector_n": 3},
  "sector_benchmarks": [
    {"sector": "Vial interurbana", "n": 4, "median_bidders": 2.5,
     "mean_bidders": 2.5, "usable": true}
  ],
  "results": [
    {"tender_id": "028_RVLC1", "n_bidders": 1, "single_bidder": true,
     "submission_window_days": 45, "short_window": false, "below_sector_norm": true,
     "structural_flag_count": 1, "structural_score": 80.0,
     "flags": [{"code": "single_bidder", "severity": "high", "dimension": "participation"}]},
    {"tender_id": "052_RVLC4", "n_bidders": 2, "single_bidder": false,
     "submission_window_days": 19, "short_window": true, "below_sector_norm": true,
     "structural_flag_count": 3, "structural_score": 70.0,
     "flags": [{"code": "thin_competition", "severity": "medium", "dimension": "participation"},
               {"code": "below_sector_norm", "severity": "medium", "dimension": "participation"},
               {"code": "short_submission_window", "severity": "medium", "dimension": "timing"}]}
  ],
  "warnings": ["No competing-bid amounts in any record — price-dispersion indicators skipped (expected for Chile-style panels)."]
}

The response returns, per record, the computed StructuralIndicators and the list of fired flags with their severities, plus the panel's sector benchmarks. (This endpoint does not emit the series' X-Replay-Match header — that header is carried by the platform's replay-scored modules; the structural scan, being a pure function of its input panel, is reproduced instead by re-running the same panel against the published workbook.) The indicator catalogue (codes, severities, definitions) is available without scanning a panel at GET /workspaces/{workspace_id}/procurement/structural/indicators.

Appendix B — Indicator reference and flag formulas

The composite structural_score is 100 − Σ severity_weight(flag), floored at 0, with severity weights high = 20, medium = 10, low = 5. The price flags require bid-level amounts and are computed only where those exist (Brazil in this paper, not Chile).

Appendix C — Glossary

• CRI (Corruption Risk Index) — a composite of structural procurement indicators, built around single-bidding, validated against independent corruption measures.
• Hazard ratio (HR) — the multiplicative effect of a covariate on the instantaneous rate of the event (here, first renegotiation). HR = 1 means no effect.
• Concordance (Harrell's C) — the probability the model ranks a random event-pair in the correct order; 0.5 is a coin flip, 1.0 is perfect.
• Cox proportional-hazards model — a semi-parametric survival model for the effect of covariates on the hazard, without specifying the baseline hazard's shape.
• Kaplan–Meier / log-rank — a non-parametric survival-curve estimator and the test comparing two such curves.
• Discrete-time hazard (cloglog) — a survival model fit on annual person-period data with a complementary log-log link.
• Censoring — an observation whose event has not occurred by the end of observation; here, right-censored at the register snapshot or contract end.
• Modificación — in the DGC register, any recorded amendment to a concession contract; the unit of the renegotiation outcome in this paper, and inclusive of state-initiated Nuevas obras, which the primary outcome excludes.
• Convenio complementario — the specific Chilean contractual instrument formalising a negotiated change to a concession (a subtype of modification); historically the principal vehicle through which renegotiation was executed.
• Relicitação — the Brazilian process of handing a concession back for re-tender.
• FCCL — fiscal commitments and contingent liabilities; the framework for the public obligations a PPP creates.
• CA (contracting authority) — the public entity that runs the tender and signs the contract; distinct from the MoF under the orchestrator partition.

Appendix D — Data-availability and role-cell mapping

The recurring obstacle is the absence of structured, bidder-level concession-tender data. The table records, for seven jurisdictions, whether the participation record and the renegotiation outcome are available in structured form, the source we checked, and our confidence in the assessment.

Availability assessed against the named portals as of mid-2026. "Verified present/absent" means we confirmed the portal does or does not carry a structured bidder-level field; "not found in structured form" means we did not locate a structured source and the data, where it exists, is in scanned PDFs. This is a seven-jurisdiction convenience sample, not an exhaustive global census — partial exceptions exist (the EU's TED notices carry some concession bidder data) — but it spans the regimes a researcher reaches first, and the pattern is the policy argument of §11 in one table: even strong-governance regimes publish little or no structured bidder-level concession data, and a programmatic renegotiation outcome is available in only one of the seven. The orchestration contrast the planning outline proposed (Chile CO-LEAD vs Brazil GATE under Paper 3's role-cells) is not tested here, for the reasons set out in §10.

Appendix E — Two scoped extensions (blueprints, not results)

The roadmap of §11.6 names two concrete next tests. Because this paper's discipline is that empirical claims rest on data actually built, we set these out as blueprints — with the real data sources, join keys, sample bounds, and the obstacles we verified against the live portals — rather than as results. Neither is run here; both are bounded by the same data-darkness the paper documents.

E.1 The decisive test — EU concessions via TED (a lower-renegotiation regime)

The test §11.6 calls for is a non-endemic renegotiation regime, where a competition flag has room to carry information. The European Union is the most promising structured candidate — and, usefully, single-bidding on EU concessions runs at about 23% (2023), essentially identical to Chile's 23%, so the flag side is comparable while the outcome regime is not.

The schema supports the test end to end. In the EU's eForms standard, a concession award notice (notice subtype 32, "concession award notice — concessions directive") carries a bidder count in business term BT-759 "Received Submissions Count" at the lot-result level — the single-bidder flag, natively. A concession contract-modification notice (subtype 40), the vehicle for Article 43 of the Concessions Directive (2014/23/EU), is the renegotiation event; and — the pleasant surprise — it carries an explicit back-reference to the original contract (BT-1501 "Modified Contract Reference"), so the award→renegotiation join is a schema key, not a fuzzy buyer-and-title match. The data is openly queryable (the TED Search API requires no key) and downloadable in bulk (TED XML packages and the Public Procurement Data Space launched September 2024).

Reconnaissance — we queried it. A live query of the TED API (June 2026) confirms both the promise and the ceiling. The bidder-count side is healthy: 3,410 concession-award notices are retrievable in the eForms era, with BT-759 populated in essentially all of them (about 80% non-zero, the zeros being genuinely uncontested awards) — single-bidding is directly measurable on EU concessions, and runs at the ~23% the EU's own Scoreboard reports. The outcome side is the wall: the entire European Union has published just 196 concession-modification notices in eForms, and after dropping back-references that are internal identifiers or broken placeholders, only about 160 link back to a notice at all — at most a few hundred clean award→modification dyads, every one of them post-October-2023. That is not a survival panel; it is a demonstration. And the binding constraint is not the API or the schema — both work — but disclosure: the most complete procurement database in the world has filed 196 concession-modification notices, and Article 43 under-filing means even that over-states what is observable. A TED extension would therefore test the null at demonstration scale and, more powerfully, re-document the data-darkness argument in the one regime that should be immune to it.

The obstacles are themselves the paper's thesis. Three bound this to a small-N, explicitly-caveated design rather than a high-powered panel: (i) a schema discontinuity — the clean BT-759 / BT-1501 structure is eForms-native (mandatory from October 2023), so a clean join is effectively a post-2024 panel, the legacy schema before it having no equivalent key; (ii) small N — concessions are ~12% of directive-covered procurement (~€63bn/yr, low-thousands of award notices, concentrated in France and Italy) and only above the ~€5.4M threshold; and decisively (iii) modification-notice under-filing — if authorities do not file Article 43 notices, the renegotiation event is unobserved and the regime looks artificially low-renegotiation. That third obstacle is not noise to clean away; it is the §11 argument in a second jurisdiction — the gate is blind not because the regime renegotiates less, but because the renegotiation is not disclosed — so a TED extension would need a disclosure-completeness audit before any low-renegotiation reading could be trusted.

Falsifiable prediction. If renegotiation in the EU concession regime is genuinely less endemic than in Chile and adequately disclosed, the single-bidder flag should carry a positive, significant renegotiation hazard there — the result this paper failed to find in Chile. Finding it would not overturn the Chilean null; it would confirm the mechanism of §7.3: the flag bites where the base rate leaves it room to.

E.2 The replication — Brazilian federal toll roads via ANTT (a second endemic regime, with price dispersion)

The most efficient second hazard panel is not a new country but the completion of one already half-built. The Brazil panel (250 PPI concessions, 92 transport) has the participation flags but no programmatic renegotiation outcome (§4.3). Hand-coding that outcome for the ~57 federal toll-road concessions would give a second endemic-regime hazard test — and, because Brazilian auctions publish competing bids, the first test of the price-dispersion flag the paper can only demonstrate in Chile (§3.3).

Where the outcome lives, and the join. There is no structured federal dataset of contract amendments — a 2023 audit by Brazil's federal audit court (TCU) found explicitly that ANTT does not publish concession data in open formats and recommended it build one. The renegotiation record exists as "Extrato de Termo Aditivo" acts on ANTTlegis (searchable by act type and year) and as full-text PDFs on each concessionaire's page; the first-amendment date is recoverable per concession, by hand. The join to the PPI panel has no shared key — PPI carries no CNPJ, edital, or contract number — so it is a manual crosswalk on concessionaire/highway name and auction year, roughly 57 dossiers, on the order of a focused person-month.

The hard design choices. Two are load-bearing. First, what counts as a renegotiation event: a termo aditivo ranges from a trivial asset-registry tweak to a major economic rebalancing, so the outcome must be defined on amendment type, not raw count, or it is noise — the same problem §5.3 handles for Chilean Nuevas obras, but without a clean typology. Second, regulator heterogeneity: a São Paulo (ARTESP) sub-panel is no easier (also PDF/RAR, no structured fields), and pooling federal ANTT with state ARTESP concessions would require a regulator fixed effect and a federal-only sensitivity run. PPI's own status flags (Cancelled / Distressed / Concluded) are too coarse to be the outcome and would enter only as a secondary end-of-life check.

What it would add. A replication of the null in a second endemic regime (strengthening §7.3), and the first hazard test of price dispersion — the strongest flag in the integrity literature, and the one Chile cannot see. It is Paper-5b-scale work, set out here so the path is concrete rather than gestured at.

References & notes